notebooks/Agents_and_Tools/llm_agents_pharma.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# LLM Agents for Clinical Trial Management\n",
    "\n",
    "In this notebook, we develop an agent-based Large Language Model (LLM) application to automate the evaluation of patients for potential clinical trials. By leveraging documents related to patients' medical histories, clinical policies, and trial inclusion criteria, such an agentic system can significantly enhance the quality of evaluations and reduce the time required.\n",
    "\n",
    "We use limited demo databases and simple agentic tools and pipelines to demonstrate the capabilities of such LLM systems. For real-world applications, more comprehensive medical databases, in collaboration with clinical specialists, will be necessary. Additionally, more advanced tools (e.g., advanced Retrieval-Augmented Generation (RAG) methods) must be incorporated for real applications.\n",
    "\n",
    "**Objective:** Automate clinical trial data management, analysis, compliance checks, and reporting.\n",
    "\n",
    "## Key Components of This Design: \n",
    "1. Patient Data Collection: Gather comprehensive medical histories and relevant patient information.\n",
    "2. Patient Data Analysis: Analyze patient data to identify key health indicators and relevant medical conditions.\n",
    "3. Clinical Compliance Verification: Ensure patient data complies with clinical policies and trial eligibility criteria.\n",
    "4. Trial Matching: Match patients with suitable clinical trials based on their medical profiles.\n",
    "5. Hallucination Grader: Guardrailing the LLM outputs to ensure relevance and accuracy, preventing hallucinations.\n",
    "6. Human-in-the-Loop Interventions: Allow clinical experts to review and modify intermediate and final outcomes.\n",
    "\n",
    "\n",
    "## Visual Representation of the Pipeline:\n",
    "\n",
    "![pipeline](../../images/llm_pharma_diagram.png) \n",
    "\n",
    "\n",
    "## Technologies and Tools Used:\n",
    "- Local Models: Embedding of policies and trial information.\n",
    "- Local Vector Stores: Information retrieval using metadata.\n",
    "- Information Retrieval: Using metadata to find relevant documents.\n",
    "- Chains of LLM Calls: Sequential LLM calls to perform complex tasks.\n",
    "- Python Shell: LLM integration for performing numerical calculations.\n",
    "- Agentic Graph: Deploying a workflow for evaluating patients and trials.\n",
    "- Human-in-the-Loop: Clinical experts can intervene and modify the evaluation process and results.\n",
    "\n",
    "## Video Demonstration:\n",
    "[Watch the demo videos on YouTube.](https://www.youtube.com/playlist?list=PLMtE8Mev6Cct7n4NRsKVTgP8N5IzDSjbq)\n",
    "\n",
    "These sample videos demonstrate the application use and its functionalities such as Profile (re-)generation, Policy Evaluation, Human in-the-loop interventions, Hallucination Guardrail, and Retrieval Grader."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2\n",
    "import getpass\n",
    "import os\n",
    "\n",
    "\n",
    "def _set_if_undefined(var: str):\n",
    "    if not os.environ.get(var):\n",
    "        os.environ[var] = getpass.getpass(f\"Please provide your {var}\")\n",
    "\n",
    "# Loading the requied environment variables\n",
    "_set_if_undefined(\"OPENAI_API_KEY\")\n",
    "_set_if_undefined(\"LANGCHAIN_API_KEY\")\n",
    "\n",
    "# Optional, add tracing in LangSmith\n",
    "os.environ[\"LANGCHAIN_TRACING_V2\"] = \"true\"\n",
    "os.environ[\"LANGCHAIN_PROJECT\"] = \"Clinical Trial Management\""
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Parameters"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain_openai import ChatOpenAI\n",
    "\n",
    "\n",
    "policy_vstore_created = False\n",
    "trials_vstore_created = False\n",
    "trials_db_downloaded = False\n",
    "chromadb_path = \"../../chroma_db\"\n",
    "policy_collection = \"policy-chroma\"\n",
    "trial_collection = \"trial-chroma\"\n",
    "\n",
    "modelID = \"gpt-3.5-turbo\"\n",
    "n_retrieved_policies = 5 # Number of most relevant policies to retrieve per patient\n",
    "n_retrieved_trials = 6\n",
    "\n",
    "model = ChatOpenAI(temperature = 0.0, model=modelID)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "import warnings\n",
    "warnings.filterwarnings(\"ignore\")\n",
    "\n",
    "import pprint\n",
    "# A function for printing nicely\n",
    "def nprint(text, indent=2):\n",
    "    pp = pprint.PrettyPrinter(indent=indent)\n",
    "    pp.pprint(text)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Creating the demo datasets"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Patients database\n",
    "\n",
    "We create a sample dataset of patients with their medical history, previous trials, and other relevant information.   \n",
    "We use random names, diseases, and past or ongoing trials.  \n",
    "This is a demo dataset limited information. However, for real cases, a more complit medical history of patients will be used."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>patient_id</th>\n",
       "      <th>name</th>\n",
       "      <th>age</th>\n",
       "      <th>medical_history</th>\n",
       "      <th>previous_trials</th>\n",
       "      <th>trial_status</th>\n",
       "      <th>trial_completion_date</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>1</td>\n",
       "      <td>Alice Johnson</td>\n",
       "      <td>66</td>\n",
       "      <td>depressive disorder</td>\n",
       "      <td>NCT02483538</td>\n",
       "      <td>Withdrawn</td>\n",
       "      <td>2023-05-29</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>2</td>\n",
       "      <td>Alice Johnson</td>\n",
       "      <td>78</td>\n",
       "      <td>myelomonocytic leukemia</td>\n",
       "      <td>NCT04366575</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>3</td>\n",
       "      <td>Michael Anderson</td>\n",
       "      <td>35</td>\n",
       "      <td>Depression</td>\n",
       "      <td>NCT09379910</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>4</td>\n",
       "      <td>James Johnson</td>\n",
       "      <td>23</td>\n",
       "      <td>myeloma</td>\n",
       "      <td>NCT07318260</td>\n",
       "      <td>Withdrawn</td>\n",
       "      <td>2023-06-30</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>5</td>\n",
       "      <td>Alice Garcia</td>\n",
       "      <td>66</td>\n",
       "      <td>lymphoblastic leukemia</td>\n",
       "      <td>NCT05620361</td>\n",
       "      <td>Completed</td>\n",
       "      <td>2023-01-16</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>6</td>\n",
       "      <td>Jane Brown</td>\n",
       "      <td>52</td>\n",
       "      <td>mesothelioma</td>\n",
       "      <td>NCT09453025</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>7</td>\n",
       "      <td>Emma Doe</td>\n",
       "      <td>27</td>\n",
       "      <td>liposarcoma</td>\n",
       "      <td>NCT06730532</td>\n",
       "      <td>Completed</td>\n",
       "      <td>2023-05-20</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>8</td>\n",
       "      <td>Alice Doe</td>\n",
       "      <td>44</td>\n",
       "      <td>lymphoblastic leukemia</td>\n",
       "      <td>NCT04593772</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>9</td>\n",
       "      <td>Daniel Smith</td>\n",
       "      <td>77</td>\n",
       "      <td>soft tissue sarcoma</td>\n",
       "      <td>NCT04222635</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>9</th>\n",
       "      <td>10</td>\n",
       "      <td>John Brown</td>\n",
       "      <td>47</td>\n",
       "      <td>mesothelioma</td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>10</th>\n",
       "      <td>11</td>\n",
       "      <td>John Brown</td>\n",
       "      <td>22</td>\n",
       "      <td>gastric cancer</td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>11</th>\n",
       "      <td>12</td>\n",
       "      <td>James Martinez</td>\n",
       "      <td>33</td>\n",
       "      <td>lymphoblastic leukemia</td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>12</th>\n",
       "      <td>13</td>\n",
       "      <td>Jane Davis</td>\n",
       "      <td>56</td>\n",
       "      <td>myeloid leukemia</td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>13</th>\n",
       "      <td>14</td>\n",
       "      <td>John Doe</td>\n",
       "      <td>59</td>\n",
       "      <td>Chronic Pain</td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>14</th>\n",
       "      <td>15</td>\n",
       "      <td>Emily Martinez</td>\n",
       "      <td>51</td>\n",
       "      <td>lymphoblastic leukemia</td>\n",
       "      <td>NCT05554154</td>\n",
       "      <td>Withdrawn</td>\n",
       "      <td>2024-03-23</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>15</th>\n",
       "      <td>16</td>\n",
       "      <td>Daniel Wilson</td>\n",
       "      <td>42</td>\n",
       "      <td>Chronic Pain</td>\n",
       "      <td>NCT02902950</td>\n",
       "      <td>Completed</td>\n",
       "      <td>2024-05-16</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>16</th>\n",
       "      <td>17</td>\n",
       "      <td>Emily Wilson</td>\n",
       "      <td>22</td>\n",
       "      <td>mesothelioma</td>\n",
       "      <td>NCT07762208</td>\n",
       "      <td>Completed</td>\n",
       "      <td>2022-07-22</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>17</th>\n",
       "      <td>18</td>\n",
       "      <td>John Johnson</td>\n",
       "      <td>58</td>\n",
       "      <td>Heart Disease</td>\n",
       "      <td>NCT01114491</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>18</th>\n",
       "      <td>19</td>\n",
       "      <td>James Johnson</td>\n",
       "      <td>20</td>\n",
       "      <td>suicidal ideation</td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>19</th>\n",
       "      <td>20</td>\n",
       "      <td>John Anderson</td>\n",
       "      <td>77</td>\n",
       "      <td>Anxiety</td>\n",
       "      <td>NCT04639247</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td></td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "    patient_id              name  age          medical_history  \\\n",
       "0            1     Alice Johnson   66      depressive disorder   \n",
       "1            2     Alice Johnson   78  myelomonocytic leukemia   \n",
       "2            3  Michael Anderson   35               Depression   \n",
       "3            4     James Johnson   23                  myeloma   \n",
       "4            5      Alice Garcia   66   lymphoblastic leukemia   \n",
       "5            6        Jane Brown   52             mesothelioma   \n",
       "6            7          Emma Doe   27              liposarcoma   \n",
       "7            8         Alice Doe   44   lymphoblastic leukemia   \n",
       "8            9      Daniel Smith   77      soft tissue sarcoma   \n",
       "9           10        John Brown   47             mesothelioma   \n",
       "10          11        John Brown   22           gastric cancer   \n",
       "11          12    James Martinez   33   lymphoblastic leukemia   \n",
       "12          13        Jane Davis   56         myeloid leukemia   \n",
       "13          14          John Doe   59             Chronic Pain   \n",
       "14          15    Emily Martinez   51   lymphoblastic leukemia   \n",
       "15          16     Daniel Wilson   42             Chronic Pain   \n",
       "16          17      Emily Wilson   22             mesothelioma   \n",
       "17          18      John Johnson   58            Heart Disease   \n",
       "18          19     James Johnson   20        suicidal ideation   \n",
       "19          20     John Anderson   77                  Anxiety   \n",
       "\n",
       "   previous_trials trial_status trial_completion_date  \n",
       "0      NCT02483538    Withdrawn            2023-05-29  \n",
       "1      NCT04366575      Ongoing                        \n",
       "2      NCT09379910      Ongoing                        \n",
       "3      NCT07318260    Withdrawn            2023-06-30  \n",
       "4      NCT05620361    Completed            2023-01-16  \n",
       "5      NCT09453025      Ongoing                        \n",
       "6      NCT06730532    Completed            2023-05-20  \n",
       "7      NCT04593772      Ongoing                        \n",
       "8      NCT04222635      Ongoing                        \n",
       "9                                                      \n",
       "10                                                     \n",
       "11                                                     \n",
       "12                                                     \n",
       "13                                                     \n",
       "14     NCT05554154    Withdrawn            2024-03-23  \n",
       "15     NCT02902950    Completed            2024-05-16  \n",
       "16     NCT07762208    Completed            2022-07-22  \n",
       "17     NCT01114491      Ongoing                        \n",
       "18                                                     \n",
       "19     NCT04639247      Ongoing                        "
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import pandas as pd\n",
    "from datetime import datetime, timedelta\n",
    "import numpy as np\n",
    "import json\n",
    "import random\n",
    "\n",
    "# Generate the sample data again for display\n",
    "columns = [\"patient_id\", \"name\", \"age\", \"medical_history\", \"previous_trials\", \"trial_status\", \"trial_completion_date\"]\n",
    "data = []\n",
    "\n",
    "# Given names and surnames\n",
    "names = [\"John\", \"Jane\", \"Alice\", \"Michael\", \"Emily\", \"Daniel\", \"Sophia\", \"James\", \"Emma\", \"Oliver\"]\n",
    "surnames = [\"Doe\", \"Smith\", \"Johnson\", \"Brown\", \"Davis\", \"Garcia\", \"Martinez\", \"Anderson\", \"Thomas\", \"Wilson\"]\n",
    "\n",
    "# Generate all possible unique combinations of names and surnames\n",
    "combinations = [(name, surname) for name in names for surname in surnames]\n",
    "\n",
    "# Shuffle the combinations to ensure randomness\n",
    "random.shuffle(combinations)\n",
    "\n",
    "# Select the first 100 unique combinations\n",
    "unique_names = combinations[:100]\n",
    "\n",
    "# Generate the full names\n",
    "full_names = [f\"{name} {surname}\" for name, surname in unique_names]\n",
    "\n",
    "with open('../../source_data/diseases_list.json', 'r') as file:\n",
    "        trial_diseases =  json.load(file)\n",
    "\n",
    "list_trial_diseases = list({disease for diseases in trial_diseases.values() for disease in diseases})\n",
    "\n",
    "other_medical_conditions = [\"Hypertension\", \"Diabetes\", \"Asthma\", \"Heart Disease\", \"Arthritis\",\n",
    "                      \"Chronic Pain\", \"Anxiety\", \"Depression\", \"Obesity\"]\n",
    "\n",
    "all_conditions = list(set(list_trial_diseases + other_medical_conditions))\n",
    "\n",
    "trial_statuses = [\"Completed\", \"Ongoing\", \"Withdrawn\"]\n",
    "\n",
    "def random_date(start, end):\n",
    "    return start + timedelta(days=random.randint(0, int((end - start).days)))\n",
    "\n",
    "# start_date must be 2 years before now\n",
    "start_date = datetime.now() - timedelta(days=365 * 2)\n",
    "# start_date = datetime(2020, 1, 1)\n",
    "\n",
    "# end_date must be a month before now\n",
    "end_date = datetime.now() - timedelta(days=10)\n",
    "# end_date = datetime(2023, 1, 1)\n",
    "\n",
    "for i in range(1, 101):\n",
    "    name = random.choice(full_names)\n",
    "    age = random.randint(20, 80)\n",
    "    medical_history = random.choice(all_conditions)\n",
    "    \n",
    "    if random.choice([True, False]):\n",
    "        previous_trials = f\"NCT0{random.randint(1000000, 9999999)}\"\n",
    "        trial_status = random.choice(trial_statuses)\n",
    "        trial_completion_date = random_date(start_date, end_date).strftime('%Y-%m-%d')\n",
    "    else:\n",
    "        previous_trials = \"\"\n",
    "        trial_status = \"\"\n",
    "        trial_completion_date = \"\"\n",
    "    if trial_status == \"Ongoing\":\n",
    "        trial_completion_date = \"\"\n",
    "\n",
    "    data.append((i, name, age, medical_history, previous_trials, trial_status, trial_completion_date))\n",
    "\n",
    "df = pd.DataFrame(data, columns=columns)\n",
    "# save df to csv\n",
    "df.to_csv(\"patients.csv\", index=False)\n",
    "df.head(20)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Creating a local SQLite database for patients"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "100\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>patient_id</th>\n",
       "      <th>name</th>\n",
       "      <th>age</th>\n",
       "      <th>medical_history</th>\n",
       "      <th>previous_trials</th>\n",
       "      <th>trial_status</th>\n",
       "      <th>trial_completion_date</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>1</td>\n",
       "      <td>Alice Johnson</td>\n",
       "      <td>66</td>\n",
       "      <td>depressive disorder</td>\n",
       "      <td>NCT02483538</td>\n",
       "      <td>Withdrawn</td>\n",
       "      <td>2023-05-29</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>2</td>\n",
       "      <td>Alice Johnson</td>\n",
       "      <td>78</td>\n",
       "      <td>myelomonocytic leukemia</td>\n",
       "      <td>NCT04366575</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td>None</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>3</td>\n",
       "      <td>Michael Anderson</td>\n",
       "      <td>35</td>\n",
       "      <td>Depression</td>\n",
       "      <td>NCT09379910</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td>None</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>4</td>\n",
       "      <td>James Johnson</td>\n",
       "      <td>23</td>\n",
       "      <td>myeloma</td>\n",
       "      <td>NCT07318260</td>\n",
       "      <td>Withdrawn</td>\n",
       "      <td>2023-06-30</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>5</td>\n",
       "      <td>Alice Garcia</td>\n",
       "      <td>66</td>\n",
       "      <td>lymphoblastic leukemia</td>\n",
       "      <td>NCT05620361</td>\n",
       "      <td>Completed</td>\n",
       "      <td>2023-01-16</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>6</td>\n",
       "      <td>Jane Brown</td>\n",
       "      <td>52</td>\n",
       "      <td>mesothelioma</td>\n",
       "      <td>NCT09453025</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td>None</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>7</td>\n",
       "      <td>Emma Doe</td>\n",
       "      <td>27</td>\n",
       "      <td>liposarcoma</td>\n",
       "      <td>NCT06730532</td>\n",
       "      <td>Completed</td>\n",
       "      <td>2023-05-20</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>8</td>\n",
       "      <td>Alice Doe</td>\n",
       "      <td>44</td>\n",
       "      <td>lymphoblastic leukemia</td>\n",
       "      <td>NCT04593772</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td>None</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>9</td>\n",
       "      <td>Daniel Smith</td>\n",
       "      <td>77</td>\n",
       "      <td>soft tissue sarcoma</td>\n",
       "      <td>NCT04222635</td>\n",
       "      <td>Ongoing</td>\n",
       "      <td>None</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>9</th>\n",
       "      <td>10</td>\n",
       "      <td>John Brown</td>\n",
       "      <td>47</td>\n",
       "      <td>mesothelioma</td>\n",
       "      <td>None</td>\n",
       "      <td>None</td>\n",
       "      <td>None</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   patient_id              name  age          medical_history previous_trials  \\\n",
       "0           1     Alice Johnson   66      depressive disorder     NCT02483538   \n",
       "1           2     Alice Johnson   78  myelomonocytic leukemia     NCT04366575   \n",
       "2           3  Michael Anderson   35               Depression     NCT09379910   \n",
       "3           4     James Johnson   23                  myeloma     NCT07318260   \n",
       "4           5      Alice Garcia   66   lymphoblastic leukemia     NCT05620361   \n",
       "5           6        Jane Brown   52             mesothelioma     NCT09453025   \n",
       "6           7          Emma Doe   27              liposarcoma     NCT06730532   \n",
       "7           8         Alice Doe   44   lymphoblastic leukemia     NCT04593772   \n",
       "8           9      Daniel Smith   77      soft tissue sarcoma     NCT04222635   \n",
       "9          10        John Brown   47             mesothelioma            None   \n",
       "\n",
       "  trial_status trial_completion_date  \n",
       "0    Withdrawn            2023-05-29  \n",
       "1      Ongoing                  None  \n",
       "2      Ongoing                  None  \n",
       "3    Withdrawn            2023-06-30  \n",
       "4    Completed            2023-01-16  \n",
       "5      Ongoing                  None  \n",
       "6    Completed            2023-05-20  \n",
       "7      Ongoing                  None  \n",
       "8      Ongoing                  None  \n",
       "9         None                  None  "
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import os\n",
    "import sqlite3\n",
    "import pandas as pd\n",
    "\n",
    "overwrite = True\n",
    "\n",
    "database_file = 'patients_database.db'\n",
    "df = pd.read_csv('patients.csv')\n",
    "if overwrite or not os.path.exists(database_file):\n",
    "    if os.path.exists(database_file):\n",
    "        os.remove(database_file)\n",
    "    conn = sqlite3.connect(database_file)\n",
    "    cursor = conn.cursor()\n",
    "    cursor.execute('''\n",
    "    CREATE TABLE IF NOT EXISTS patients (\n",
    "        patient_id INTEGER PRIMARY KEY,\n",
    "        name TEXT,\n",
    "        age INTEGER,\n",
    "        medical_history TEXT,\n",
    "        previous_trials TEXT,\n",
    "        trial_status TEXT,\n",
    "        trial_completion_date TEXT\n",
    "    )\n",
    "    ''')\n",
    "\n",
    "    # Insert DataFrame into SQLite table\n",
    "    df.to_sql('patients', conn, if_exists='append', index=False)\n",
    "\n",
    "    # Commit and close the connection\n",
    "    conn.commit()\n",
    "else:\n",
    "    conn = sqlite3.connect(database_file)\n",
    "    cursor = conn.cursor()\n",
    "\n",
    "query = 'SELECT * FROM patients'\n",
    "\n",
    "# Execute the query and fetch all results\n",
    "cursor.execute(query)\n",
    "rows = cursor.fetchall()\n",
    "\n",
    "# Optionally, you can get the column names\n",
    "column_names = [description[0] for description in cursor.description]\n",
    "\n",
    "# Convert the results to a Pandas DataFrame for better readability\n",
    "df = pd.DataFrame(rows, columns=column_names)\n",
    "print(len(df))\n",
    "\n",
    "# Display the DataFrame\n",
    "conn.close()\n",
    "\n",
    "db = database_file\n",
    "df.head(10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "# df[df['medical_history'].str.contains('Obes')]\n",
    "# df[df['medical_history'].str.contains('Pain')]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Clinical Trials database\n",
    "\n",
    "We download a sample clinical trials database from a [publicly available source]('https://raw.githubusercontent.com/futianfan/clinical-trial-outcome-prediction/main/data/raw_data.csv').   \n",
    "Details are provided in the [helper function](../../src/helper_functions.py)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Only trials with status recruiting are selected.\n",
      "The database for trials is saved to ../../data/trials_data.csv \n",
      " It has 30 rows.\n",
      "---- Sample 1 ----\n",
      "\n",
      "        Inclusion Criteria:\n",
      "\n",
      "          -  Patients have provided a signed Informed Consent Form\n",
      "\n",
      "          -  ECOG performance status of 0-1\n",
      "\n",
      "          -  BMI ≥ 18.5\n",
      "\n",
      "          -  Age: 18-65 years old\n",
      "\n",
      "          -  Histologically confirmed diagnosis of metastatic pancreatic cancer\n",
      "\n",
      "          -  No prior palliative chemotherapy\n",
      "\n",
      "          -  Measurable disease in at least 1 diameter by CT scan or MRI as per RECIST 1.1 criteria\n",
      "\n",
      "          -  Life expectancy ≥ 3 months\n",
      "\n",
      "          -  Patient has adequate bone marrow and organ function\n",
      "\n",
      "               -  Absolute Neutrophil Count (ANC) ≥ 2.0 x 109/L\n",
      "\n",
      "               -  Platelets ≥ 90 x 109/L\n",
      "\n",
      "               -  Hemoglobin ≥ 90 g/L\n",
      "\n",
      "          -  Patient has adequate liver function\n",
      "\n",
      "               -  AST and ALT not more than 2.5 times ULN (not more than 5.0 times ULN if there is\n",
      "                  liver metastasis)\n",
      "\n",
      "               -  Serum bilirubin ≤ 1.2 x ULN\n",
      "\n",
      "          -  Creatinine ≤ 1.25 times ULN\n",
      "\n",
      "          -  Good compliance\n",
      "\n",
      "        Exclusion Criteria:\n",
      "\n",
      "          -  Pregnant or lactating women\n",
      "\n",
      "          -  Brain metastasis or only with bone metastasis.\n",
      "\n",
      "          -  Patients with severe gastrointestinal hemorrhage which need frequent blood\n",
      "             transfusions.\n",
      "\n",
      "          -  Refuse to take appropriate contraceptive measures (including male patients).\n",
      "\n",
      "          -  Allergic to Oxaliplatin, Irinotecan, Leucovorin or 5-Fluorouracil.\n",
      "\n",
      "          -  Severe systemic disease out of control such as unstable or uncompensated respiratory,\n",
      "             cardiac, liver, renal diseases.\n",
      "\n",
      "          -  Patient has a concurrent malignancy or has a malignancy within 5 years of study\n",
      "             enrollment, (with the exception of non-melanoma skin cancer or cervical carcinoma in\n",
      "             situ).\n",
      "\n",
      "          -  Psychiatric illness that would prevent the patient from giving informed consent.\n",
      "\n",
      "          -  Patient is concurrently using other antineoplastic agent\n",
      "\n",
      "          -  Patient has used investigational antineoplastic agent within 4 weeks prior to entry.\n",
      "\n",
      "          -  Known HIV-positivity.\n",
      "\n",
      "          -  No history of chronic diarrhea, nausea or vomit.\n",
      "\n",
      "          -  No ≥ grade 2 sensory peripheral neuropathy.\n",
      "\n",
      "          -  A history of transmural myocardial infarction (within 6 months prior to entry),\n",
      "             congestive heart failure, and unstable angina.\n",
      "\n",
      "          -  Infectious disease or inflammation with body temperature ≥ 38 ℃.\n",
      "      \n",
      "---- Sample 2 ----\n",
      "\n",
      "        Inclusion Criteria:\n",
      "\n",
      "          -  HCV patients who will undergo treatment by direct acting antiviral agents.\n",
      "\n",
      "        Exclusion Criteria:\n",
      "\n",
      "          -  Patients with decompensated liver disease\n",
      "      \n"
     ]
    }
   ],
   "source": [
    "import sys\n",
    "module_dir = os.path.abspath('../../src')  # Gets the absolute path to the src directory\n",
    "sys.path.append(module_dir)\n",
    "from helper_functions import dataset_create_trials\n",
    "\n",
    "overwrite = False\n",
    "if trials_db_downloaded and not overwrite:\n",
    "    print(f'Trials database already downloaded and stored in {trials_path}')\n",
    "    df_trials = pd.read_csv(trials_path)\n",
    "else:\n",
    "    df_trials, trials_path = dataset_create_trials(status='recruiting')\n",
    "    trials_db_downloaded = True\n",
    "\n",
    "# print the ctriteria column value froom 4 random rows of the dataferam. use a for loop\n",
    "df_samples = df_trials.sample(4).reset_index(drop=True)\n",
    "for i in range(2):\n",
    "    print(f'---- Sample {i+1} ----')\n",
    "    print(df_samples.iloc[i]['criteria'])\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Creating Vectore Stores"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Vstore for policy database\n",
    "\n",
    "To create a vector store for the policy database, we will use perform local embedding via NomicEmbeddings.   \n",
    "Also, to treat each policy separately, we split the policies main document into per policy chunks.   \n",
    "Note that this is a problem specific strategy, and may not be appropriate for all RAG applications."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "12\n",
      "Vectore store loaded from existing chromadb\n",
      "12\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "Document(page_content='\\n#### 1. **General Eligibility Criteria**\\n   - **Age Restrictions**: Patients must be between 18 and 75 years old.\\n   - **Gender**: No restrictions unless specified by the trial protocol.\\n   - **Informed Consent**: All patients must provide informed consent prior to participation.\\n   - **Residency**: Patients must be residents of the country where the trial is conducted.\\n   - **Health Insurance**: Patients must have valid health insurance coverage.\\n', metadata={'source': '../../source_data/instut_trials_policy.md'})"
      ]
     },
     "execution_count": 30,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import re\n",
    "\n",
    "from langchain.text_splitter import RecursiveCharacterTextSplitter\n",
    "from langchain_community.document_loaders import WebBaseLoader\n",
    "from langchain_community.vectorstores import Chroma\n",
    "from langchain_nomic.embeddings import NomicEmbeddings\n",
    "from langchain_community.document_loaders import UnstructuredMarkdownLoader\n",
    "from langchain_core.documents import Document\n",
    "import chromadb\n",
    "\n",
    "# vstore_delete = False\n",
    "vstore_load = True\n",
    "markdown_path = \"../../source_data/instut_trials_policy.md\"\n",
    "\n",
    "with open(markdown_path, \"r\") as f:\n",
    "    policy_text = f.read()\n",
    "\n",
    "\n",
    "doc_splits = [\n",
    "    Document(page_content= txt, metadata={\"source\": markdown_path}) for txt in re.split(r\"(?=\\n###)\", policy_text)\n",
    "    ]\n",
    "\n",
    "# the title of markdown is not required in the split\n",
    "doc_splits = doc_splits[1:]    \n",
    "\n",
    "print(len(doc_splits))\n",
    "doc_splits\n",
    "\n",
    "persistent_client = chromadb.PersistentClient(path = chromadb_path)\n",
    "\n",
    "if vstore_load == False:\n",
    "    vectorstore.delete_collection()\n",
    "\n",
    "\n",
    "# if chromadb_path exists\n",
    "if os.path.exists(chromadb_path) and vstore_load:\n",
    "    vectorstore = Chroma(\n",
    "        client=persistent_client,\n",
    "        collection_name=policy_collection,\n",
    "        embedding_function=NomicEmbeddings(model=\"nomic-embed-text-v1.5\",\n",
    "                                inference_mode='local'),\n",
    "    )\n",
    "    print('Vectore store loaded from existing chromadb')\n",
    "    # vectorstore._collection.count()\n",
    "else:\n",
    "    vectorstore = Chroma.from_documents(\n",
    "        documents=doc_splits,\n",
    "        client=persistent_client,\n",
    "        collection_name=policy_collection,\n",
    "        # persist_directory=chromadb_path,\n",
    "        embedding=NomicEmbeddings(model=\"nomic-embed-text-v1.5\",\n",
    "                                inference_mode='local'),\n",
    "    )\n",
    "    print('Vectore store created and stored in persistent chromadb client')\n",
    "print(vectorstore._collection.count())\n",
    "doc_splits[0]\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Vstore for trials"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "For simplicy, we just work on 3 main categories of cancer, lukemia, and mental health problems.   \n",
    "This selection gives us a managable set of trials as a demo application.   \n",
    "However, for a more proper application, categories can be extended with the help of domain specialists.\n",
    "\n",
    "The vectore store is created using chromadb and as a persistent client."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'nctid': 'NCT00715611', 'status': 'recruiting', 'diseases': \"['mesothelioma']\", 'disease_category': 'cancer', 'drugs': \"['pemetrexed and cisplatin or carboplatin']\"}\n",
      "Loading the vectorstore from persistent client\n",
      "18\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "Document(page_content='\\n        Inclusion Criteria:\\r\\n\\r\\n          -  Provide written informed consent to participate on the study\\r\\n\\r\\n          -  Patients must have a pathologically confirmed diagnosis either at MSKCC or at the\\r\\n             participating site of stage I-III malignant pleural mesothelioma\\r\\n\\r\\n          -  Epithelioid or biphasic histology subtype (Note: patients with biphasic histology can\\r\\n             have < 10% sarcomatoid)\\r\\n\\r\\n          -  No evidence of metastatic disease.\\r\\n\\r\\n          -  Patient age ≥ 18 years but ≤ 80 years at the time of consent.\\r\\n\\r\\n          -  Karnofsky performance status ≥ 80%\\r\\n\\r\\n          -  Pulmonary Function Tests:\\r\\n\\r\\n               1. For all patients: DLCO > 40% predicted (corrected for Hgb)\\r\\n\\r\\n               2. For patients enrolled post-P/D, only: FEV1 >/= 35% (corrected for Hgb) (Note:\\r\\n                  patients enrolled prior to P/D will have PFTs repeated pre-IMRT. If this criteria\\r\\n                  is not met, they will be removed from study)\\r\\n\\r\\n          -  In cases of concern about decreased renal function and potential high radiation dose\\r\\n             to the kidneys, an optional nuclear medicine kidney function scan may be performed\\r\\n             prior to radiation therapy to determine the functional contribution of each kidney.\\r\\n\\r\\n          -  Patient enrolled prior to chemotherapy must have adequate organ function as indicated\\r\\n             by the following laboratory values:\\r\\n\\r\\n               1. Absolute neutrophil count ≥1.5 K/mcL\\r\\n\\r\\n               2. Platelets ≥100 K/mcL\\r\\n\\r\\n               3. Serum total bilirubin ≤ 1.5 X ULN\\r\\n\\r\\n               4. AST (SGOT) or ALT (SGPT) ≤ 3.0 X ULN Note: patients enrolled after chemotherapy\\r\\n                  do not have to meet the above criteria\\r\\n\\r\\n        Exclusion Criteria:\\r\\n\\r\\n          -  > 10% Sarcomatoid or desmoplastic histology\\r\\n\\r\\n          -  Continuous oxygen use\\r\\n\\r\\n          -  Prior nephrectomy on the contralateral side of MPM\\r\\n\\r\\n          -  Prior intrapleural therapy (except pleurodesis) or intrapleural therapy at the time of\\r\\n             P/D (i.e.: intrapleural chemotherapy, photodynamic therapy, intrapleural betadine)\\r\\n\\r\\n          -  Prior thoracic radiation therapy preventing hemithoracic pleural IMRT\\r\\n\\r\\n          -  Bulky disease in the fissure preventing lung-sparing pleural IMRT\\r\\n\\r\\n          -  Patients undergoing extrapleural pneumonectomy\\r\\n\\r\\n          -  Patients with an active infection that require systemic antibiotics, antiviral, or\\r\\n             antifungal treatments\\r\\n\\r\\n          -  Patients with a concurrent active malignancy (except squamous or basal cell carcinoma\\r\\n             of the skin)\\r\\n\\r\\n          -  Patients with serious unstable medical illness\\r\\n\\r\\n          -  Presence of third space fluid that cannot be controlled by drainage\\r\\n\\r\\n          -  For patients who develop or have baseline clinically significant pleural effusions\\r\\n             before or during initiation of pemetrexed therapy; consideration should be given to\\r\\n             drain the effusion prior to chemotherapy administration\\r\\n\\r\\n          -  No acute congestive heart failure\\r\\n\\r\\n          -  Pregnant or lactating women\\r\\n\\r\\n          -  Men or women not using effective contraception\\r\\n\\r\\n        Reproductive risks Patients should not become pregnant or father a baby while on this study\\r\\n        because the drugs in this study can affect an unborn baby. Women should not breast-feed a\\r\\n        baby while on this study. Women of childbearing age will be counseled to use birth control\\r\\n        while on this study.\\r\\n      ', metadata={'nctid': 'NCT00715611', 'status': 'recruiting', 'diseases': \"['mesothelioma']\", 'disease_category': 'cancer', 'drugs': \"['pemetrexed and cisplatin or carboplatin']\"})"
      ]
     },
     "execution_count": 31,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from helper_functions import disease_map\n",
    "\n",
    "vstore_delete = False\n",
    "\n",
    "trial_docs = []\n",
    "for i, row in df_trials.iterrows():\n",
    "    disease = disease_map(row['diseases'])\n",
    "    if disease == 'other_conditions':\n",
    "        continue\n",
    "    doc = Document(\n",
    "        page_content=row['criteria'],\n",
    "        metadata={\n",
    "            \"nctid\": row['nctid'],\n",
    "            \"status\": row['status'],\n",
    "            # \"why_stop\": row['why_stop'],\n",
    "            # \"label\": row['label'],\n",
    "            # \"phase\": row['phase'],\n",
    "            \"diseases\": str(row['diseases']),\n",
    "            \"disease_category\": disease[0],\n",
    "            \"drugs\": row['drugs'],            \n",
    "        }\n",
    "    )\n",
    "    trial_docs.append(doc)\n",
    "# trial_docs\n",
    "\n",
    "print(trial_docs[0].metadata)\n",
    "\n",
    "list_remove = set()\n",
    "for i, doc in enumerate(trial_docs):\n",
    "    if len(doc.page_content)>10000:\n",
    "        list_remove.add(i)\n",
    "        # print(doc.metadata)\n",
    "    if doc.metadata['disease_category'] == 'other_conditions':\n",
    "        list_remove.add(i)\n",
    "        # print(doc.metadata)\n",
    "# remove list_remove indexes from trial_docs\n",
    "trial_docs = [doc for i, doc in enumerate(trial_docs) if i not in list_remove]\n",
    "\n",
    "# print(len(trial_docs[11].page_content))\n",
    "# trial_docs.pop(11)\n",
    "\n",
    "persistent_client = chromadb.PersistentClient(path = chromadb_path)\n",
    "\n",
    "if vstore_delete == True:\n",
    "    vectorstore_trials.delete_collection()\n",
    "    trials_vstore_created = False\n",
    "    vstore_delete = False\n",
    "    print(\"vstore deleted\")\n",
    "\n",
    "vectorstore_trials = Chroma(\n",
    "    client=persistent_client,\n",
    "    collection_name=trial_collection,\n",
    "    embedding_function=NomicEmbeddings(model=\"nomic-embed-text-v1.5\",\n",
    "                            inference_mode='local'),\n",
    ")\n",
    "if vectorstore_trials._collection.count() == 0:\n",
    "    vectorstore_trials = Chroma.from_documents(\n",
    "        documents=trial_docs,\n",
    "        client=persistent_client,\n",
    "        collection_name=trial_collection,\n",
    "        # persist_directory=chromadb_path,\n",
    "        embedding=NomicEmbeddings(model=\"nomic-embed-text-v1.5\",\n",
    "                                inference_mode='local'),\n",
    "    )\n",
    "    trials_vstore_created = True\n",
    "    print(\"Vectorstore is created now\")\n",
    "else:\n",
    "    print(\"Loading the vectorstore from persistent client\")\n",
    "\n",
    "print(vectorstore_trials._collection.count())\n",
    "trial_docs[0]\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Tools:\n",
    "Here we define some tools that the agent needs to use for evaluation of the patient, general policies, and clinic trials.\n",
    "Below is the list of tools:\n",
    "- Patient Data Extractor\n",
    "- Profile Maker\n",
    "- Policy Retriever\n",
    "- Policies to Questions\n",
    "- Policy Evaluator with Python Tools (multiply, date_today, date_difference, date_convert, date_split, number_compare)\n",
    "- Trials Retriever \n",
    "- Trial Grader\n",
    "- Hallucination Grader\n",
    "- Profile Rewriter"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Patient Data Extractor\n",
    "This tool extracts patient data from the sqlite database using the patient_id."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain_core.tools import tool\n",
    "from langchain_openai import ChatOpenAI\n",
    "\n",
    "def get_patient_data(patient_id: int) -> dict:\n",
    "    \"\"\"Fetch all fields for the patient based on the given patient_id as an integer.\n",
    "\n",
    "    Returns:\n",
    "        A dictionary containing the patient's medical history.        \n",
    "    \"\"\"    \n",
    "\n",
    "    conn = sqlite3.connect(db)\n",
    "    cursor = conn.cursor()    \n",
    "    query = 'SELECT * FROM patients WHERE patient_id=?'\n",
    "    cursor.execute(query, (patient_id,))\n",
    "    patient_data = cursor.fetchone()\n",
    "    # rows = cursor.fetchall()\n",
    "    column_names = [column[0] for column in cursor.description]\n",
    "    conn.close()\n",
    "    if patient_data is None:\n",
    "        return None\n",
    "    else:    \n",
    "        results = dict(zip(column_names, patient_data))    \n",
    "    return results\n",
    "\n",
    "# Also possible to add a new patient\n",
    "def add_patient_data(patient_data: dict):    \n",
    "    \"\"\"Adds a new patient to the SQLite database.\"\"\"\n",
    "    \n",
    "    name = patient_data['name']\n",
    "    age = patient_data['age']\n",
    "    medical_history = patient_data['medical_history']\n",
    "    previous_trials = patient_data['previous_trials']\n",
    "    trial_status = patient_data['trial_status']\n",
    "    last_trial_dates = patient_data['last_trial_dates']\n",
    "\n",
    "    conn = sqlite3.connect(db)\n",
    "    cursor = conn.cursor()\n",
    "    \n",
    "    # Insert the new patient data into the database\n",
    "    cursor.execute('''\n",
    "    INSERT INTO patients (name, age, medical_history, previous_trials, trial_status, last_trial_dates)\n",
    "    VALUES (?, ?, ?, ?, ?, ?)\n",
    "    ''', (name, age, medical_history, previous_trials, trial_status, last_trial_dates))\n",
    "    \n",
    "    conn.commit()\n",
    "    conn.close()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'patient_id': 1,\n",
       " 'name': 'Alice Johnson',\n",
       " 'age': 66,\n",
       " 'medical_history': 'depressive disorder',\n",
       " 'previous_trials': 'NCT02483538',\n",
       " 'trial_status': 'Withdrawn',\n",
       " 'trial_completion_date': '2023-05-29'}"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sample_patient_id = 1\n",
    "patient_data = get_patient_data(sample_patient_id)\n",
    "patient_data"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Profile Maker\n",
    "This tool creates a a text description of the patient data as the parient profile.   \n",
    "It also suggest possible relevant medical trial categories based on the patient's disease history. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain_core.output_parsers import StrOutputParser\n",
    "from langchain.prompts import PromptTemplate\n",
    "\n",
    "parser = StrOutputParser()\n",
    "prompt_profile = PromptTemplate(\n",
    "    template=\"\"\"\n",
    "    You are the Clinical Research Coordinator in the screening phase of a clinical trial. \n",
    "    Use the following patient data to write the patient profile for the screening phase.\n",
    "    The patient profile is a summary of the patient's information in continuous text form.    \n",
    "    If they had no previous trial participation, exclude trial status and trial completion date.\\n\n",
    "    Do not ignore any available information.\\n \n",
    "    Also suggest medical trials that can be related to patient's disease history.\\n    \n",
    "    Write the patient profile in 3 to 4 short sentences.\\n\\n\n",
    "    {patient_data}\"\"\",\n",
    "    input_variables=[\"patient_data\"],\n",
    ")\n",
    "# In a separate section of the profile, also suggest medical trial categories related to patient's disease history.\\n    \n",
    "# model = ChatOpenAI(temperature = 0.0, model=modelID)\n",
    "\n",
    "chain_profile = prompt_profile | model | parser\n",
    "# to preserve the patient's private information"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "('Patient is a 66-year-old individual with a history of depressive disorder. '\n",
      " 'They have previously participated in the clinical trial NCT02483538, but '\n",
      " 'withdrew before completion on May 29, 2023. Given their medical history, the '\n",
      " 'patient may be suitable for clinical trials focusing on treatments for '\n",
      " 'depressive disorders or mental health interventions.')\n"
     ]
    }
   ],
   "source": [
    "# to preserve the patient's private information\n",
    "if patient_data.get('name'):\n",
    "    del patient_data['patient_id']\n",
    "    del patient_data['name']\n",
    "\n",
    "patient_profile = chain_profile.invoke({'patient_data': patient_data})\n",
    "nprint(patient_profile) # patient_profile"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Policy Retriever\n",
    "A retriever tool that retrieves relevant policies from policy documents based on the patient profile."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "5\n",
      "('\\n'\n",
      " '#### 3. **Previous Trial Participation**   \\n'\n",
      " '   - **Trial Frequency**: Patients should not have participated in more than '\n",
      " 'two clinical trials in the past year excluding this trial to prevent '\n",
      " 'overexposure to experimental treatments.\\n'\n",
      " '   - **Active Trials**: Patients currently enrolled in another active trial '\n",
      " 'are not eligible until they have completed or withdrawn from the ongoing '\n",
      " 'trial.\\n'\n",
      " '   - **Recent Trials**: Patients cannot have participated in any other '\n",
      " 'clinical trial within the past 5 months from now.\\n'\n",
      " '   - **Trial Outcomes**:\\n'\n",
      " '     - Patients who were withdrawn from previous trials due to adverse '\n",
      " 'effects are not eligible.\\n'\n",
      " '     - Patients who have completed a trial must have a documented recovery '\n",
      " 'period before participating in a new trial.\\n')\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[Document(page_content='\\n#### 3. **Previous Trial Participation**   \\n   - **Trial Frequency**: Patients should not have participated in more than two clinical trials in the past year excluding this trial to prevent overexposure to experimental treatments.\\n   - **Active Trials**: Patients currently enrolled in another active trial are not eligible until they have completed or withdrawn from the ongoing trial.\\n   - **Recent Trials**: Patients cannot have participated in any other clinical trial within the past 5 months from now.\\n   - **Trial Outcomes**:\\n     - Patients who were withdrawn from previous trials due to adverse effects are not eligible.\\n     - Patients who have completed a trial must have a documented recovery period before participating in a new trial.\\n', metadata={'source': '../../source_data/instut_trials_policy.md'}),\n",
       " Document(page_content='\\n#### 1. **General Eligibility Criteria**\\n   - **Age Restrictions**: Patients must be between 18 and 75 years old.\\n   - **Gender**: No restrictions unless specified by the trial protocol.\\n   - **Informed Consent**: All patients must provide informed consent prior to participation.\\n   - **Residency**: Patients must be residents of the country where the trial is conducted.\\n   - **Health Insurance**: Patients must have valid health insurance coverage.\\n', metadata={'source': '../../source_data/instut_trials_policy.md'}),\n",
       " Document(page_content='\\n#### 4. **Medication and Treatment Restrictions**\\n   - **Concurrent Medications**: Patients must not be on any medication that could interfere with the study drug, unless specified by the trial protocol.\\n   - **Substance Abuse**: Patients with a history of substance abuse within the past 12 months are excluded.\\n   - **Vaccinations**: Patients must not have received live vaccines within 30 days prior to the start of the trial.\\n', metadata={'source': '../../source_data/instut_trials_policy.md'}),\n",
       " Document(page_content='\\n#### 2. **Medical History Restrictions**\\n   - **Severe Heart Disease** (e.g., recent myocardial infarction, heart failure)\\n   - **Uncontrolled Diabetes Mellitus**     \\n   - **Recent Surgeries**: Patients who have undergone major surgery within the current year are not eligible.   \\n   - **Infectious Diseases**: Patients with active tuberculosis, hepatitis B or C, or HIV are excluded unless the trial is specifically designed for these conditions.\\n', metadata={'source': '../../source_data/instut_trials_policy.md'}),\n",
       " Document(page_content='\\n#### 6. **Special Considerations**\\n   - **Pregnancy and Nursing**: Pregnant or nursing women are excluded unless the trial is specifically designed for this population.\\n   - **Genetic Testing**: Trials involving genetic testing must have explicit consent from patients for genetic data collection and analysis.\\n   - **Patient Monitoring**: Patients must agree to regular monitoring and follow-up visits as specified by the trial protocol.\\n   - **BMI Restrictions**: Patients with a BMI outside the range of 18.5-35 may be excluded unless otherwise specified.\\n   - **Comorbid Conditions**: Patients with multiple comorbid conditions that may interfere with the study outcomes are excluded.\\n', metadata={'source': '../../source_data/instut_trials_policy.md'})]"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "retriever_policy = vectorstore.as_retriever(search_kwargs={\"k\": n_retrieved_policies})\n",
    "\n",
    "# sample policy retrieval\n",
    "docs_retrieved = retriever_policy.get_relevant_documents(patient_profile)\n",
    "print(len(docs_retrieved))\n",
    "doc_txt = docs_retrieved[0].page_content\n",
    "nprint(doc_txt)\n",
    "docs_retrieved"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Policies to Questions\n",
    "Each main policy document also consists of a list of sub-policies related to the same category. For example, a policy may contain eligible and non-eligible sub-policies for a give patient. This is a challenging condition for an llm to determine the patient's eligibility in a single shot attempt.   \n",
    "To assist the llm and prevent possible mistakes, we convert each subpolicy to a direct yes/no question with a yes answer indicating the patient is not eligible.   \n",
    "We define an llm chain for that purpose."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Policies --> questions\n",
    "from langchain_core.pydantic_v1 import BaseModel, Field\n",
    "\n",
    "prompt_rps = PromptTemplate(\n",
    "    template=\"\"\" You are a Principal Investigator (PI) for clinical trials. \n",
    "        The following document contains a policy document about participation in clinial trials:\n",
    "        \\n\\n{policy}\\n\\b\n",
    "\n",
    "        Your task is to rephrase each single policy from the document above into a single yes/no question. \n",
    "        Form each question so that a yes answer indicates the patient's ineligibility.\n",
    "        Do not create more questions than given number of policies.\n",
    "\n",
    "        Example: Patients who have had accidents in the past 10 months are not eligible\n",
    "        rephrased: Did patient have an accident in the past 10 months?\n",
    "\n",
    "        Example: Patients with active tuberculosis, hepatitis B or C, or HIV are excluded unless the trial is specifically designed for these conditions.\n",
    "        rephrased: Did patient have active tuberculosis, hepatitis B or C, or HIV?\n",
    "        \"\"\",\n",
    "    input_variables=[\"policy\"],\n",
    ")\n",
    "\n",
    "class policy_relevance(BaseModel):\n",
    "    \"policy relevance score\"\n",
    "\n",
    "    relevant: str = Field(description=\"is policy relevant? 'yes' or 'no'.\")\n",
    "    reason: str\n",
    "\n",
    "parser = StrOutputParser()\n",
    "\n",
    "policy_rps_chain = prompt_rps | model | parser\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "('1. Has the patient participated in more than two clinical trials in the past '\n",
      " 'year excluding this trial?\\n'\n",
      " '2. Is the patient currently enrolled in another active trial?\\n'\n",
      " '3. Has the patient participated in any other clinical trial within the past '\n",
      " '5 months from now?\\n'\n",
      " '4. Was the patient withdrawn from previous trials due to adverse effects?\\n'\n",
      " '5. Has the patient completed a trial without a documented recovery period '\n",
      " 'before participating in a new trial?')\n"
     ]
    }
   ],
   "source": [
    "\n",
    "policy = docs_retrieved[0].page_content\n",
    "policy_qs = policy_rps_chain.invoke({\"policy\": policy})\n",
    "nprint(policy_qs)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Policy Evaluator with Python Tools\n",
    "\n",
    "The policy evaluator uses Python tools to evaluate the patient's eligibility.\n",
    "More often the patient's profiles or the policy documents contain date or numeric information. However, llms are not able to process these types of information properly. Therefore, we use Python tools for the model to perform numerical or date calculations and convertions.\n",
    "\n",
    "Note: You can also try the python REPL tool to give LLM more flexibility and freedon to do all kinds of calculations. However, this tool needs fine-tuned guardrails and also clear instructions on how to use it. Otherwise llm may use it even for non-numerical simulations."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain_core.tools import StructuredTool\n",
    "from datetime import date, datetime\n",
    "from langchain_core.prompts import ChatPromptTemplate\n",
    "from langgraph.prebuilt import create_react_agent\n",
    "\n",
    "\n",
    "def policy_tools(policy_qs: str, patient_profile: str):\n",
    "    system_message = \"\"\"\n",
    "    You are a Principal Investigator (PI) for evaluating patients for clinical trials.\n",
    "    You are asked to compare the patient profile document to the institution policy questions.\n",
    "    You must determine if the patient is eligible based on the following documnents.\n",
    "\n",
    "    \\n #### Here is the patient profile document: \\n {patient_profile}\\n\\n\n",
    "\n",
    "    If the answer to any policy question is yes, then the patient is not eligible.\\n\n",
    "    If the answer to the question is not provided in the patient profile, answer 'no'.\\n\n",
    "\n",
    "    Give a binary 'yes' or 'no' score in the response to indicate whether the patient is eligible according to the given policy ONLY.\n",
    "    If the patient is not eligible then also include the reason in your response.\n",
    "\n",
    "    You have access to the following tools:\n",
    "    {tool_names}\n",
    "    \"\"\"\n",
    "\n",
    "    prompt = ChatPromptTemplate.from_messages(\n",
    "        [\n",
    "            (\"system\", system_message),\n",
    "            (\"human\", \"{input}\"),\n",
    "            # Placeholders fill up a **list** of messages\n",
    "            (\"placeholder\", \"{agent_scratchpad}\"),\n",
    "        ]\n",
    "    )\n",
    "\n",
    "    class CalculatorInput(BaseModel):\n",
    "        num1: float = Field(description=\"first number\")\n",
    "        num2: float = Field(description=\"second number\")\n",
    "\n",
    "    def multiply(num1: float, num2: float) -> float:\n",
    "        \"multiplies two input numbers together, num1 and num2\"\n",
    "        return (num1 * num2)\n",
    "\n",
    "    multiply_tool = StructuredTool.from_function(\n",
    "        func=multiply,\n",
    "        name=\"multiply\",\n",
    "        description=\"multiply numbers\",\n",
    "        args_schema=CalculatorInput,\n",
    "        # return_direct=True,    \n",
    "    )\n",
    "\n",
    "    @tool(\"date_today-tool\")\n",
    "    def date_today() -> datetime.date:\n",
    "        \"returns today date\"\n",
    "        return datetime.today().date()    \n",
    "\n",
    "    def date_difference(date1: date, date2: date) -> int:\n",
    "        \"The number of months date1 is before date2\"\n",
    "        # \"How many months before date2 is date1?\"\n",
    "        month_difference = (date2.year - date1.year) * 12 + date2.month - date1.month\n",
    "        return f'{month_difference} months'\n",
    "\n",
    "    class dates(BaseModel):\n",
    "        date1: date = Field(description=\"first date\")\n",
    "        date2: date = Field(description=\"second date\")\n",
    "\n",
    "    date_difference_tool = StructuredTool.from_function(\n",
    "        func=date_difference,\n",
    "        name=\"date_difference\",\n",
    "        description=\"The number of months first date is before second date\",\n",
    "        args_schema=dates,\n",
    "        # return_direct=True,    \n",
    "    )\n",
    "\n",
    "    class date_class(BaseModel):\n",
    "        date: str = Field(description=\"A date string in the format YYYY-MM-DD\")    \n",
    "\n",
    "    @tool(\"date_convert-tool\", args_schema=date_class)\n",
    "    def date_convert(date: str) -> date:\n",
    "        \"Converts a date string to a date object\"\n",
    "        # \"How many months before date2 is date1?\"\n",
    "        date = datetime.strptime(date, \"%Y-%m-%d\").date()\n",
    "        return date\n",
    "\n",
    "    @tool(\"date_split-tool\", args_schema=date_class)\n",
    "    def date_split(date: str) -> date:\n",
    "        \"Extracts the year and month from a date string\"\n",
    "        # \"How many months before date2 is date1?\"\n",
    "        date = datetime.strptime(date, \"%Y-%m-%d\").date()\n",
    "        year = date.year\n",
    "        month = date.month\n",
    "        return f'year: {year}, month: {month}'\n",
    "\n",
    "    @tool(\"number_comparison-tool\", args_schema=CalculatorInput)\n",
    "    def number_compare(num1: float, num2: float) -> bool:\n",
    "        \"Determines if first number is less than the second number\"\n",
    "        # \"How many months before date2 is date1?\"\n",
    "        num1_less_num1 = num1 < num2\n",
    "        return num1_less_num1\n",
    "\n",
    "    tools = [multiply_tool, date_today, date_difference_tool, date_split, number_compare]\n",
    "\n",
    "    tool_names=\", \".join([tool.name for tool in tools])\n",
    "    prompt = prompt.partial(tool_names=tool_names)\n",
    "    prompt = prompt.partial(patient_profile=patient_profile)\n",
    "\n",
    "    model_agent = ChatOpenAI(temperature=0, model=\"gpt-4o\")\n",
    "    react_agent = create_react_agent(model_agent, tools, prompt, debug=False)\n",
    "\n",
    "    message = f\"\"\" Here are the policy questions: \\n{policy_qs}\"\"\"\n",
    "    messages = react_agent.invoke({\"messages\": [(\"human\", message)]})\n",
    "    result = messages[\"messages\"][-1].content    \n",
    "    return result"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# sample policy evaluation\n",
    "import langchain\n",
    "langchain.debug = True\n",
    "policy = docs_retrieved[0].page_content\n",
    "# nprint(policy)\n",
    "result = policy_tools(policy, patient_profile)\n",
    "nprint(result)\n",
    "langchain.debug = False"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Trials Retriever \n",
    "\n",
    "We use SelfQueryRetriever module that also uses metadata information. This retrival can quickly find relevant trials based on metadata such as disease category, drug names, etc.   \n",
    "The value of such retrival becomes important for real applications, where patient profiles contain more comphrehensive medical histories."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain.chains.query_constructor.base import AttributeInfo\n",
    "from langchain.retrievers.self_query.base import SelfQueryRetriever\n",
    "\n",
    "metadata_field_info = [\n",
    "    AttributeInfo(\n",
    "        name=\"disease_category\",\n",
    "        description=\"Defines the disease group of patients related to this trial. One of ['cancer', 'leukemia', 'mental_health']\",\n",
    "        # description=\"The trial is for patients when their disease is related to this category. One of ['cancer', 'leukemia', 'mental_health']\",\n",
    "        type=\"string\",\n",
    "    ),\n",
    "    AttributeInfo(\n",
    "        name=\"drugs\",\n",
    "        description=\"List of drug names used in the trial\",\n",
    "        type=\"str\",\n",
    "    ),    \n",
    "]\n",
    "\n",
    "document_content_description = \"The list of patient conditions to include or exclude them from the trial\"\n",
    "retriever_trial_sq = SelfQueryRetriever.from_llm(\n",
    "    model,\n",
    "    vectorstore_trials,\n",
    "    # vectorstore_trials_mpnet,\n",
    "    document_content_description,\n",
    "    metadata_field_info\n",
    "    # enable_limit=True\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'disease_category': 'mental_health', 'diseases': \"['suicidal ideation', 'major depressive disorder', 'healthy participants']\", 'drugs': \"['ketamine', 'saline']\", 'nctid': 'NCT02037503', 'status': 'recruiting'}\n",
      "\n",
      "        Inclusion Criteria:\n",
      "\n",
      "        Suicidal Ideation group:\n",
      "\n",
      "          1. Any person admitted to the emergency room department after a suicide attempt , defined\n",
      "             as requiring medical intervention - not just a psychiatrist ( surgical or\n",
      "             pharmacological treatment but also the need for observation ) .\n",
      "\n",
      "          2. The need for medical intervention will be defined by the ER ED physician\n",
      "\n",
      "          3. Ages 18-65\n",
      "\n",
      "        For the depression group:\n",
      "\n",
      "          1. Diagnosed with major depression according to DSM VI.\n",
      "\n",
      "          2. Ongoing depression (according to DSM criteria) despite treatment with at least two\n",
      "             antidepressants in adequate dosages and for longer than three weeks.\n",
      "\n",
      "          3. Ages 18-65\n",
      "\n",
      "        For the romantic relationship breakup:\n",
      "\n",
      "        1. Participants that have experienced a meaningful romantic relationship break-up within\n",
      "        the past 12 months\n",
      "\n",
      "        -\n",
      "\n",
      "        Exclusion Criteria for all groups:\n",
      "\n",
      "          1. Psychotic state instate in the examination\n",
      "\n",
      "          2. Diagnosis of schizophrenia / schizoaffective disorder\n",
      "\n",
      "          3. Drug or alcohol abuse as is revealed in by blood/urine tests\n",
      "\n",
      "          4. Patient in which, according to the examiner, there is primary or secondary gain.\n",
      "\n",
      "          5. Patient , which, at the time of his admission , is without any pharmacological\n",
      "             treatment.\n",
      "      \n"
     ]
    }
   ],
   "source": [
    "# sample trial retrieval\n",
    "question = f\"\"\"\n",
    "Which trials are relevant to the patient with the following medical history?\\n\n",
    "patient_profile: {patient_profile}\n",
    "\"\"\"        \n",
    "docs_retrieved = retriever_trial_sq.get_relevant_documents(question)\n",
    "print(docs_retrieved[0].metadata)\n",
    "print(docs_retrieved[0].page_content)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Trial Grader\n",
    "\n",
    "Each retrieved trial is checked for relevance to the patient's medical profile. The final output would be in the form of a binary relevance score, the explanation behind the score, and further information that is needed from the patient's medical history."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "class grade(BaseModel):\n",
    "        \"\"\"The result of the trial's relevance check as relevance score and explanation.\"\"\"\n",
    "\n",
    "        relevance_score: str        \n",
    "        explanation: str = Field(description=\"Reasons to the given relevance score.\")        \n",
    "        further_information: str\n",
    "\n",
    "prompt_grader = PromptTemplate(\n",
    "        template=\"\"\" \n",
    "        You are a Principal Investigator (PI) for evaluating patients for clinical trials.\\n\n",
    "        Your task is to evaluate the relevance of a clinical trial to the given patient's medical profile. \\n\n",
    "        \n",
    "        \n",
    "        The clinical trial is related to these diseases: {trial_diseases} \\n\n",
    "        Here are the inclusion and exclusion criteria of the trial: \\n\\n {document} \\n\\n\n",
    "        \n",
    "        ===============                \n",
    "        Use the following steps to determine relevance and provide the necessary fields in your response: \\n\n",
    "        1- If the patient's profile meets any exclusion criteria, then the trial is not relevant --> relevance_score = 'No'. \\n\n",
    "        2- If the patient has or had the trial's inclusion diseases, then it is relevant --> relevance_score = 'Yes'.\\n        \n",
    "        3- If the patient did not have the trial's inclusion diseases, then it is not relevant --> relevance_score = 'No'.\\n\n",
    "                       \n",
    "        Example 1: \n",
    "The patient has Arthritis and the trial is related to pancreatic cancer. --> relevance_score = 'No' \\n\n",
    "        \n",
    "        Example 2: \n",
    "The patient has pancreatic cancer and the trial is also related to carcinoma pancreatic cancer. --> relevance_score = 'Yes' \\n\n",
    "\n",
    "        Example 3: \n",
    "The patient has pancreatic cancer and the trial is related to breast cancer or ovarian cancer. --> relevance_score = 'No'. \\n \n",
    "\n",
    "        Bring your justification in the explanation. \\n\n",
    "\n",
    "        Mention further information that is needed from the patient's medical history related to the trial's criteria \\n\n",
    "\n",
    "        ===============\n",
    "        Here is the patient's medical profile: {patient_profile} \\n\\n\n",
    "        \"\"\",\n",
    "        input_variables=[\"document\", \"patient_profile\", \"trial_diseases\"],\n",
    "    )\n",
    "\n",
    "llm_with_tool = model.with_structured_output(grade)\n",
    "retrieval_grader = prompt_grader | llm_with_tool"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "('relevance_score', 'No')\n",
      "( 'explanation',\n",
      "  'The patient is 66 years old, which exceeds the age limit of 65 specified in '\n",
      "  'the inclusion criteria for both the suicidal ideation and depression '\n",
      "  \"groups. Additionally, the patient's medical history mentions a history of \"\n",
      "  'depressive disorder, which could potentially exclude them from the trial if '\n",
      "  'they have already been diagnosed and treated for major depression according '\n",
      "  'to DSM VI criteria.')\n",
      "( 'further_information',\n",
      "  \"Further information needed: Confirmation of the patient's current mental \"\n",
      "  'health status and whether they meet the specific criteria for major '\n",
      "  'depression according to DSM VI.')\n"
     ]
    }
   ],
   "source": [
    "doc = docs_retrieved[0]\n",
    "doc_txt = doc.page_content\n",
    "trial_diseases = doc.metadata.get(\"diseases\")\n",
    "trial_score = retrieval_grader.invoke({\"patient_profile\": patient_profile, \"document\": doc_txt, \"trial_diseases\": trial_diseases})\n",
    "\n",
    "for score in trial_score:\n",
    "    nprint(score)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Hallucination Grader\n",
    "When agent cross-matches the patient's medical profile, it sometimes relates the input information (patient profile or trial data) to its trained knowledge. Or in some cases it makes non-existing assumptions about the patient's medical history.\n",
    "   \n",
    "These cases, known as llm hallucinations, can be prevented by cross-matching the input information with the model's completion. We created a guardrail using this strategy.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "No\n",
      "('The patient is 66 years old, which exceeds the age limit of 65 specified in '\n",
      " 'the inclusion criteria for both the suicidal ideation and depression groups. '\n",
      " \"Additionally, the patient's medical history mentions a history of depressive \"\n",
      " 'disorder, which could potentially exclude them from the trial if they have '\n",
      " 'already been diagnosed and treated for major depression according to DSM VI '\n",
      " 'criteria.')\n",
      "yes\n",
      "(\"The LLM generation is grounded in the patient's medical profile by \"\n",
      " \"accurately referencing the patient's age, history of depressive disorder, \"\n",
      " 'and previous participation in a clinical trial.')\n"
     ]
    }
   ],
   "source": [
    "class GradeHallucinations(BaseModel):\n",
    "    \"\"\"Binary score and explanation for whether the LLM's generated answer is grounded in / supported by the facts in the patient's medical profile.\"\"\"\n",
    "    # \"\"\"Binary score for hallucination present in generation answer.\"\"\"\n",
    "\n",
    "\n",
    "    binary_score: str = Field(\n",
    "        description=\"Answer is grounded in the patient's medical profile, 'yes' or 'no'\"\n",
    "    )\n",
    "    Reason: str = Field(description=\"Reasons to the given relevance score.\")\n",
    "\n",
    "\n",
    "# LLM with function call\n",
    "llm_with_tool = model.with_structured_output(GradeHallucinations)\n",
    "\n",
    "# Prompt\n",
    "system = \"\"\"You are a grader assessing whether an LLM generation is grounded in / supported by the facts in the patient's medical profile. \\n \n",
    "     Give a binary score 'yes' or 'no'. 'Yes' means that the answer is grounded in / supported by the facts in the patient's medical profile.\"\"\"\n",
    "hallucination_prompt = ChatPromptTemplate.from_messages(\n",
    "    [\n",
    "        (\"system\", system),\n",
    "        (\"human\", \"Patient's medical profile: \\n\\n {patient_profile} \\n\\n LLM generation: {explanation}\"),\n",
    "    ]\n",
    ")\n",
    "\n",
    "hallucination_grader = hallucination_prompt | llm_with_tool\n",
    "print(trial_score.relevance_score)\n",
    "explanation = trial_score.explanation\n",
    "nprint(explanation)\n",
    "# docs = patient_profile\n",
    "score = hallucination_grader.invoke({\"patient_profile\": patient_profile, \"explanation\": explanation})\n",
    "grade = score.binary_score\n",
    "print(grade)\n",
    "nprint(score.Reason)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Profile Rewriter\n",
    "\n",
    "In some cases, which is often the case in this demo, the patient's medical history is not complete or does not directly indicate relevant clinical trials. However, there could still be a connecton between its disease(s) and some of the trial descriptions. Therefore, Insread of manually rewriting the patient's medical history, we can use the LLM to extpand the trial suggetions in their profile. This can also be used as a draft for a clinical specialist to further complete the patient's profile.\n",
    "\n",
    "The Profile Rewriter is an llm based tool that tries to link the patient's medical history to a predefined set of clinical categories."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "\"The patient is a 66-year-old with a medical history of depressive disorder. They have participated in the trial NCT02483538, which was withdrawn before completion on May 29, 2023.\\n\\nSuggested relevant trials:\\nmental_health: Depressive disorder is a mental health condition, and the patient's medical history indicates a potential relevance to trials in this category.\""
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "system = \"\"\"\n",
    "A trial cross match resulted in no trials for the patient.\n",
    "As a clinical specialist write a medical profile for this patient and see if their disease(s) can be relevant to any of these categories of mental_health, cancer, or leukemia.\n",
    "If yes, then suggest relevant medical trial categories for the agent.\n",
    "If no, then do not add anything there.\n",
    "\n",
    "Your output must be as below:\n",
    "<a text summart of original profile>\n",
    "Suggested relevant trials:\n",
    "<bullet points of relevant medical trial categoriyes from the above with a one line reason>\n",
    "\n",
    "Only include categories which can be related to patient diseases in more often cases.\n",
    "Disregard categories which ocasionaly or in some cases can be relevant to patient diseases.\n",
    "\n",
    "example:\n",
    "The patient is a X-year-old with a medical history of Y. They have participated ........  previous trials, and their trial status and completion date .........\n",
    "Suggested relevant trials:\n",
    "category X: [patient's disease] can be related to X due to Y.\n",
    "\"\"\"\n",
    "re_write_prompt = ChatPromptTemplate.from_messages(\n",
    "    [\n",
    "        (\"system\", system),\n",
    "        (\n",
    "            \"human\",\n",
    "            \"Here is a patient data:\\n\\n {patient_data} \\n write a patient profile.\",\n",
    "        ),\n",
    "    ]\n",
    ")\n",
    "\n",
    "profile_rewriter_chain = re_write_prompt | model | StrOutputParser()\n",
    "profile_rewriter_chain.invoke({\"patient_data\": patient_data})"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Graph\n",
    "To implement the pipeline of extracting and evaluating the patient's information and clinical policies and trials, we use [LangGraph](https://langchain-ai.github.io/langgraph/).   \n",
    "This is a comprehensive tool-chain for creating agent-based pipelines in Langchain. \n",
    "The main two components are:\n",
    "- Agent state\n",
    "- Agent graph\n",
    "  \n",
    "Following their [tutorials](https://langchain-ai.github.io/langgraph/tutorials/), you can create many different agentic designs with Langgraph. Even for the same applications you may create different graph architectures."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## State\n",
    "To help agent (and also us) to keep track of the state of information in the graph, we define an agent state and its relevant values.   \n",
    "These values are passed from one node of the graph to the other or will be modified during iterations."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "from typing import Annotated, List\n",
    "from typing_extensions import TypedDict\n",
    "from langgraph.graph.message import AnyMessage, add_messages\n",
    "from langchain_core.messages import AIMessage, SystemMessage, HumanMessage, ToolMessage\n",
    "from langchain_core.documents import Document\n",
    "\n",
    "class AgentState(TypedDict):\n",
    "    # messages: Annotated[Sequence[BaseMessage], operator.add]\n",
    "    last_node: str\n",
    "    patient_prompt: str\n",
    "    patient_id: int\n",
    "    patient_data: dict\n",
    "    patient_profile: str\n",
    "    policy_eligible: bool\n",
    "    policies: List[Document]\n",
    "    checked_policy: Document\n",
    "    unchecked_policies: List[Document]\n",
    "    policy_qs: str\n",
    "    rejection_reason: str    \n",
    "    revision_number: int\n",
    "    max_revisions: int\n",
    "    trial_searches: int\n",
    "    max_trial_searches: int            \n",
    "    trials: List[Document]\n",
    "    relevant_trials: list[dict]\n",
    "    ask_expert: str"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Agent\n",
    "The main Agent graph is defined by nodes and edges. Nodes are the main steps of the pipeline, and edges define the transitions between the nodes (steps).\n",
    "\n",
    "In our architecture, the steps of the pipeline are most of the time deterministic. We also tried to define nodes of the graph according to the main steps of the clinical evaluation pipeline. This helps you to follow the logics and steps easier. \n",
    "  \n",
    "However, you can also use a single centric assistant node and let the agent to decide which tool must be called given the current state in the process. This may look more professional, but needs more effort to built robustly."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langgraph.checkpoint.sqlite import SqliteSaver\n",
    "from langgraph.prebuilt import ToolNode\n",
    "from langgraph.prebuilt import tools_condition\n",
    "from langchain_core.pydantic_v1 import BaseModel, Field\n",
    "from operator import itemgetter\n",
    "from typing import Literal\n",
    "from langgraph.graph import StateGraph, END\n",
    "\n",
    "\n",
    "model = ChatOpenAI(model=\"gpt-3.5-turbo\")  #reduce inference cost\n",
    "\n",
    "memory = SqliteSaver.from_conn_string(\":memory:\")\n",
    "\n",
    "class Patient_ID(BaseModel):\n",
    "    patient_id: int\n",
    "    # no_id: bool\n",
    "\n",
    "def patient_collector_node(state: AgentState):\n",
    "    patient_data_prompt = \"\"\"You are a helpful assistance in extrcting patient's medical history.\n",
    "Based on the following request identify and return the patient's ID number.\n",
    "\"\"\"\n",
    "\n",
    "    response = model.with_structured_output(Patient_ID).invoke([\n",
    "        SystemMessage(content=patient_data_prompt),\n",
    "        HumanMessage(content=state['patient_prompt'])\n",
    "    ])\n",
    "    patient_id = response.patient_id\n",
    "    # print(response)\n",
    "    print(f\"Patient ID: {patient_id}\")\n",
    "    patient_data = get_patient_data(patient_id)\n",
    "    print(patient_data)\n",
    "    if patient_data is not None:        \n",
    "        if patient_data.get('name'):\n",
    "            del patient_data['patient_id']\n",
    "            del patient_data['name']\n",
    "        patient_profile = chain_profile.invoke({'patient_data': patient_data})\n",
    "\n",
    "    return {\n",
    "        \"last_node\": \"patient_collector\",\n",
    "        \"patient_data\": patient_data,\n",
    "        \"patient_profile\": patient_profile,\n",
    "        \"patient_id\": patient_id,\n",
    "        \"revision_number\": state.get(\"revision_number\", 1) + 1,\n",
    "        'policy_eligible': 'N/A'\n",
    "        }\n",
    "\n",
    "def policy_search(state: AgentState) -> List[Document]:    \n",
    "    # question = f\"\"\"\n",
    "    # Does the following patient profile comply with the institutional policies?\\n\n",
    "    # {patient_profile}\n",
    "    # \"\"\"\n",
    "    docs_retrieved = retriever_policy.get_relevant_documents(patient_profile)    \n",
    "    return {\n",
    "        \"last_node\": \"policy_search\",\n",
    "        \"policies\": docs_retrieved,\n",
    "        \"unchecked_policies\": docs_retrieved.copy(),\n",
    "    }\n",
    "\n",
    "def policy_evaluator(state: AgentState, policy_doc: Document = None):\n",
    "    if policy_doc is None:\n",
    "        policy_doc = state['unchecked_policies'][0]\n",
    "    \n",
    "    policy_header = policy_doc.page_content.split('\\n', 2)[1]\n",
    "    print(f'Evaluating Policy:\\n {policy_header}')\n",
    "    \n",
    "    policy = policy_doc.page_content\n",
    "    \n",
    "    policy_qs = policy_rps_chain.invoke({\"policy\": policy})\n",
    "    \n",
    "    # result = policy_repl(policy_qs, state['patient_profile'])\n",
    "    result = policy_tools(policy_qs, state['patient_profile'])\n",
    "    # nprint(result['output'])\n",
    "\n",
    "    class eligibility(BaseModel):\n",
    "        \"\"\" Give the patient's eligibility result.\"\"\"\n",
    "\n",
    "        eligibility: str = Field(description=\"Patient's eligibility for the clinical trial. 'yes' or 'no'\")\n",
    "        reason: str = Field(description=\"The reason(s) only if the patient is not eligible for clinical trials. Othereise use N/A\")\n",
    "\n",
    "        class Config:\n",
    "            schema_extra = {\n",
    "                \"example\": {\n",
    "                    \"eligibility\": 'yes',\n",
    "                    \"reason\": \"N/A\",\n",
    "                },\n",
    "                \"example 2\": {\n",
    "                    \"eligibility\": 'no',\n",
    "                    \"reason\": \"The patient is pregnant at the moment.\",\n",
    "                },                \n",
    "            }\n",
    "\n",
    "    llm_with_tools = model.bind_tools([eligibility])\n",
    "    message = f\"\"\"Evaluation of the patient's eligibility:\n",
    "    {result}\\n\\n\n",
    "    Is the patient eligible according to this policy?\"\"\"\n",
    "    response = llm_with_tools.invoke(message)\n",
    "    policy_eligible = response.tool_calls[0]['args']['eligibility']             \n",
    "\n",
    "    state[\"policy_eligible\"] = policy_eligible.lower() == 'yes'\n",
    "    state[\"rejection_reason\"] = response.tool_calls[0]['args']['reason']\n",
    "    \n",
    "    # remove the first element of unchecked_policies\n",
    "    state['unchecked_policies'].pop(0)\n",
    "    state[\"revision_number\"] = state.get(\"revision_number\", 1) + 1\n",
    "    state['checked_policy'] = policy_doc\n",
    "    state['policy_qs'] = policy_qs\n",
    "    state['last_node'] = 'policy_evaluator'\n",
    "\n",
    "    return state\n",
    "\n",
    "def trial_search(state: AgentState):\n",
    "    \"\"\"\n",
    "    This node searches the trial database to retrieve a list of clinical trials that match the patient's medical history.\n",
    "    \"\"\"\n",
    "\n",
    "    patient_profile = state['patient_profile']\n",
    "\n",
    "    question = f\"\"\"\n",
    "    Which trials are relevant to the patient with the following medical history?\\n\n",
    "    patient_profile: {patient_profile}\n",
    "    \"\"\"        \n",
    "    docs_retrieved = retriever_trial_sq.get_relevant_documents(question)\n",
    "    # print(state)\n",
    "    trial_searches = state.get('trial_searches') or 0\n",
    "    return {\n",
    "        'last_node': 'trial_search',\n",
    "        'trials': docs_retrieved,\n",
    "        'trial_searches': trial_searches + 1,\n",
    "        }\n",
    "\n",
    "def grade_trials(state: AgentState, trials: List[Document] = None):\n",
    "    \"\"\"\n",
    "    Determines whether the retrieved documents are relevant to the question.\n",
    "\n",
    "    Args:\n",
    "        state (dict): The current graph state\n",
    "\n",
    "    Returns:\n",
    "        state (dict): Updates documents key with only filtered relevant documents\n",
    "    \"\"\"\n",
    "\n",
    "    print(\"----- CHECKING THE TRIALS RELEVANCE TO PATIENT PROFILE ----- \")\n",
    "    \n",
    "    if trials is None:\n",
    "        trials = state['trials']\n",
    "    patient_profile = state['patient_profile']\n",
    "\n",
    "    # Score each doc\n",
    "    relevant_trials = []\n",
    "    for trial in trials:\n",
    "        doc_txt = trial.page_content\n",
    "        trial_diseases = trial.metadata['diseases']\n",
    "        nctid = trial.metadata['nctid']\n",
    "        print(f\"---GRADER: TRIAL {nctid}: ---\") \n",
    "        trial_score = retrieval_grader.invoke(\n",
    "            {\n",
    "                \"patient_profile\": patient_profile, \n",
    "                \"document\": doc_txt, \n",
    "                \"trial_diseases\": trial_diseases\n",
    "            }\n",
    "        )\n",
    "            \n",
    "        relevance_score = trial_score.relevance_score\n",
    "        trial_score_dic = dict(trial_score)\n",
    "        trial_score_dic['nctid'] = nctid                    \n",
    "\n",
    "        if relevance_score.lower() == \"yes\":   \n",
    "            # Hullucination check         \n",
    "            explanation = trial_score.explanation            \n",
    "            factual_score = hallucination_grader.invoke({\"patient_profile\": patient_profile, \"explanation\": explanation})\n",
    "            factual_score_grade = factual_score.binary_score            \n",
    "            if factual_score_grade == \"no\":\n",
    "                print(\"--- HALLUCINATION: MODEL'S EXPLANATION IS NOT GROUNDED IN PATIENT PROFILE --> REJECTED---\")\n",
    "                trial_score_dic['relevance_score'] = 'no'\n",
    "                trial_score_dic['explanation'] = \"Agent's Hallucination\"\n",
    "\n",
    "        if relevance_score.lower() == \"yes\" and factual_score_grade == \"yes\":\n",
    "            print(f\"---TRIAL RELEVANT---\")                        \n",
    "        else:\n",
    "            print(f\"--- TRIAL NOT RELEVANT---\")\n",
    "\n",
    "        relevant_trials.append(trial_score_dic)\n",
    "        \n",
    "    return {\n",
    "        'last_node': 'grade_trials',\n",
    "        \"relevant_trials\": relevant_trials\n",
    "        }\n",
    "\n",
    "def profile_rewriter(state: AgentState):\n",
    "    patient_data = state['patient_data']\n",
    "    patient_profile_rewriten = profile_rewriter_chain.invoke({\"patient_data\": patient_data})\n",
    "    # print in capitals\n",
    "    print(\"--- PROFILE REWRITER: PATIENT'S PROFILE REWRITEN ---\")\n",
    "    # state['patient_profile'] = patient_profile_rewriten\n",
    "    return {\n",
    "        'last_node': 'profile_rewriter',\n",
    "        'patient_profile': patient_profile_rewriten\n",
    "    }\n",
    "    \n",
    "\n",
    "# ========= continue functions =========\n",
    "def should_continue_patient(state: AgentState):\n",
    "    # end = False\n",
    "    if state.get(\"patient_data\"):\n",
    "        # print('patient data found')\n",
    "        print(state.get(\"patient_profile\", 'No profile generated'))\n",
    "        return \"policy_search\"\n",
    "    else:\n",
    "        print('Patient data not found in the database')\n",
    "        return END\n",
    "        \n",
    "    if state[\"revision_number\"] > state[\"max_revisions\"]:\n",
    "        print('max revision reached')\n",
    "        # print(state.get(\"patient_data\", 1))\n",
    "        return END        \n",
    "\n",
    "def should_continue_policy(state: AgentState):\n",
    "    if state[\"revision_number\"] > state[\"max_revisions\"]:\n",
    "        print('max revision reached')        \n",
    "        return END\n",
    "    more_policies = len(state[\"unchecked_policies\"]) > 0\n",
    "    if state[\"policy_eligible\"]:\n",
    "        if more_policies:\n",
    "            print(\"-------------- Check the next policy --------------\")\n",
    "            return \"policy_evaluator\"\n",
    "        else:\n",
    "            print(\"-------------- Patient passed the general trials policies --------------\")\n",
    "            return \"trial_search\"\n",
    "    else:\n",
    "        print(\"Patient did not pass this policy.\\n\")\n",
    "        print(state['checked_policy'].page_content)\n",
    "        # print(\"Patient did not pass the general trials policies.\\n\")        \n",
    "        print('Rejection Reason: ')\n",
    "        nprint(state['rejection_reason'])\n",
    "        print(\"You can correct the patient's medical profile if required.\")        \n",
    "        return END\n",
    "\n",
    "def should_continue_trials(state: AgentState):\n",
    "    relevant_trials = state[\"relevant_trials\"]    \n",
    "    has_relevant_trial = any(trial['relevance_score'] == 'Yes' for trial in state[\"relevant_trials\"])\n",
    "\n",
    "    if state[\"trial_searches\"] > state[\"max_trial_searches\"]:\n",
    "        print('max trial searches reached')\n",
    "        next = END    \n",
    "    # elif not state[\"relevant_trials\"]:\n",
    "    elif has_relevant_trial == False:\n",
    "        print('----- No relevant trials found --> rewriting the patient profile -----')\n",
    "        next = \"profile_rewriter\"        \n",
    "    else:\n",
    "        next = END\n",
    "        print('to end')\n",
    "    return next\n",
    "\n",
    "\n",
    "# ==== define graph nodes====\n",
    "builder = StateGraph(AgentState)\n",
    "builder.set_entry_point(\"patient_collector\")\n",
    "builder.add_node(\"patient_collector\", patient_collector_node)\n",
    "builder.add_node(\"policy_search\", policy_search)\n",
    "builder.add_node(\"policy_evaluator\", policy_evaluator)\n",
    "builder.add_node(\"trial_search\", trial_search)\n",
    "builder.add_node(\"grade_trials\", grade_trials)\n",
    "builder.add_node(\"profile_rewriter\", profile_rewriter)\n",
    "\n",
    "# ===== define graph edges====\n",
    "builder.add_conditional_edges(\n",
    "    \"patient_collector\", \n",
    "    should_continue_patient, \n",
    "    {END: END    \n",
    "    ,\"policy_search\": \"policy_search\"\n",
    "    }\n",
    ")\n",
    "\n",
    "builder.add_conditional_edges(\n",
    "    \"policy_evaluator\", \n",
    "    should_continue_policy, \n",
    "    {\n",
    "    \"trial_search\": \"trial_search\"\n",
    "    , \"policy_evaluator\": \"policy_evaluator\"\n",
    "    , END: END\n",
    "    }\n",
    ")\n",
    "\n",
    "builder.add_edge(\"policy_search\", \"policy_evaluator\")\n",
    "builder.add_edge(\"trial_search\", \"grade_trials\")\n",
    "builder.add_edge(\"profile_rewriter\",\"trial_search\")\n",
    "# builder.set_finish_point(\"grade_trials\")\n",
    "\n",
    "builder.add_conditional_edges(\n",
    "    \"grade_trials\", \n",
    "    should_continue_trials, \n",
    "    {\n",
    "    \"profile_rewriter\": \"profile_rewriter\"    \n",
    "    , END: END\n",
    "    }\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Initiate the graph"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "\n",
    "graph = builder.compile(\n",
    "    checkpointer=memory\n",
    ")\n",
    "\n",
    "from IPython.display import Image\n",
    "\n",
    "Image(graph.get_graph().draw_png())"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Invoke the Graph with no GUI:\n",
    "You can invoke the pipeline using the `stream` method. This is useful for debugging and the development phase.\n",
    "\n",
    "The thread variable acts like different graph run profiles.\n",
    "\n",
    "To use the LangGraph persistent and time travel capabilities, use the `resume_from_state` function from the `helper_functions` module.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import langchain\n",
    "langchain.debug = False\n",
    "sample_patient_id = 2\n",
    "thread = {\"configurable\": {\"thread_id\": \"0\"}}\n",
    "events_gen = graph.stream({\n",
    "    'patient_prompt': f\"Is patient {sample_patient_id} eligible for any medical trial?\",    \n",
    "    \"max_revisions\": 10,\n",
    "    \"revision_number\": 1,\n",
    "    'max_trial_searches':3,\n",
    "}, thread)\n",
    "for event in events_gen:\n",
    "    print(event)\n",
    "# event = events_gen.__next__()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Application GUI\n",
    "For the sake of quick development, a GUI is designed based on the `gradio` python package and an adoption from [This tutorial course](https://learn.deeplearning.ai/courses/ai-agents-in-langgraph). You can also use [streamlit](https://docs.streamlit.io/develop/tutorials/llms/build-conversational-apps) as another tool for GUI development.\n",
    "\n",
    "In the applicaiton, the process is interupted at several intermediate points using the `interrupt_after` argument. At these points, the user (the clinical expert) can decide whether to alter the patient's profile by adding or removing relevant/non-relevant information. The user can also skip some of rejections at these intrupt points to continue with the process.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "initial run\n",
      "Running on local URL:  http://127.0.0.1:7959\n",
      "\n",
      "To create a public link, set `share=True` in `launch()`.\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<div><iframe src=\"http://127.0.0.1:7959/\" width=\"100%\" height=\"500\" allow=\"autoplay; camera; microphone; clipboard-read; clipboard-write;\" frameborder=\"0\" allowfullscreen></iframe></div>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/plain": []
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from helper_gui import trials_gui\n",
    "\n",
    "memory = SqliteSaver.from_conn_string(\":memory:\")\n",
    "graph = builder.compile(\n",
    "    checkpointer=memory\n",
    "    , interrupt_after=['patient_collector', 'policy_search', 'trial_search', 'grade_trials', 'profile_rewriter']\n",
    "    # , interrupt_after=['grade_trials','profile_rewriter']\n",
    ")\n",
    "\n",
    "try:\n",
    "    app.demo.close()\n",
    "except:\n",
    "    print('initial run')\n",
    "app = trials_gui(graph)    \n",
    "app.demo.launch(server_name=\"127.0.0.1\", server_port=7959)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Possible Further Improvements for Real Application:\n",
    "\n",
    "To enhance the robustness of the pipeline and effectively manage diverse patient and trial conditions, it is essential to implement advanced methodologies, particularly when handling large databases of patients, trials, and policies.\n",
    "\n",
    "Recommended Improvements:\n",
    "1. Graph-Based Retrieval-Augmented Generation (RAG):\n",
    "- Implement graph-based RAG to focus on the relationships between entities across different databases. For instance, map connections between patients, diseases, drugs, and trials to provide more accurate and relevant information retrieval.\n",
    "\n",
    "2. Advanced RAG pipeline:\n",
    "- The output of RAG pipelines can be improved by benefiting from more advanced techniques such as Adaptive RAG, Corrective RAG, and Self-RAG. Such pipelines are more robust againts hallucinations and wrong inferences in LLMs outputs.\n",
    "\n",
    "4. Advanced Chain of Thought Processing:\n",
    "- Employ a more sophisticated chain of thought process to cross-match the patient's profile with the detailed information provided in each trial or policy. This approach enhances the decision-making capability of the system by considering multiple factors simultaneously.\n",
    "\n",
    "1. Fine-Tuning the Large Language Model:\n",
    "- Fine-tune the large language model using a diverse set of medical trial data to improve the accuracy and relevance of its inferences. This customization ensures the model better understands the nuances and specifics of clinical trial data.\n",
    "\n",
    "1. Cyclic Graphs for Iterative Evaluation:\n",
    "- Utilize cyclic graphs to iteratively evaluate the patient's profile against various trials and policies. This iterative approach allows the system to refine its matches through multiple rounds of evaluation, ultimately identifying the best possible trial for the patient."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": ".llmenv",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.12.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}