AutoPrognosis - A system for automating the design of predictive modeling pipelines tailored for clinical prognosis.

🔑 Features

• 🚀 Automatically learns ensembles of pipelines for classification, regression or survival analysis tasks.
• 🌀 Easy to extend pluginable architecture.
• 🔥 Interpretability and uncertainty quantification tools.
• 🩹 Data imputation using HyperImpute.
• ⚡ Build demonstrators using Streamlit.
• 📓 Python and R tutorials available.
• 📖 Read the docs

Note : The R bindings have been tested using R version 4.2 and Python 3.8.

🚀 Installation

Using pip

The library can be installed from PyPI using

$ pip install autoprognosis

or from source, using

$ pip install .

Redis (Optional, but recommended)

AutoPrognosis can use Redis as a backend to improve the performance and quality of the searches.

For that, install the redis-server package following the steps described on the official site.

Environment variables

The library can be configured from a set of environment variables.

Variable	Description
N_OPT_JOBS	Number of cores to use for hyperparameter search. Default : 1
N_LEARNER_JOBS	Number of cores to use by inidividual learners. Default: all cpus
REDIS_HOST	IP address for the Redis database. Default 127.0.0.1
REDIS_PORT	Redis port. Default: 6379

Example: export N_OPT_JOBS = 2 to use 2 cores for hyperparam search.

💥 Sample Usage

Advanced Python tutorials can be found in the Python tutorials section.

R examples can be found in the R tutorials section.

List the available classifiers

from autoprognosis.plugins.prediction.classifiers import Classifiers
print(Classifiers().list_available())

Create a study for classifiers

from sklearn.datasets import load_breast_cancer

from autoprognosis.studies.classifiers import ClassifierStudy
from autoprognosis.utils.serialization import load_model_from_file
from autoprognosis.utils.tester import evaluate_estimator


X, Y = load_breast_cancer(return_X_y=True, as_frame=True)

df = X.copy()
df["target"] = Y

study_name = "example"

study = ClassifierStudy(
    study_name=study_name,
    dataset=df,  # pandas DataFrame
    target="target",  # the label column in the dataset
)
model = study.fit()

# Predict the probabilities of each class using the model
model.predict_proba(X)

(Advanced) Customize the study for classifiers

from pathlib import Path

from sklearn.datasets import load_breast_cancer

from autoprognosis.studies.classifiers import ClassifierStudy
from autoprognosis.utils.serialization import load_model_from_file
from autoprognosis.utils.tester import evaluate_estimator


X, Y = load_breast_cancer(return_X_y=True, as_frame=True)

df = X.copy()
df["target"] = Y

workspace = Path("workspace")
study_name = "example"

study = ClassifierStudy(
    study_name=study_name,
    dataset=df,  # pandas DataFrame
    target="target",  # the label column in the dataset
    num_iter=100,  # how many trials to do for each candidate
    timeout=60,  # seconds
    classifiers=["logistic_regression", "lda", "qda"],
    workspace=workspace,
)

study.run()

output = workspace / study_name / "model.p"
model = load_model_from_file(output)

# <model> contains the optimal architecture, but the model is not trained yet. You need to call fit() to use it.
# This way, we can further benchmark the selected model on the training set.
metrics = evaluate_estimator(model, X, Y)

print(f"model {model.name()} -> {metrics['str']}")

# Train the model
model.fit(X, Y)

# Predict the probabilities of each class using the model
model.predict_proba(X)

List the available regressors

from autoprognosis.plugins.prediction.regression import Regression
print(Regression().list_available())

Create a Regression study

# third party
import pandas as pd

# autoprognosis absolute
from autoprognosis.utils.serialization import load_model_from_file
from autoprognosis.utils.tester import evaluate_regression
from autoprognosis.studies.regression import RegressionStudy

# Load dataset
df = pd.read_csv(
    "https://archive.ics.uci.edu/ml/machine-learning-databases/00291/airfoil_self_noise.dat",
    header=None,
    sep="\\t",
)
last_col = df.columns[-1]
y = df[last_col]
X = df.drop(columns=[last_col])

df = X.copy()
df["target"] = y

# Search the model
study_name="regression_example"
study = RegressionStudy(
    study_name=study_name,
    dataset=df,  # pandas DataFrame
    target="target",  # the label column in the dataset
)
model = study.fit()

# Predict using the model
model.predict(X)

Advanced Customize the Regression study

# stdlib
from pathlib import Path

# third party
import pandas as pd

# autoprognosis absolute
from autoprognosis.utils.serialization import load_model_from_file
from autoprognosis.utils.tester import evaluate_regression
from autoprognosis.studies.regression import RegressionStudy

# Load dataset
df = pd.read_csv(
    "https://archive.ics.uci.edu/ml/machine-learning-databases/00291/airfoil_self_noise.dat",
    header=None,
    sep="\\t",
)
last_col = df.columns[-1]
y = df[last_col]
X = df.drop(columns=[last_col])

df = X.copy()
df["target"] = y

# Search the model
workspace = Path("workspace")
workspace.mkdir(parents=True, exist_ok=True)

study_name="regression_example"
study = RegressionStudy(
    study_name=study_name,
    dataset=df,  # pandas DataFrame
    target="target",  # the label column in the dataset
    num_iter=10,  # how many trials to do for each candidate. Default: 50
    num_study_iter=2,  # how many outer iterations to do. Default: 5
    timeout=50,  # timeout for optimization for each classfier. Default: 600 seconds
    regressors=["linear_regression", "xgboost_regressor"],
    workspace=workspace,
)

study.run()

# Test the model
output = workspace / study_name / "model.p"

model = load_model_from_file(output)
# <model> contains the optimal architecture, but the model is not trained yet. You need to call fit() to use it.
# This way, we can further benchmark the selected model on the training set.

metrics = evaluate_regression(model, X, y)

print(f"Model {model.name()} score: {metrics['str']}")

# Train the model
model.fit(X, y)


# Predict using the model
model.predict(X)

List available survival analysis estimators

from autoprognosis.plugins.prediction.risk_estimation import RiskEstimation
print(RiskEstimation().list_available())

Create a Survival analysis study

# third party
import numpy as np
from pycox import datasets

# autoprognosis absolute
from autoprognosis.studies.risk_estimation import RiskEstimationStudy
from autoprognosis.utils.serialization import load_model_from_file
from autoprognosis.utils.tester import evaluate_survival_estimator

df = datasets.gbsg.read_df()
df = df[df["duration"] > 0]

X = df.drop(columns = ["duration"])
T = df["duration"]
Y = df["event"]

eval_time_horizons = np.linspace(T.min(), T.max(), 5)[1:-1]

study_name = "example_risks"

study = RiskEstimationStudy(
    study_name=study_name,
    dataset=df,
    target="event",
    time_to_event="duration",
    time_horizons=eval_time_horizons,
)

model = study.fit()

# Predict using the model
model.predict(X, eval_time_horizons)

Advanced Customize the Survival analysis study

# stdlib
import os
from pathlib import Path

# third party
import numpy as np
from pycox import datasets

# autoprognosis absolute
from autoprognosis.studies.risk_estimation import RiskEstimationStudy
from autoprognosis.utils.serialization import load_model_from_file
from autoprognosis.utils.tester import evaluate_survival_estimator

df = datasets.gbsg.read_df()
df = df[df["duration"] > 0]

X = df.drop(columns = ["duration"])
T = df["duration"]
Y = df["event"]

eval_time_horizons = np.linspace(T.min(), T.max(), 5)[1:-1]

workspace = Path("workspace")
study_name = "example_risks"

study = RiskEstimationStudy(
    study_name=study_name,
    dataset=df,
    target="event",
    time_to_event="duration",
    time_horizons=eval_time_horizons,
    num_iter=10,
    num_study_iter=1,
    timeout=10,
    risk_estimators=["cox_ph", "survival_xgboost"],
    score_threshold=0.5,
    workspace=workspace,
)

study.run()

output = workspace / study_name / "model.p"

model = load_model_from_file(output)
# <model> contains the optimal architecture, but the model is not trained yet. You need to call fit() to use it.
# This way, we can further benchmark the selected model on the training set.

metrics = evaluate_survival_estimator(model, X, T, Y, eval_time_horizons)

print(f"Model {model.name()} score: {metrics['str']}")

# Train the model
model.fit(X, T, Y)

# Predict using the model
model.predict(X, eval_time_horizons)

🔆 Tutorials

Plugins

• Imputation
• Preprocessing
• Classification
• Pipelines
• Interpretability
• Survival Analysis
• Regression

AutoML

• Classification tasks
• Classification tasks with imputation
• Survival analysis tasks
• Survival analysis tasks with imputation
• Regression tasks
• Classifiers with explainers
• Multiple imputation example

Building a demonstrator

• Classification demonstrator
• Survival analysis demonstrator

⚡ Plugins

Imputation methods

from autoprognosis.plugins.imputers import  Imputers

imputer = Imputers().get(<NAME>)

Name	Description
hyperimpute	Iterative imputer using both regression and classification methods based on linear models, trees, XGBoost, CatBoost and neural nets
mean	Replace the missing values using the mean along each column with SimpleImputer
median	Replace the missing values using the median along each column with SimpleImputer
most_frequent	Replace the missing values using the most frequent value along each column with SimpleImputer
missforest	Iterative imputation method based on Random Forests using IterativeImputer and ExtraTreesRegressor
ice	Iterative imputation method based on regularized linear regression using IterativeImputer and BayesianRidge
mice	Multiple imputations based on ICE using IterativeImputer and BayesianRidge
softimpute	Low-rank matrix approximation via nuclear-norm regularization
EM	Iterative procedure which uses other variables to impute a value (Expectation), then checks whether that is the value most likely (Maximization) - EM imputation algorithm
gain	GAIN: Missing Data Imputation using Generative Adversarial Nets

Preprocessing methods

from autoprognosis.plugins.preprocessors import Preprocessors

preprocessor = Preprocessors().get(<NAME>)

Name	Description
maxabs_scaler	Scale each feature by its maximum absolute value. MaxAbsScaler
scaler	Standardize features by removing the mean and scaling to unit variance. - StandardScaler
feature_normalizer	Normalize samples individually to unit norm. Normalizer
normal_transform	Transform features using quantiles information.QuantileTransformer
uniform_transform	Transform features using quantiles information.QuantileTransformer
minmax_scaler	Transform features by scaling each feature to a given range.MinMaxScaler

Classification

from autoprognosis.plugins.prediction.classifiers import Classifiers

classifier = Classifiers().get(<NAME>)

Name	Description
neural_nets	PyTorch based neural net classifier.
logistic_regression	LogisticRegression
catboost	Gradient boosting on decision trees - CatBoost
random_forest	A random forest classifier. RandomForestClassifier
tabnet	TabNet : Attentive Interpretable Tabular Learning
xgboost	XGBoostClassifier

Survival Analysis

from autoprognosis.plugins.prediction.risk_estimation import RiskEstimation

predictor = RiskEstimation().get(<NAME>)

Name	Description
survival_xgboost	XGBoost Survival Embeddings
loglogistic_aft	Log-Logistic AFT model
deephit	DeepHit: A Deep Learning Approach to Survival Analysis with Competing Risks
cox_ph	Cox’s proportional hazard model
weibull_aft	Weibull AFT model.
lognormal_aft	Log-Normal AFT model
coxnet	CoxNet is a Cox proportional hazards model also referred to as DeepSurv

Regression

from autoprognosis.plugins.prediction.regression import Regression

regressor = Regression().get(<NAME>)

Name	Description
tabnet_regressor	TabNet : Attentive Interpretable Tabular Learning
catboost_regressor	Gradient boosting on decision trees - CatBoost
random_forest_regressor	RandomForestRegressor
xgboost_regressor	XGBoostClassifier
neural_nets_regression	PyTorch-based neural net regressor.
linear_regression	LinearRegression

Explainers

from autoprognosis.plugins.explainers import Explainers

explainer = Explainers().get(<NAME>)

Name	Description
risk_effect_size	Feature importance using Cohen's distance between probabilities
lime	Lime: Explaining the predictions of any machine learning classifier
symbolic_pursuit	[Symbolic Pursuit](Learning outside the black-box: at the pursuit of interpretable models)
shap_permutation_sampler	SHAP Permutation Sampler
kernel_shap	SHAP KernelExplainer
invase	INVASE: Instance-wise Variable Selection

Uncertainty

from autoprognosis.plugins.uncertainty import UncertaintyQuantification
model = UncertaintyQuantification().get(<NAME>)

Name	Description
cohort_explainer
conformal_prediction
jackknife

🔨 Test

After installing the library, the tests can be executed using pytest

$ pip install .[testing]
$ pytest -vxs -m "not slow"

Citing

If you use this code, please cite the associated paper:

@misc{https://doi.org/10.48550/arxiv.2210.12090,
  doi = {10.48550/ARXIV.2210.12090},
  url = {https://arxiv.org/abs/2210.12090},
  author = {Imrie, Fergus and Cebere, Bogdan and McKinney, Eoin F. and van der Schaar, Mihaela},
  keywords = {Machine Learning (cs.LG), Artificial Intelligence (cs.AI), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {AutoPrognosis 2.0: Democratizing Diagnostic and Prognostic Modeling in Healthcare with Automated Machine Learning},
  publisher = {arXiv},
  year = {2022},
  copyright = {Creative Commons Attribution 4.0 International}
}

References

1. AutoPrognosis: Automated Clinical Prognostic Modeling via Bayesian Optimization with Structured Kernel Learning
2. Prognostication and Risk Factors for Cystic Fibrosis via Automated Machine Learning
3. Cardiovascular Disease Risk Prediction using Automated Machine Learning: A Prospective Study of 423,604 UK Biobank Participants