Automated mood disorder symptoms monitoring from multivariate time-series sensory data: Getting the full picture beyond a single number

This codebase was jointly developed by Filippo Corponi and Bryan M. Li. It is part of the paper "Automated mood disorder symptoms monitoring from multivariate time-series sensory data: Getting the full picture beyond a single number", published in Nature Translational Psychiatry. If you find this code or any of the ideas in the paper useful, please consider starring this repository and citing:

@article{corponi2024automated,
  title={Automated mood disorder symptoms monitoring from multivariate time-series sensory data: Getting the full picture beyond a single number},
  author={Corponi, Filippo and Li, Bryan M and Anmella, Gerard and Mas, Ariadna and Pacchiarotti, Isabella and Valent{\'\i}, Marc and Grande, Iria and Benabarre, Antoni and Garriga, Marina and Vieta, Eduard and others},
  journal={Translational Psychiatry},
  volume={14},
  number={1},
  pages={161},
  year={2024},
  publisher={Nature Publishing Group UK London}
}

Installation

• Create a new conda environment with Python 3.10.

  conda create -n timebase python=3.10

• Activate timebase virtual environment

  conda activate timebase

• Install all dependencies and packages with setup.sh script, works on both Linus and macOS.

  sh setup.sh

Data

• data/README.md details the structure of the dataset.
• Preprocess dataset. In the present work, time alignment was fixed to 0 --time_alignment 0, whereas window length was explored in powers of 2 (ranging from 2² to 2¹⁰). Below is an example command for with --segment_length 64. Please note that all the datasets corresponding to the different window lengths to be explored during tuning should be created in advance.

  python preprocess_ds.py --segment_length 64 --time_alignment 0 --output_dir data/preprocessed/ta0_sl64 --overwrite

  • please see --help for all available options.

Mood Disorder Symptoms Inference

1) Deep Learning

Below is an example command to create a supervised neural network for the task at hand.

python train.py --batch_size 64 --epochs 200 --seed 1234 --min_epochs 50 --lr_patience 10 --dataset data/preprocessed/ta0_sl64 --emb_dim 128 --lr 0.001 --weight_decay 0.001 --verbose 1 --num_units 128 --dropout 0 --model bilstm --task_mode 0 --imb_mode 0 --emb_type 0 --output_dir runs/test --clear_output_dir --save_predictions --test_time

• please see --help for all available options.

Hyperparameters tuning

Hyperparameters tuning relies on Weight & Biases. Once a sweep has been created on wandb website, tuning can be run as shown below, replacing <sweep_id> and <group_name> with the values created online on wandb.

python sweep.py --sweep_id <sweep_id> --wandb_group <group_name> --num_trials 100 --output_dir runs/test

Once tuning has been carried out and the set of hyperparameters associated with the heightest validation set performance has been identified, plug the appropriate hyperparameters in the command shown in 1) Deep Learning.

Post-hoc analyses

Assuming the best model output was saved to runs/<best_model>, in order to perform the post-hoc analyses on the best performing model, run:

python post_hoc_analysis.py --experiment_dir runs/<best_model>

To produce the graphical model on item residuals, once the above has run, run the following. This assumes R programming language has been install into the machine:

Rscript ResidualNetwork.R -f runs/<best_model>/residuals.csv

2) Classical Machine Learning

Below is an example command to tune and test random forest classifiers (one for each item in the Hamilton Depression Rating Scale and Young Mania Rating Scale).

python train_baseline.py --path2preprocessed data/preprocessed --model random_forest --output_dir runs/rf  --clear_output_dir