ICU Mortality Time-Series Prediction

Overview

This project aims to predict ICU patient mortality using time-series data from the PhysioNet 2012 Challenge dataset. The dataset consists of 37 physiological variables recorded over 48 hours, along with static patient information. The objective is to develop machine learning models to assist clinical decision-making.

How to run this code to get all results from A-Z

Create a new environment using the requirements or use the student-cluster and make sure to add optuna

1_EDA

• make sure to get the data on the studen cluster in ml4h/p1 and match it to the path given in 01-data-processing-exploration.ipynb
• Run 01-data-processing-exploration.ipynb to do the preprocessing and get the parquets in ./data

2_SupervisedML

• Run 2.1.2_basic_and_tsfresh_FINAL.ipynb to get the results of the 5 models with basic features as well as added features.
• Run Q2_RNN/2_RNN.ipynb to get the results for the LSTM and BiLSTM.
• Run 02-ransformers and 02-3-transformer-(tokens) to obtain the results for the transformers' task in 2.

3_RepresentationLearning

• Run 3.2_LabelScarcity_Complete_FINAL to do the LSTM and Autoencoder comparison as well as the scarcity experiments and get the patient_embeddings_* for 3.3

4_FoundationModels

• Uncomment all cells in Q4.1_LLM4TS_problem.ipynb to get the embeddings and run the notebook the get the results stated in the report
• Visualize the embeddings with 04-3-embeddings.visualizations.ipynb
• Uncomment all cells and run to get the Chronos results for q4.3. make sure to have 17GB of disk space, otherwise *.pt file cannot be written to disk (needed for q4.3.2)