Machine Learning on temporal data from ICU and chronic disease to predict future values and outcomes

This repository contains the code associated with my MSc project. Two main challenges are adressed: (1) Predict Traumatic Brain Injury patients outcomes using the MIMIC-IV database, (2) Conduct a survival analysis of Chronic pulmonary aspergillosis (CPA) patients using numerical features extracted from CT-scans.

Additional_analysis

This folder contains the files for the annex CPA analysis. model.ipynb is a notebook containing (1) ML models for 5-year mortality prediction, (2) univariate and multivariate survival analysis for 5-year mortality.

Archives

Contains former files such as the preprocessing using for the demo version of MIMIC-III used at the beginning of the project.

MIMIC-IV

Contains the TBI management study files, using MIMIC-IV database.

Database analysis

Contains files for exploratory data analysis of the TBI cohort in MIMIC-IV.

SQL Preprocessing Pipeline

• height_weight_table.sql: computes the height/weight table by performing unit transformation.
• preprocessing_pipeline.sql: SQL preprocessing for TBI cohort extraction, data aggregation, feature extraction, feature engineering. It outputs tables provided in google drive which will be used to train the different models.
• preprocessing_pipeline_augmented.sql: SQL preprocessing for augmented cohort, used for Transfer Learning source task.

Tasks code

Contains code to run for the LOS and BP regression tasks.

• data_augmentation.py: class for data augmentation using VAE.
• evaluation.py: class for models evaluation
• GRU.py: class with the GRU architecture
• hyp_tuning.py: class for hyperparameter tuning using Hyperopt
• preprocessing.py: class for extra-preprocessing (data imputation, outlier removal...)
• main_LOS.py: running this file will train the best model obtained for LOS binary classification
• main_transfer_learning.py: running this file will train the best GRU model for Blood Pressures regression
• main_task_2.py: running this file will train the best ML model for Blood Pressures regression
• VAE.py: VAE architecture for data augmentation
• VAE_model.pkl: already trained VAE model