Generative Machine Learning for Precision Medicine

Project Description

Introduction

This Master's thesis investigates the application of advanced generative machine learning models to improve clinical decision-making in time-sensitive medical scenarios, such as acute stroke. The focus is on using Variational Autoencoders (VAEs) and Generative Adversarial Imputation Networks (GAINs) to enhance the accuracy of clinical decision-support systems by effectively managing missing data.

Project Objectives

The project aims to design and implement generative models like VAEs and explore the potential of GAINs for patient outcome prediction. These models are integrated into a prognostic tool that compensates for incomplete data inputs, thereby supporting robust clinical decision-making. The fidelity of these generative models is critically assessed by measuring the quality of imputation on simulated datasets and evaluating their impact on the accuracy and efficiency of the prognostic pipeline.

Data Sources

The study utilizes a comprehensive set of patient datasets, including the Haddassah stroke dataset for stroke patients and Cardiovascular Disease dataset, which provide a realistic foundation for developing and testing the predictive models.

Methodology

The methodology involves developing generative models that can effectively impute missing entries in patient data, integrating these models into predictive frameworks to enhance prognosis accuracy, and determining the best procedures for calibrating the predictive models. This approach allows for an in-depth evaluation of how well each model performs under different conditions of data completeness and complexity.

Expected Outcomes

The project expects to significantly improve the precision of patient outcome predictions through the developed generative model-based pipeline, providing a measure of uncertainty due to missing data. It aims to demonstrate the potential of VAEs and GAINs in handling missing data challenges effectively in clinical datasets.

Significance of the Project

The research contributes to the development of advanced tools for clinical decision support systems, enhancing decision-making for better patient outcomes. It offers robust methods for addressing missing data challenges in critical healthcare situations and provides practical insights into the application of generative models in precision medicine.

References:
- Variational Autoencoder with Arbitrary Conditioning
- GAIN: Missing Data Imputation Using Generative Adversarial Nets