Publication Date

Spring 2024

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

William Andreopoulos

Second Advisor

Navrati Saxena

Third Advisor

Thomas Austin

Keywords

Personalized Medical Predictions, BEHRT model, Graph Attention Networks

Abstract

Over the past few years, personalized medicine has gained traction due to its ability to solve medical issues efficiently for a person based on their personal characteristics. This project aims to design a machine-learning model that can generate predictions of lab test scores in the future based on past medical history. The model is trained using the MIMIC-4 (Medical Information Mart for Intensive Care) dataset that consists of medical records of over 40,000 patients. The proposed model, MOE-BEHRT, consists of Bidirectional Encoder Representations from Transformers on Electronic Health Records (BEHRT) and Mixture of Experts (MOE). The BEHRT model was originally designed for multi-class classification. In this project, we’ve fine-tuned the BEHRT model to predict future lab test scores based on each patient’s medical history and patient similarity data. Cosine similarity metric and Graph Attention Networks (GATv2) are used to identify the similarities between patients based on their medical history. Based on the experimental results, the MOE-BEHRT model performs better than the Linear Regression by 6.85% on Blood Sugar tests, 1.684% on Cholestrol and 3.81755% on Alkaline Phosphate tests on Root Mean Square Error (RMSE). On comparing the RMSE results of MOE-BEHRT with a neural network, we observed that MOE-BEHRT performed better by 87.66% on Blood Sugar tests, 26.9867% on Cholesterol and 3.54855% on Alkaline Phosphate.

Download

Available for download on Friday, May 23, 2025

Share

COinS