Knowledge Distillation for Input-Driven Switching State-Space Models for Clinical Decision Making
Name
sheriff-fareeds-meng-eecs-2026-thesis.pdf
Size
3.72 MB
Format
Adobe PDF
Checksum (MD5)
69c9bd0479b80d13cdb7cd8dcc0d97c5
Author(s)
Sheriff, Fareed
Advisor(s)
Lehman, Li-wei
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
We propose a framework for knowledge distillation of deep learning models into interpretable probabilistic graphical models (PGMs) with input-driven dynamics for sequential treatment decision making in healthcare in order to model evolution of patient health states in response to therapeutic actions in a clinical setting. While deep learning models like LSTMs achieve superior predictive performance in healthcare applications, their black-box nature limits clinical adoption. Conversely, interpretable models such as autoregressive Hidden Markov Models (AR-HMMs) offer transparency but do not necessarily learn representations that are predictive of patient outcomes for informed treatment decision making. To bridge this gap, we extend recent work on Constrained Variational Inference Knowledge Distillation to incorporate input-driven dynamics, allowing models to account for the effect of treatments (inputs) on patient state evolution over time. Our preliminary results on the MIMIC-IV dataset show promise for this approach, with knowledge-distilled AR-HMMs performing significantly better than baseline models while maintaining interpretability. This work lays methodological groundwork for future clinical decision support research by enabling more accurate and interpretable modeling of treatment-dependent patient dynamics.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
