SHM Converge Scientific Abstract ePoster Gallery-Interpretable, Data-Driven Patient Grouping for Sepsis Treatment Decision Support

Pdf Summary

This document describes an interpretable, data-driven decision support system for sepsis treatment. Sepsis is a life-threatening response to infection that can cause organ failure and death, and current treatment guidelines are limited. The goal is to use machine learning and optimization to recommend real-time treatments based on patient characteristics.<br /><br />The study uses about 11,000 ICU sepsis patient records from a Boston-area hospital, covering demographics, vital signs, lab values, interventions, and outcomes. Data were aggregated hourly, yielding about 780,000 observations. Patients were modeled as a Markov Decision Process, where actions include starting or stopping IV fluids, vasopressors, antibiotics, oxygen cannula, ventilator support, and dialysis. The objective is to minimize SOFA score and improve outcomes such as discharge, while avoiding death and readmission.<br /><br />Two grouping methods were compared for creating “meta-states”:<br />- Method 1: Minimal Representation Learning (MRL), which groups patients with similar SOFA scores and splits them so each group responds similarly to treatments.<br />- Method 2: K-means clustering, which groups patients by demographics, vitals, and lab values.<br /><br />MRL performed better, with lower next-state prediction error (0.346 vs. 0.446), suggesting it predicts patient transitions more accurately. It also produced more differentiated meta-states and lower median optimal values, indicating better separation between favorable and unfavorable patient groups.<br /><br />Simulated trajectories for 1,097 patients using Method 1 suggested improved outcomes: more discharges, fewer deaths, fewer readmissions, shorter lengths of stay, and lower average SOFA scores.<br /><br />Overall, the conclusion is that MRL-based grouping combined with optimization may provide a more interpretable and effective sepsis treatment decision-support tool than standard clustering methods.

Asset Subtitle

Angela G. Lin

Meta Tag

Author List Angela G. Lin, Dessislava Pachamanova, Georgia Perakis, Lien Le

Category Research

Concept Sepsis

Concept Minimal Representation Algorithm

Concept Markov Decision Process

Concept Clinical Decision Support

Concept Meta-State

Distinguished Non-Finalist

Presenter Organization Massachusetts Institute of Technology Operations Research Center

Presenting Author Angela G. Lin

Track Technology in Hospital Medicine

Keywords

sepsis treatment

decision support system

machine learning

optimization

Markov Decision Process

ICU patient records

SOFA score

Minimal Representation Learning

K-means clustering

interpretable model

Sepsis

Minimal Representation Algorithm

Markov Decision Process

Clinical Decision Support

Meta-State