Integrating clinical notes for enhanced mortality prediction in ICU

Predicting patient outcomes in Intensive Care Units (ICUs) is a critical task for im- proving patient management and treatment planning. Traditional predictive models primarily rely on structured time-series data (e.g., vital signs, lab results) to forecast outcomes like in-hospital mortality. Ho...

Full description

Saved in:
Bibliographic Details
Main Author: Pang, Kelvin
Other Authors: Fan Xiuyi
Format: Final Year Project
Language:English
Published: Nanyang Technological University 2024
Subjects:
Online Access:https://hdl.handle.net/10356/181084
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Predicting patient outcomes in Intensive Care Units (ICUs) is a critical task for im- proving patient management and treatment planning. Traditional predictive models primarily rely on structured time-series data (e.g., vital signs, lab results) to forecast outcomes like in-hospital mortality. However, these models often overlook the rich, un- structured information available in clinical notes, which can provide important context about a patient’s condition that is not reflected in structured data alone. This study explores the viability of integrating unstructured clinical text data into mor- tality prediction models using the STraTS (Self-Supervised Transformer for Sparse and Irregularly Sampled Multivariate Clinical Time-Series) model. The primary ob- jective is to assess whether text embeddings derived from clinical notes—processed using Clinical Longformer and BioWordVec—can enhance predictive performance in comparison to the baseline STraTS model, which uses structured time-series and demographic data. The predictive performance of three model variants, the baseline STraTS model, Clin- ical Longformer, and BioWordVec models, will be evaluated. Models were evaluated using ROC-AUC, PR-AUC, and (min(Re,Pr)). While the baseline model outperformed both Clinical Longformer and BioWordVec, the Clinical Longformer demonstrated potential in optimizing the precision-recall trade-off, a critical factor in mortality pre- diction tasks involving imbalanced datasets. The findings suggest that unstructured text data offers complementary value, while structured data remains a more reliable predictor of mortality.