HIBERT: A Hybrid Clustering BERT for Interpretable Opioid Overdose Risk Prediction

Presentation Time: 08:12 AM - 08:24 AM

Abstract Keywords: Clinical Decision Support, Deep Learning, Machine Learning, Public Health
Primary Track: Foundations
Programmatic Theme: Clinical Research Informatics

Drug overdose, mostly from opioids, is a continuing crisis in the US. Highly accurate models for early detection of opioid overdose (OD) risk are crucial for early intervention and prevention. While deep learning has shown promise in using electronic health records (EHRs) for OD risk prediction, its clinical utility is often limited by challenges posed by data sparsity, heterogeneity, and label imbalance, and lack of interpretability. We present HIBERT, a hybrid BERT model that combines the transformer model with deep clustering. HIBERT uses a multiple BERT architecture integrating specialized BERT modules for distinct EHR feature categories, and incorporates deep significance clustering to generate clinically meaningful risk stratification. HIBERT outperforms conventional and state-of-the-art models based on evaluation with the Health Facts database and identifies four distinct risk clusters, in addition to ranked critical features. It provides actionable, personalized OD risk assessment with improved interpretability.

Speaker(s):
Zihan Ding, Master's of Science
Stony Brook University

Author(s):
Zihan Ding, Master's of Science - Stony Brook University; Xinyu Dong, MS - Stony Brook University; Yinan Liu, Ms. - Stony Brook University; Tengfei Ma, PhD; Xia Zhao, MS - Stony Brook University Hospital; Rachel Wong, MD, MPH, MBA, MS - Stony Brook University, School of Medicine; Richard Rosenthal, MD; Fusheng Wang, Ph.D. - Stony Brook University;