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

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Public Health

Background:

• Opioid overdose (OD) is a critical public health issue in the US.
• Accurate early detection of OD risk is vital for intervention and prevention.
• Existing deep learning models for OD risk prediction face challenges like data sparsity, heterogeneity, label imbalance, and lack of interpretability.

Purpose of the Study:

• To develop and evaluate HIBERT, a hybrid BERT model for enhanced opioid overdose risk prediction.
• To address the limitations of current models by integrating deep clustering for clinically meaningful risk stratification.
• To provide actionable, personalized OD risk assessments with improved interpretability.

Main Methods:

• Developed HIBERT, a hybrid model combining a transformer architecture (BERT) with deep clustering.
• Utilized a multiple BERT architecture with specialized modules for different electronic health record (EHR) feature categories.
• Incorporated deep significance clustering for risk stratification and identified critical predictive features.

Main Results:

• HIBERT demonstrated superior performance compared to conventional and state-of-the-art models on the Health Facts database.
• The model identified four distinct opioid overdose risk clusters.
• Key features contributing to OD risk were ranked, enhancing model interpretability.

Conclusions:

• HIBERT offers a promising approach for accurate and interpretable opioid overdose risk prediction.
• The model's ability to stratify risk and identify critical features provides actionable insights for personalized interventions.
• This hybrid AI model has the potential to improve clinical utility in managing the opioid crisis.