Temporal Drivers of Opioid-Related ED Visits: An Ensemble Machine Learning Study With Consensus-Based Feature...

Area of Science:

• Health Informatics
• Pharmacogenomics
• Public Health

Background:

• Identifying modifiable risk factors for prescription opioid use disorder (OUD)-related emergency department (ED) visits is crucial.
• Existing models often lack pharmacogenomic (PGx) data and robust feature validation.

Purpose of the Study:

• To develop and validate a predictive model for prescription OUD-related ED visits using comprehensive claims data.
• To incorporate pharmacogenomic burden and pre-index temporal dynamics into risk prediction.
• To identify actionable risk features and optimal intervention timing.

Main Methods:

• Analysis of Virginia All-Payer Claims Database (2016-2019) for 6,929,576 patients.
• Machine learning models (CatBoost, XGBoost, XGBoost-RF) trained on PGx burden and temporal dynamics.
• Consensus Filter (SHAP ∩ FFA) and Dynamic Time Warping (DTW) for risk feature identification and trajectory analysis.

Main Results:

• The model achieved significant predictive lift (2.3× to 3.4×) and an AUROC of 0.800 in the 25-44 age band.
• Top risk features included gabapentin, long-term opioid use (Z79.891), and pharmacogenomic drug burden (pgx_num_drugs).
• DTW identified three distinct utilization archetypes with a mean pre-index time-to-target of 6.8 months.

Conclusions:

• The study provides an observational framework for risk attribution in prescription OUD-related ED care.
• Findings specify intervention targets (medication, care) and optimal timing (surveillance windows).
• The association-versus-causation distinction is maintained to support clinical use and future causal validation.