Interpretable Forecasting of Overdose-Response Demand for Accountable Public-Health Prioritization

Public-health agencies responding to drug overdoses need near-term forecasts that are accurate, interpretable, and actionable under real resource constraints. In this paper, we build a model that forecasts next-month overdose-response calls across 235 Seattle census tracts, treating the outcome explicitly as recorded emergency-response demand. We evaluate a four-feature Capacity-Aware Response Estimation Explainable Boosting Machine (CARE-EBM) once on a January–June 2026 rolling-origin holdout: the CARE-EBM reduces RMSE from 0.726 to 0.663 relative to the three-month persistence baseline, and raises R² from 0.655 to 0.713. Under a capacity-aware ranking metric, Best-Possible Reach, the CARE-EBM also outperforms the 3-month average baseline model, from 0.834 to 0.896 at 10 tracts and from 0.793 to 0.858 at 20 tracts. Additionally, we construct socioeconomic and demand-trajectory typologies and cross them in a discordance matrix that shows where neighborhood context and demand trends align or diverge — giving practitioners the fuller picture needed to interpret forecasts responsibly. Because decisions like these require human oversight, we frame the system as shadow-mode decision support: predictions and context inform public-health staff, with explicit monitoring, use restrictions, and pause criteria, rather than acting on their own.