overdose | Fabrice Harel-Canada

Nowcasting Neighborhood Overdose Deaths in LA County

Thu, 10 Sep 2026 00:00:00 -0700

Overdose death records take 90–120 days to finalize, so public health teams are always looking at the past. As part of Rapid Overdose Surveillance Los Angeles (ROSLA), this project builds nowcasts of overdose mortality across Los Angeles County ZIP codes to close that gap.

Models. Gradient-boosted Poisson count models trained in a walk-forward backtest on finalized Medical Examiner records from 2012 to 2025, so every estimate uses only data that would have been available at the time.
Signals. Each neighborhood’s own overdose history is combined with public data: treatment-facility access, deprivation indices, homeless counts, drug-checking results, EMS calls, and search trends.
Knowing what’s knowable. When a ZIP code averages about one death per quarter, random chance dominates. The project measures the ceiling that counting noise places on any model, and uses it to choose the geography and time period where a real rise becomes detectable.

How well each approach identifies the 20 highest-burden ZIP codes, scored against realized counts (open) and against four-quarter deaths that smooth out counting noise (filled).

Manuscript in preparation.

Detecting Emerging Drugs in LA County Overdose Deaths

Wed, 09 Sep 2026 00:00:00 -0700

The drug supply behind overdose deaths keeps changing: para-fluorofentanyl, bromazolam, carfentanil, xylazine. Death records usually can’t keep up. In Los Angeles County, the modeling panel carries just 8 coded substance flags and lumps everything else into “Others.” The medical examiner’s free-text cause-of-death narratives are far richer, and this project uses them to answer a simple question: which substances are emerging, and where?

Extraction. A lexicon of 4,089 aliases mapped to 3,127 canonical substances, built from the DEA NFLIS catalog plus curated misspellings and metabolite rollups, turns 8 coded columns into 219 distinct substances at 99.8% decedent coverage.
Detection. An empirical-Bayes (gamma-Poisson) shrinkage ranking flags substances whose share of deaths is rising. It combines case-weighting, dual-gate normalization, and a closed-form spatial-clustering term, alongside a tree-temporal scan over the pharmacological hierarchy. After prototyping more than 20 detector variants, the final design reaches 0.85 precision.
Validation. Every variant is back-tested quarter by quarter using only the data available at the time, and scored against hand-curated emergence timelines and synthetic Monte Carlo profiles with known ground truth.
Characterization. Spatial and space-time scans show whether a flagged substance is concentrated geographically, and polysubstance analysis separates substances that cause deaths from ones that are merely present.

Known emerging substances in LA County. Yellow bands mark the quarters the detector was in alarm; dotted lines mark the first recorded LA death. para-Fluorofentanyl was flagged one quarter after its first death, while xylazine, with only 10 deaths, never crossed the threshold.

Improving Drug Identification in Overdose Death Surveillance by Using Clinical Natural Language Processing Models

Sun, 08 Feb 2026 00:00:00 -0800