50 — Machine Learning for Drug Overdose Surveillance

This paper explores two “fast subset scans” — namely, the Gaussian Process Subset Scan (GPSS) and Multidimensional Tensor Scan (MDTS) in order to scan population records for anomalous signals that might suggest a predisposition for drug-overdose. These methods ingested features such as location and time of an overdose event (in the case of GPSS) or personal characteristics and risk factors, such as gender and age group (in the case of MDTS).

In each scan, the general population is compared against $H_0$ — that a overdose does not occur — and then subsets of the general population are taken that sample from a subset of the parameter space according to the subset scan protocol. The problem of characterizing overdoses now becomes a parameter optimization problem (where the parameters account for the subset splitting of the larger population).

As an example of findings: MDTS-driven analysis found that fentanyl — an dangerous opioid variant — was responsible for clusters of overdose-related deaths in western Pennsylvania, USA. MDTS analysis subdivided the overdose population parameter space to best explain the clustering: One cluster of overdoses likely resulted from a batch of heroin with high concentrations of fentanyl in it; another cluster likely resulted from fentanyl being sold “disguised as heroin.”

By combining these findings with domain and regional expertise in localities across the world, it may become easier and more effective to stop widespread opioid overdose earlier.