Predicting and differentiating accidental and self-harm drug poisonings using health records data

Preventing drug poisonings or overdoses will involve both generic poisoning prevention strategies and strategies specific to self-harm poisonings. This research used health records to develop statistical models predicting drug poisoning in general and self-harm poisoning in particular after an outpatient mental health specialty visit or primary care visit with mental health diagnosis. Records regarding visits from 2015 to 2019 in four health systems were used to develop models predicting poisoning following a two-step process: prediction of any poisoning or overdose followed by differentiating self-harm from accidental or undetermined intent poisoning or overdose. Separate random forest models for mental health specialty and primary care mental health visits were developed in 70% random samples of visits and validated in the remaining 30%. Among 19,130,028 visits, 114,911 (0.60%) were followed by any poisoning and 52,063 (0.27%) by self-harm poisoning. Models had moderate performance predicting any poisoning (Area under the Receiver Operating Curve or AUC = 0.778 among mental health specialty visits and 0.767 among primary care mental health visits) and good performance for specifically predicting self-harm poisoning (AUCs = 0.836 and 0.858 respectively). Mental health visits with risk above the 95th percentile accounted for 39.7% of self-harm poisoning (sensitivity) and actual risk of 2.57% (positive predictive value). Predictors of any poisoning included prior self-harm, accidental poisoning, and mental health service use. Specific predictors of self-harm poisoning included age and prior mental health diagnoses or treatments. In conclusion, among outpatients with mental health diagnoses, prediction models using records data have moderate accuracy predicting any overdose or poisoning and good accuracy predicting self-harm poisoning. Self-harm and accidental poisoning have both shared and distinct predictors.