Abstract
BACKGROUND: Drug poisoning (overdose) is a public health crisis, particularly among people living with HIV and hepatitis C (HCV) co-infection. Identifying potential predictors of drug poisoning could help decrease drug-related deaths. METHODS: Data from the Canadian Co-infection Cohort were used to predict death due to drug poisoning within 6 months of a cohort visit. Participants were eligible for analysis if they ever reported drug use. Supervised machine learning (stratified random forest with undersampling to account for imbalanced data) was used to develop a classification algorithm using 40 sociodemographic, behavioural, and clinical variables. Predictors were ranked in order of importance, and odds ratios and 95% confidence intervals (CIs) were generated using a generalized estimating equation regression. RESULTS: Of 2,175 study participants, 1,998 met the eligibility criteria. There were 94 drug poisoning deaths, 53 within 6 months of a last visit. When applied to the entire sample, the model had an area under the curve (AUC) of 0.9965 (95% CI, 0.9941-0.9988). However, the false-positive rate was high, resulting in a poor positive predictive value (1.5%). Our model did not generalize well out of sample (AUC 0.6, 95% CI 0.54-0.68). The top important variables were addiction therapy (6 months), history of sexually transmitted infection, smoking (6 months), ever being on prescription opioids, and non-injection opioid use (6 months). However, no predictor was strong. CONCLUSIONS: Despite rich data, our model was not able to accurately predict drug poisoning deaths. Larger datasets and information about changing drug markets could help improve future prediction efforts.