Predicting fatal drug poisoning (overdose) among people living with HIV-HCV co-infection

Background: Drug poisoning (overdose) is an important public health crisis, particularly among people living with HIV and hepatitis C (HIV-HCV) co-infection. Direct-acting antivirals result in high HCV cure rates, successfully reducing liver-related mortality. However, increased rates of drug poisoning deaths will negate these benefits. Investigating the potential predictors for drug poisoning could help to reduce mortality by identifying groups most at-risk. Objective: The objective of this thesis was to predict six-month drug poisoning mortality among people with HIV-HCV coinfection using socioeconomic, behavioural, and clinical factors, including factors that are routinely measured in clinical practice, as well as those recorded for research purposes.Methods: Data from the Canadian Co-infection Cohort (CCC) were used. Participants were followed up at six-month intervals when they completed questionnaires on socio-demographic, behavioural and clinical factors. Participants were eligible for analysis if they ever reported injection or non-injection drug use between 2003 and 2023. The outcome was death due to drug poisoning within six months of a participant’s cohort visit. We selected a total of 40 predictors. We used a supervised machine learning model, random forest, to develop a classification algorithm. Due to imbalanced data, we used a stratified random forest approach with undersampling. Predictors of drug poisoning were ranked in order of importance and odds ratios (OR) and 95% confidence intervals (CIs) were generated using a generalized estimating equation (GEE) regression with the top five important predictors. Four sensitivity analyses were conducted.Results: Of 2,132 total CCC participants, 1,998 met the eligibility criteria for this analysis. Of those eligible, 1,764 (88.3%) reported ever using injection drugs and 1,807 (90.4%) reported ever using non-injection drugs. From a total of 94 drug poisoning deaths, 53 occurred within six months of a participant’s last visit. When applied to the out-of-bag sample, the model had an area under the curve (AUC) of 0.61 (95% CI: 0.54, 0.68), indicating poor performance. When applied to the entire sample, the model performed better with an AUC of 0.9965 (95% CI: 0.9941, 0.9988). When ranking the predictors by importance, the top five variables were: addiction therapy in the past six months (6m), history of sexually transmitted infection, smoking (6m), ever being on prescription opioids, and non-injection opioid use (6m). However, the mean decrease in accuracy was low for all variables, indicating that no predictor was very strong. Additionally, the ORs generated by the GEE of the top important variables were close to the null, and almost all 95% CIs associated with these ORs crossed the null, preventing any definitive conclusions to be made on the direction of the association.Discussion: Ranking variables by importance pointed to some interesting clues as to who might be at risk for fatal drug poisonings, however, due to the challenges we faced in prediction, these results must be interpreted with caution. Our model performed poorly when withholding a sample of the data, and even the most important predictors had little impact on the overall accuracy. These results suggest that drug poisoning deaths may be a random event within the cohort and could reflect the toxicity of the drug supply. Alternatively, the low number of events and imbalanced data would benefit from exploring alternative approaches to investigate this question.Conclusion: Understanding the predictors of short-term risk of drug poisoning is an important first step for developing clinical tools to target at-risk patients. However, our model performed relatively poorly. As we are unable to identify specific predictors of who is most at risk, efforts need to be placed elsewhere, such as interventions to reduce the toxicity of the supply

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.