Background/Objectives: This study aimed to evaluate the frequency and predictors of adverse drug events (ADEs) related to medication abuse, misuse, and dependence, along with serious adverse events (SAEs), and to develop machine learning models to detect serious abuse, misuse, and dependence cases. Methods: This study included 455,415 ADE reports involving medications with high dependence potential reported to the Korea Adverse Event Reporting System (KIDS KAERS DB) between 2013 and 2022. Multivariate logistic regression was used to identify predictors. Three machine learning algorithms, random forest (RF), support vector machine, and eXtreme Gradient Boosting, were developed and evaluated. Results: Higher reporting likelihood of abuse-, misuse-, and dependence-related ADEs was observed with concomitant use of acetaminophen (OR 3.60, 95% CI 2.40–5.39), antidepressants (OR 1.75, 95% CI 1.17–2.61), antipsychotics (OR 4.97, 95% CI 3.21–7.17), and anticonvulsants (OR 3.42, 95% CI 2.42–4.81). Reports from the general public were associated with higher odds of abuse, misuse, and dependence than those from healthcare professionals (OR 4.59, 95% CI 3.04–6.94). Ketamine (ROR 14.03) and bromazepam (ROR 13.02) showed the highest likelihood of being classified as SAEs. Cardiovascular (ROR 30.36) and respiratory disorders (ROR 17.03) demonstrated the highest SAE reporting likelihood. RF model demonstrated the best predictive performance (AUC-ROC 0.928; accuracy 94.4%), with reporter type identified as a key feature. Conclusions: RF model demonstrated optimal predictive performance, with reporter type as the most important feature for detecting serious cases. This study emphasizes the importance of incorporating patient-reported data and polypharmacy surveillance to facilitate early detection of serious cases.
Kim et al. (Thu,) studied this question.