Machine Learning Applications in Predicting Opioid-Associated Adverse Events: A Systematic Review

Machine learning has increasingly been applied to predict adverse outcomes in healthcare, including those associated with opioids due to its ability to handle complex interactions and potential for generating actionable predictions. This systematic review aimed to evaluate the types and quality of ML methods used in opioid safety research. A search was conducted using Ovid MEDLINE, PubMed and SCOPUS databases from inception to October 2023 to identify studies applying supervised ML techniques to predict opioid-related harms. Prediction Model Risk of Bias Assessment Tool (PROBAST) was used to assess risk of bias. Model characteristics and performance measures were extracted using Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS). From 907 abstracts, 74 full-text articles were reviewed and 44 included. The most commonly predicted outcomes were postoperative opioid use (n=16, 21%), opioid overdose (n=8, 18%), opioid use disorder (n=8, 18%) and persistent opioid use (n=5, 11%) with varying definitions. Most studies originated from North America (96%), with external validation reported in only 7%. Model performance ranged from moderate to strong, but calibration was reported less frequently (41% missing). Although imbalanced datasets were frequent, only 45% addressed the issue methodologically. Transparent reporting of model development was often incomplete, with key aspects such as calibration, imbalance correction, and handling of missing data absent. Infrequent external validation limited the generalizability of current models. Addressing these aspects is critical for transparency, interpretability, and future implementation of the results.