AB1656 THE USE OF MACHINE LEARNING METHODS IN OPIOID-ASSOCIATED OUTCOMES RESEARCH: A SYSTEMATIC REVIEW

Abstract

Background Prescription opioids have been a considerable contributor towards the opioid epidemic, a major public health concern with profound economic repercussions and mortality. In order to address the opioid crisis, a need exists to improve understanding on how to personalise prescribing incorporating individual factors to estimate the risk of opioid-related harms. Machine learning (ML) techniques, both supervised and unsupervised, have potential identify patients at risk through the development of risk prediction models and a data-driven redefinition of patient classes. This might be particularly relevant when addressing the changing dynamics of the opioid epidemic, due to its ability to ascertain non-linear relationships between various risk factors and the risk of opioid-related harms. Objectives To perform a systematic review of the use of ML methods in prescription opioid-related harms research. Methods A systematic literature search was conducted using the databases Ovid MEDLINE, PubMed and SCOPUS databases to identify relevant articles that were published from inception to October 2022. Articles were eligible for review if they described studies that applied ML techniques with a primary outcome that was related to prescription opioid harms and using electronic health records as a main data source for the research. We excuded studies that focused patients with cancer or younger than 18 years old. Results A total of 760 abstracts were reviewed, yielding 75 articles for full text review. Four main categories were identified in this area: development of ML algorithms for prediction (n=81, 81%), identification of risk factors n=7, 9%), patient subtype clustering (n=5, 7%) and the application of natural language processing to understand user-generated text for pharmacovigilance (n=2, 3%). Within the studies that developed ML models for prediction, most of them (n=16, 21%) had prolonged postoperative opioid use as their main outcome (the majority looking at arthroplasty, arthroscopy and spine surgery patient populations). All these studies had skewed outcome proportions but only three studies (4%) addressed it (with either oversampling techniques or reporting the area under the precision-recall curve). Only one study had been externally validated. Other primary outcomes of prediction models developed included opioid use disorder (n=12, 16%), opioid overdose prediction (n=9, 12%), opioid administration and prescribing (n=6, 8%) and chronic opioid use (n=3, 4%). For risk factor identification, ML techniques were limited to understanding the drivers of opioid-overdose (n=2, 3%), opioid dependency (n=2, 3%) and opioid administration and prescribing (n=3, 4%). A very limited number of studies used unsupervised and semi-supervised ML algorithms to address opioid-associated outcomes (n=8, 11%). Conclusion To date, application of ML techniques besides logistic regression to classify patients who experience opioid-related harms has been limited, with most publications using electronic health records from the United States in the last 5 years. The majority focused on prediction algorithms for postoperative opioid use and have limited implementation in clinical practice. The current literature lacks external validation studies for developed prediction models using ML. This is especially important if implemented outside of the United States, since the use of opioids is affected by a diverse set of individual and contextual factors that can substantially vary across countries. Acknowledgements Funded by a FOREUM Career Research Grant and NIHR. MJ is supported by an NIHR Advanced Fellowship [NIHR301413]. The views expressed in this publication are those of the authors and not necessarily those of the NIHR, NHS or the UK Department of Health and Social Care. Disclosure of Interests None Declared.