ABSTRACT
Background: Chronic opioid use following postoperative pain poses a significant clinical risk and global economic burden. Various prediction models have been developed to augment clinical decision-making and mitigate risks associated with opioid use. Our aim is to identify, describe and compare currently available models to predict chronic opioid use following surgery.
Methods: We searched OVID MEDLINE, OVID Embase, Pubmed and Scopus databases to identify eligible studies using pre-determined criteria following PRISMA guidelines. Two independent investigators performed title and abstract screening and then a full-text review of relevant studies to determine inclusion in the systematic review. Data extraction of included studies was performed following guidance from CHecklist for critical Appraisal and data extraction for systematic Reviews of prediction Model Studies (CHARMS). Risk of bias was assessed using the Prediction model Risk of Bias Assessment Tool (PROBAST).
Results: 410 unique studies were identified from the search strategy of which 10 studies were included in the systematic review. All studies were performed in the USA and six studies were of orthopaedic patients. Commonly selected predictor variables include demographics, preoperative opioid use and pain level, psychiatric or smoking history, and perioperative factors. The included studies presented various model algorithms including regularised logistic regression, random forest, gradient boosting machines and artificial neural networks. Model performance varied from fair to excellent with the area under receiver-operator curve (AUROC) between 0.75 – 0.95. The overall risk of bias was high.
Conclusion: We have identified 10 models for the prediction of chronic opioid use following surgery with fair to excellent accuracy. These studies confirm the unrealised role of artificial intelligence in augmenting clinical decision-making. In the preoperative setting, these prediction models provide an opportunity to identify and mitigate risks associated with opioid use prior to the onset of clinical symptoms or signs. Despite this, methodological concerns of included studies and the lack of diverse patient backgrounds present a barrier to generalisation to a diverse surgical population. We recommend further research in clinical prediction modelling to facilitate a safe and evidence-based translation to clinical practice.
