Please confirm that an ethics committee approval has been applied for or granted: Not relevant (see information at the bottom of this page)

Application for ESRA Abstract Prizes: I don’t wish to apply for the ESRA Prizes

Background and Aims Machine learning enables complex patient data to be distilled into predictive diagnostic tools. This review identified studies that applied machine learning to predict acute, subacute, or chronic pain or opioid use after any surgical procedure.

Methods We searched PubMed using the following search strategy and terms: ‘machine learning’ OR ‘artificial intelligence’ AND ‘pain’ OR ‘opioid’ AND ‘surgery’ OR ‘postoperative’ AND ‘predict.’ The inclusion criteria were literature written in English that used machine learning and/or artificial intelligence to predict postoperative and/or opioid use after surgery. The exclusion criteria were reviews; protocol papers, commentaries; not a pain or opioid-related outcome; not a postoperative outcome; diagnostic or measurement tool.

Results Thirty-nine studies were included (figure 1). Nineteen studies (48.7%) utilized machine learning to predict the outcome of chronic postoperative pain or function after any surgical procedure, followed by 12 studies (30.8%) utilizing machine learning to predict chronic postoperative opioid use. The most common algorithms were GBDT (n = 28), random forest algorithms (n = 23) and regularization algorithms (n = 22). 27 studies (69.2%) used preoperative pain as a predictor in the initial model. 22 studies (69.2%) used preoperative pain as a predictor in the final model. 25 studies (64.1%) used preoperative opioid use as a predictor in the initial model. 19 studies (54.3%) used preoperative opioid use as a predictor in the final model.

Conclusions Machine learning can contribute to personalized perioperative pain management approaches. Patient-reported variables are important, salient predictors of acute, subacute, or chronic pain or opioid use after any surgical procedure.

Attachment ESRA 2023 Machine Learning Abstract_5.21.2023_final.pdf