Machine learning algorithms predict extended postoperative opioid use in primary total knee arthroplasty

• Published in: Knee surgery, sports traumatology, arthroscopy : official journal of the ESSKA Vol. 30; no. 8; pp. 2573 - 2581
• Main Authors: Klemt, Christian, Harvey, Michael Joseph, Robinson, Matthew Gerald, Esposito, John G., Yeo, Ingwon, Kwon, Young-Min
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.08.2022; Springer Nature B.V
• Subjects: Algorithms; Analgesia; Arthroplasty (knee); Artificial intelligence; Calibration; Complications; Decision analysis; Demographic variables; Drug abuse; Joint replacement surgery; Knee; Learning algorithms; Machine learning; Medicine; Medicine & Public Health; Narcotics; Opioids; Orthopaedic implants; Orthopedics; Pain; Patients; Postoperative period; Recovery; Recovery (Medical); Sports Medicine; Extended opioid use; Primary total knee arthroplasty; Artificial intelligence; Risk factors; Machine learning
• ISSN: 0942-2056; 1433-7347
• DOI: 10.1007/s00167-021-06812-4

Purpose Adequate postoperative pain control following total knee arthroplasty (TKA) is required to achieve optimal patient recovery. However, the postoperative recovery may lead to an unnaturally extended opioid use, which has been associated with adverse outcomes. This study hypothesizes that machine learning models can accurately predict extended opioid use following primary TKA. Methods A total of 8873 consecutive patients that underwent primary TKA were evaluated, including 643 patients (7.2%) with extended postoperative opioid use (> 90 days). Electronic patient records were manually reviewed to identify patient demographics and surgical variables associated with prolonged postoperative opioid use. Five machine learning algorithms were developed, encompassing the breadth of state-of-the-art machine learning algorithms available in the literature, to predict extended opioid use following primary TKA, and these models were assessed by discrimination, calibration, and decision curve analysis. Results The strongest predictors for prolonged opioid prescription following primary TKA were preoperative opioid duration (100% importance; p  < 0.01), drug abuse (54% importance; p  < 0.01), and depression (47% importance; p  < 0.01). The five machine learning models all achieved excellent performance across discrimination (AUC > 0.83), calibration, and decision curve analysis. Higher net benefits for all machine learning models were demonstrated, when compared to the default strategies of changing management for all patients or no patients. Conclusion The study findings show excellent model performance for the prediction of extended postoperative opioid use following primary total knee arthroplasty, highlighting the potential of these models to assist in preoperatively identifying at risk patients, and allowing the implementation of individualized peri-operative counselling and pain management strategies to mitigate complications associated with prolonged opioid use. Level of evidence IV.