Development and Internal Validation of a Machine Learning Model for Predicting Long-Term Opioid Therapy After Hip...

Long-term opioid therapy (LTOT) after hip fracture surgery is a common postoperative complication associated with adverse outcomes, yet tools to identify at-risk patients among opioid-naïve older adults are lacking. This study aimed to develop and internally validate a parsimonious model to predict LTOT following hip fracture surgery.

Using Danish nationwide registries, we identified 26,057 opioid-naïve patients (≥ 65 years) undergoing hip fracture surgery (2010-2020) and analysed 29 predictors, covering demographics, comorbidities, medication, lifestyle, socioeconomics and surgical factors. LTOT was defined as redeeming ≥ 2 prescriptions between 31 and 365 days of surgery. Models were developed using four machine learning methods: logistic regression, backwards elimination, elastic net penalised logistic regression and random forest. Model performance was assessed using area under the receiver operating characteristic curve (AUC), calibration slope, intercept, Brier score and decision curve analysis.

LTOT was identified in 8095 (31.1%) patients. The backward elimination algorithm identified the best performing model, selecting 8 of 29 predictors and achieving an AUC of 0.68, calibration slope of 0.99, intercept of 0.02 and Brier score of 0.20. Predictors included age, marital status, preoperative non-opioid pain medication, preoperative novel oral anticoagulants, fracture type, surgery delay, length of hospital stay and postoperative cumulated ambulation score at discharge.

A prediction model was developed and validated for use at discharge to identify patients at risk of LTOT 1 year after hip fracture. The model may support risk stratification at discharge, but requires external validation and evaluation of clinical implementation before routine use.

This study presents the first internally validated prediction model for long-term opioid use in opioid-naïve older adults after hip fracture surgery. The model functions as a simple and interpretable risk stratification tool at discharge and has been incorporated in a freely available risk calculator. It addresses the lack of clinically applicable risk stratification tools in this frail population and highlights opportunities for more targeted postoperative pain management, although feasibility testing is required before clinical implementation.