Predicting opioid consumption after surgical discharge: a multinational derivation and validation study using a foundation model

Abstract

Opioids are frequently overprescribed after surgery. We applied a tabular foundation model to predict the risk of post-discharge opioid consumption. The model was trained and internally validated on an 80:20 training/test split of the 'Opioid PrEscRiptions and usage After Surgery' (ACTRN12621001451897p) study cohort, including adult patients undergoing general, orthopaedic, gynaecological and urological operations (n = 4267), with external validation in a distinct cohort of patients discharged after general surgical procedures (n = 826). The area under the receiver operator curve was 0.84 (95% confidence interval [CI] 0.81-0.88) at internal testing and 0.77 (95% CI 0.74-0.80) at external validation. Brier scores were 0.13 (95% CI 0.12-0.14) and 0.19 (95% CI 0.17-0.2). Patients with a <50% predicted risk of opioid consumption consumed a median of 0 oral morphine equivalents in the first week after surgery. Applying this model would reduce opioid prescriptions by 4.5% globally, and counterfactual modelling suggests without increasing time in severe pain (-4.3%, 95% CI -17.7 to 8.6).

Fig. 1. Model predictive performance plots.

A, B Receiver operator curve (ROC), C, D calibration plot, E, F precision-recall (PR) curve, and G, H decision curve analysis of our model during internal (blue) and external (orange) validation. The shaded area represents 95% confidence intervals.

Fig. 2. Box plots depicting total milligram morphine equivalents consumed stratified by patients predicted to be <50% and >50% risk of consuming an opioid in the first week after surgical discharge.

Left: internal validation (blue); right: external validation (orange). The box represents the interquartile range (IQR), and the whiskers extend to 1.5× the IQR, with outlier dots plotted beyond the whisker extents.

Fig. 3. Global opioid reduction modelling.

Modelled global reductions in average oral morphine equivalents (OMEs) prescribed based on country-level case volume and predicted reductions in OMEs. Darker shades of blue indicate greater reductions in OME prescribed, while countries with missing data are shown in white.