Beyond the ‘black box’: choosing interpretable machine learning models for predicting postoperative opioid trends

Artificial intelligence encompasses machine learning and is a popular, yet controversial, topic in healthcare. Recent guidelines from national regulatory agencies underscore the critical importance of interpretability in machine learning models used in healthcare [1]. ‘Interpretability’ means that clinicians understand the reasoning behind a model's predictions, fostering trust and enabling informed clinical decision-making [Doshi-Velez et al. preprint, https://arxiv.org/abs/1702.08608]. In response to the opioid epidemic, there has been interest in using machine learning models to predict which patients will have the highest risk of postoperative opioid dependence. To be interpretable, clinicians should be able to see which specific factors (e.g. previous opioid use, type of surgery or mental health conditions) contribute to prediction. Experts have advocated for building inherently interpretable models from the start, especially in high-stakes medical contexts, rather than retrofitting explanations onto complex models after development [2]. As machine learning algorithms become integral to peri-operative management, balancing model complexity with interpretability is crucial [3]. The objective of this study was to evaluate whether simpler, more interpretable models could match complex ones in predictive accuracy and in identifying key predictors for postoperative opioid use.

Following institutional review board approval, we conducted a retrospective cohort study at a US Veterans Affairs hospital. We included adult patients who had surgery from 2015 to 2021 and had documented pre-operative and post-discharge opioid prescriptions. Patients without complete opioid prescription data were not studied.

Baseline data were extracted from electronic health records and included: patient characteristics; clinical variables (such as type of surgery and duration of hospital stay); and mental health diagnoses. We assessed three outcomes, with mean daily morphine milligram equivalents (MME) as the primary outcome and variance in MME and monthly rate of change in MME as secondary outcomes; these were all measured over 12 months before surgery and post-discharge. Opioid prescriptions were converted to MME, and mental health diagnoses were identified using ICD-10 revision codes as described in previous studies [4].

We developed three machine learning models to predict postoperative opioid use: lasso regression, which enhances accuracy and interpretability through variable selection and regularisation; decision tree, which predicts outcomes using interpretable decision rules inferred from data; and extreme gradient boosting (XGBoost), an ensemble method known for high predictive performance but lower interpretability [5].

Analyses were performed using RStudio (version 12.0, R Foundation for Statistical Computing, Vienna, Austria) involving two scenarios: models were trained using only baseline predictors without pre-operative opioid use data; and models included all baseline predictors plus pre-operative opioid use metrics. We utilised the interpretable machine learning package for feature importance analysis, with the rpart and XGBoost packages used for model implementation. Hyperparameters were optimised via grid search and cross-validation. Ten-fold cross-validation minimised overfitting and assessed generalisability. The primary evaluation metric was root mean squared error (RMSE) and mean absolute error (MAE) was also calculated. Feature importance was determined by coefficient magnitude (lasso regression), tree structure (decision tree) and built-in importance measures (XGBoost), with p < 0.05 defined as statistically significant.

The study cohort consisted of 1396 patients who were predominantly male (93.6%), aged > 70 y (49.4%) and White (77.4%) (online Supporting Information Table S1). Half of the cohort had a diagnosed mental illness, with major depression (58.2%) and substance use disorder (27.9%) being most prevalent. Surgery type varied, with orthopaedics (19.3%) and ophthalmology (9.3%) being common. The mean (SD) pre-operative MME was 681 (1340), indicating significant opioid use before surgery.

Including pre-operative opioid metrics enhanced predictive accuracy across all models (Table 1). Lasso regression showed the greatest improvement (RMSE 1263 to 711, MAE 726 to 350, p < 0.01), followed by decision tree (RMSE 1286 to 787, MAE 709 to 363, p < 0.01), while XGBoost showed modest improvements (RMSE 1352 to 1168, MAE 600 to 528, p < 0.05). For secondary outcomes (online Supporting Information Table S2), models showed modest improvements in predicting opioid use variance, with XGBoost performing best (RMSE 2,540,888 to 2,299,404), while improvements in predicting monthly rate of change were minimal.

Feature importance analysis revealed differences among models (online Supporting Information Figure S1). While XGBoost heavily weighted pre-operative mean MME, emphasising reliance on previous opioid use patterns, decision tree and lasso regression identified additional important predictors. Decision tree highlighted surgical type and duration of hospital stay alongside pre-operative opioid metrics, while lasso regression emphasised mental health diagnoses and duration of hospital stay as influential predictors.

Our study shows that simpler models can predict postoperative opioid trends effectively and provide valuable insights into key predictors [6]. Notably, lasso regression and decision tree models identified clinically relevant factors beyond opioid use history, achieving comparable accuracy while offering greater potential interpretability. The lack of transparency in complex models may limit clinical adoption and need further evaluation.

The Veterans Affairs healthcare population is known to have higher prevalence rates of mental illness, pre-operative opioid use and substance use disorders compared with typical surgical populations in the USA. Recent studies of general surgical populations in the USA report pre-operative opioid use rates of 10–30% and mental health diagnosis rates of 10–35%, compared with rates in our cohort of 78% and > 50%, respectively [6-8]. The 94.1% prevalence of chronic pain in our cohort is also notably higher than in general surgical populations (typically 25–40%). These characteristics of US veterans, particularly among those seeking surgical care at VA healthcare facilities, are important to note [4, 7]. While these features may limit the generalisability of any inferences, from the perspective of our study purpose, higher prevalence rates contribute to a richer dataset for evaluating and comparing the ability of these models to identify complex predictor relationships. Based on our results, prioritising simpler models and interpretability may enhance clinical utility without compromising performance [9]. Multicentre evaluation involving more diverse surgical populations will be necessary to validate these findings and assess model interpretability needs across different clinical settings.