Interpretable machine learning models for predicting perioperative myocardial injury in non-cardiac surgery

Figures

• 

  Figure 1. Discriminative performance of machine learning models and clinical scores for predicting perioperative myocardial injury. Bootstrapped receiver operating characteristic (ROC) curves comparing the Explainable Boosting Machine, XGBoost, Random Forest, logistic regression, and the modified Revised Cardiac Risk Index in the temporally independent validation cohort (2022–2023). The EBM achieved the highest area under the ROC curve (AUC), with machine learning models consistently outperforming the mRCRI across the full range of false-positive rates.

• 

  Figure 2. Calibration of machine learning models for predicting perioperative myocardial injury. Calibration plots comparing observed and predicted event rates across deciles of predicted risk in the temporally independent validation cohort (2022–2023). The Explainable Boosting Machine, XGBoost, Random Forest, and logistic regression models demonstrated good calibration across most risk ranges. Slight deviations at the highest risk deciles reflect the limited number of patients with very high predicted risk. Overall, the machine learning models showed consistent alignment between predicted probabilities and observed outcomes.

• 

  Figure 3. Feature importance and effect visualization for perioperative myocardial injury prediction. (A) Relative importance of the top predictors in the Explainable Boosting Machine model, expressed as percentage contribution to the model. Colours indicate laboratory values (orange), diagnoses (blue), point-of-care diagnostics (purple) and patient characteristics (grey). (B, D, F) Shape functions showing the marginal effect of age (B), leucocyte count (D), and pre-operative haemoglobin (F) on the log-odds of predicted perioperative myocardial injury risk. Shaded areas represent 95% confidence intervals. (C, E, G) Corresponding distributions of these features in patients with (yellow) and without (grey) perioperative myocardial injury. Non-linear risk patterns were observed for age and leukocyte count, while haemoglobin demonstrated a marked inverse non-linear association with increased risk at lower concentrations.

• 

  Figure 4. Comparison of EBM-derived risk predictions and modified revised cardiac risk Index categories. (A) Distribution of predicted risk probabilities from the EBM across categories of the modified RCRI. (B) Cross-tabulation of patients stratified into low and high risk by the EBM and the modified RCRI. The EBM identified additional high-risk patients not captured by the RCRI.

• 

  Figure 5. Risk distribution and stratification performance of the EBM model vs. guideline-based classification. (A) Distribution of patients with and without PMI across the EBM-predicted risk spectrum. Non-PMI cases (grey) cluster at the lower end of the predicted risk range, while PMI cases (orange) are increasingly enriched at higher predicted risk levels, demonstrating the discriminative capacity of the model. (B) Observed PMI rates and corresponding Number Needed to Evaluate (NNE) across low- and high-risk categories, comparing the EBM model (blue) with the ESC guideline-based clinical practice guideline (CPG) approach (grey). In the high-risk category, the EBM identified patients with a higher observed PMI rate (33.3%; NNE 3.00) compared with guideline-based stratification (28.6%; NNE 3.5). Values indicate observed PMI rates, NNE, and patient counts in the temporally independent validation cohort (2022–2023).

• 

  Figure 6. Conceptual framework for clinical integration and federated learning of the perioperative myocardial injury prediction model. (A) Example of real-time clinical integration using an electronic medical record (EMR) system, which harmonizes patient data into a FHIR store. The Explainable Boosting Machine model retrieves patient data, calculates perioperative myocardial injury risk, and transmits risk predictions and explanations to an early warning system for clinical monitoring. (B) Federated learning architecture across three independent clinical sites. Each site harmonizes local EMR data, trains a local instance of the machine learning model, and exchanges only model parameters not patient data enabling collaborative model refinement while preserving data privacy.