Risk Prediction for Non-cardiac Surgery Using the 12-Lead Electrocardiogram: An Explainable Deep Learning Approach

Background To improve on existing noncardiac surgery risk scores, we propose a novel approach which leverages features of the preoperative 12-lead electrocardiogram (ECG) to predict major adverse postoperative outcomes. Methods Data acquired in 37,060 adult patients prior to major noncardiac surgery were used to train a series of convolutional neural network models in the task of predicting in-hospital acute myocardial infarction (MI), in-hospital mortality (IHM), and a composite of in-hospital MI, in-hospital stroke, and 30-day mortality. Preoperative ECG waveform features were first modeled as sole inputs then integrated with clinical variables in fusion models. Model discrimination was evaluated using area under the receiver operating characteristic (AUROC) analysis, and performances were compared with the Revised Cardiac Risk Index (RCRI), a benchmark preoperative risk score To gain interpretable insight, a generative approach using counterfactual ECGs was implemented. Results The ECG fusion model had an AUROC of 0.858 (95% CI [0.845, 0.872]), 0.899 (95% CI [0.889, 0.908]), and 0.835 (95% CI [0.827, 0.843]) in predicting MI, IHM, and the composite endpoint, respectively; these AUROC values were significantly higher than in models based on ECG waveforms alone (MI: p = 0.001, IHM: p < 10 −4 , composite: p < 10 −4 ). All ECG based models had significantly higher discrimination than the RCRI. Counterfactual ECG analysis revealed morphological features relevant to outcome classification. Conclusion A deep learning approach integrating preoperative ECG waveform features significantly enhances the ability to predict major outcomes after noncardiac surgery. The use of counterfactual ECGs provides plausible explanations for classifier decisions, increasing the interpretability of the models. Clinical perspective What is new? □ A deep learning approach applied to preoperative 12-lead ECG waveforms accurately predicts major outcomes after noncardiac surgery. □ This model outperforms the benchmark Revised Cardiac Risk Index (RCRI). □ The highest predictive performance was obtained with a fusion model that combines preoperative ECG waveforms with routinely collected clinical variables. □ An exploratory approach in which counterfactual ECGs are generated provides explainability for classifier decisions. What are the clinical implications? □ In adults undergoing non-cardiac surgery, ECG waveform features are predictive of postoperative cardiovascular risk □ Risk models integrating ECG waveforms with clinical variables can serve as the basis for outcome modifying interventions across the surgical continuum. □ Achieving explainability through counterfactual ECGs represents an important step towards real-world implementation.