Multimodal deep learning to predict postoperative major adverse cardiac and cerebrovascular events after non-cardiac surgery
Hyun-Kyu Yoon, Jang Ho Ahn, Byeol Yi Kim, Hyeonhoon Lee, Woo-Young Jo, Soo-Hyuk Yoon, Hee-Pyoung Park, Hyung-Chul LeeBackground:
Major adverse cardiovascular and cerebrovascular events (MACCEs) after non-cardiac surgery can lead to substantial morbidity, mortality, and healthcare costs. Therefore, accurate and rapid risk prediction is crucial for targeted perioperative management. This study aimed to develop and validate a minimally burdensome multimodal deep learning model integrating demographic data, the International Classification of Diseases (ICD)-10 procedure codes, and raw preoperative 12-lead electrocardiogram (ECG) waveforms to predict 30-day MACCEs and to compare its performance with the established risk indices.
Materials and Methods:
This retrospective cohort study at a single tertiary academic center included adult patients who underwent non-cardiac surgery under regional or general anesthesia from 2006 to 2020. Preoperative 12-lead ECGs were acquired within 3 months before surgery. A transformer-based deep neural network processed raw ECG signals, while a gradient boosting machine (GBM) combined ECG-derived latent features with basic demographic variables (age, sex) and simplified ICD-10 procedure codes. The primary outcome was 30-day MACCEs (cardiac arrest, acute myocardial infarction, congestive heart failure, new arrhythmia, angina, stroke, or cardiovascular/cerebrovascular death). Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), precision-recall curves, sensitivity, specificity, F1 scores, and calibration metrics.
Results:
Among the 165,577 cases, 54.5% were female, the median age was 56 years, and 0.6% developed 30-day MACCEs. The multimodal GBM model demonstrated a significantly higher AUROC of 0.902 (95% confidence interval [CI], 0.898–0.906) than the baseline GBM (0.842 [0.838–0.847]). It also outperformed the Revised Cardiac Risk Index (0.813 [0.782–0.843]) and the American Society of Anesthesiologists class (0.759 [0.726–0.792]).
Conclusion:
A multimodal deep learning model combining raw ECG waveforms with minimal clinical data yielded superior 30-day MACCE risk prediction compared to that of the conventional indices. This approach could facilitate broad clinical adoption by minimizing data collection requirements while enhancing perioperative risk stratification.