Multimodal fusion of EHR and ECG based on deep learning for predicting new-onset coronary heart disease in cancer patients

Background: Cancer patients carry elevated risk of new-onset coronary heart disease (CHD), but accurate risk stratification remains limited. We aimed to develop a multimodal deep learning model for predicting new-onset CHD in cancer patients. Methods: Consecutive cancer patients from January 2010 to December 2020 were enrolled. The primary endpoint was new-onset CHD during hospitalization or follow-up. We constructed a hybrid CNN-LSTM model by fusing electronic health records (EHRs) and 12-lead electrocardiogram (ECG). Model performance was compared with seven machine learning methods (LR, DT, KNN, SVM, RF, AdaBoost, GBDT). Clinical utility and calibration were assessed by ROC, decision curve analysis (DCA) and calibration curves. Subgroup analyses were conducted in early-stage (Stage II) and late-stage (Stage III + IV) cohorts. Results: A total of 1262 patients were included in the final analysis, of whom 722 (57.2%) developed new-onset CHD, with 696 (55.2%) classified as early-stage (Stage II). The CNN-LSTM model achieved an AUC of 0.975 (95%CI: 0.962-0.988), outperforming seven conventional machine learning models. In subgroup analysis, the model yielded an AUC of 0.924 in early-stage patients and 0.888 in late-stage patients. SHAP analysis identified that PAB, CRP, age, cTnI, PT, and ECG markers including HR, QT and QTc intervals were the strongest predictive features. Conclusion: The multimodal CNN-LSTM model enables robust and interpretable prediction of new-onset CHD in cancer patients. Coagulation, myocardial injury, inflammatory, metabolic, and electrophysiological markers provide mechanistic insights and may facilitate early cardioprotective strategies.