ChemoCardioNet: An Explainable Multimodal Transformer for Early Prediction of Chemotherapy-Induced Cardiotoxicity

Chemotherapy-induced cardiotoxicity (CIC) remains a major cause of morbidity and mortality among cancer survivors, and conventional monitoring often fails to detect early subclinical cardiac injury. We propose ChemoCardioNet, an explainable multimodal deep learning framework that predicts cardiotoxicity risk before clinical manifestation by integrating electrocardiograms (ECG), echocardiography, clinical variables, and serum biomarkers. The architecture combines modality-specific encoders, including 1D convolutional transformers for ECG sequences, CNN-ViT hybrid blocks for echocardiographic frames, and multilayer perceptrons for clinical and biomarker features. A cross-attention fusion transformer aligns latent representations across modalities and captures temporal evolution across chemotherapy cycles. The model outputs a probabilistic cardiotoxicity risk score with feature-level explainability using SHAP analysis. In a cohort of 1,200 patients, ChemoCardioNet achieved an AUC of 0.92, outperforming single-modality and traditional machine learning approaches by 8-12%. The model predicted cardiotoxicity approximately two cycles before detectable echocardiographic dysfunction, supporting earlier risk identification and improved cardio-oncology monitoring.