Modality Dropout for Robust Multimodal Cardiovascular Risk Prediction Under Missing Clinical Data

Multimodal models that combine ECG, cardiac imaging, and electronic health record (EHR) data consistently outperform single-modality systems, but nearly all published architectures assume every input is present at inference time. In real hospital settings this assumption rarely holds: imaging may be unavailable, or a lab value may be missing. This paper studies modality dropout, a training-time regularisation technique that randomly zeroes out entire modality embeddings during training, as a mechanism for closing this gap. We sweep the dropout probability from 0.0 to 0.50 across a multimodal cardiovascular disease (CVD) classifier built from a 1D-CNN ECG encoder, a fine-tuned ResNet-50 imaging encoder, and an XGBoost EHR encoder, and compare the resulting robustness against three alternative missing-data strategies: zero-imputation, mean-imputation, and a learned missing-token embedding. At the selected operating point (p = 0.25), full-modality accuracy reaches 97.4% while single-modality (ECG-only) accuracy holds at 93.4%, a mean degradation of only 2.5 percentage points across all missing-input scenarios — compared to 35.3 points for a model trained without any missing-data mitigation. The results suggest that modality dropout is not just a regulariser but a practical requirement for any multimodal clinical AI system intended for real-world deployment.