PIVOT: Patient Identity via Unified Modal Twin Learning

Reliable healthcare AI is constrained by a simple but unresolved problem: the data needed to model a complete patient rarely exist in one place. Electronic health records, physiological time series, and medical images are routinely collected, yet seldom available together for the same patient because of fragmented systems, missing observations, privacy restrictions, and costly multimodal collection. Existing synthetic data methods often generate each modality independently, producing samples that appear realistic in isolation but describe clinically inconsistent versions of the same patient. We propose ModalTwin, a unified framework that learns a shared patient level latent representation from partially observed data using modality dropout encoders, Product of Experts fusion, and an InfoNCE cross modal consistency objective. ModalTwin jointly performs missing modality imputation and multimodal synthesis, generating complete twins whose modalities reflect one underlying clinical state. On 10,000 Symile MIMIC patients spanning EHR, 12 lead ECG, and chest X-ray, with up to two modalities missing, ModalTwin achieves a cross modal consistency score of 0.119 versus near zero for independent generators and a TSTR imputation AUROC of 0.652, exceeding the real data training baseline by 0.019 and outperforming three single modality VAE baselines. Membership inference AUROC is 0.492, indicating chance level attack performance. By replacing independent realism with patient level coherence, ModalTwin addresses a central bottleneck in healthcare AI.