Privacy-Preserving Multimodal Clinical Trial Matching Using Homomorphic Encryption

Patient privacy and regulatory concerns limit sharing of clinical data across institutions, which creates a major barrier to clinical trial recruitment, a currently manual process that places the burden on patients and medical providers to find suitable trials. This issue is particularly pronounced for rare and sensitive diseases that require aggregation of geographically distributed patient cohorts. We developed a privacy-preserving framework for multimodal clinical trial matching using homomorphic encryption, enabling eligibility evaluation directly on encrypted patient representations. The framework integrates structured clinical variables with encrypted image-text similarity calculation over embeddings from pretrained medical foundation models, enabling zero-shot classification in the encrypted domain. In simulated trial-matching workflows comprising up to 10,000 patients, 100 clinical trials, and 36 eligibility variables, encrypted evaluation reproduced plaintext eligibility classifications with 99.9982% concordance across 1,000,000 patient-trial pairs. In this setting, encrypted index setup took 13.39 seconds, and matching required 0.28 seconds per trial and 28.44 seconds in total, compared to 3.31 seconds in total for plaintext matching. The encrypted index required 177 MiB to store 10,000 records. Runtime scaled approximately linearly with cohort size. These findings demonstrate the feasibility of privacy-preserving multimodal clinical trial matching, which could securely and efficiently facilitate initial screening of eligible patients to improve clinical trial yield.