Secure and Private Vertical Federated Learning for Predicting Personalized CVA Outcomes

Cerebrovascular accident (CVA) outcome predictions could improve patient-centered care by informing individual patients on rehabilitation and expected outcomes. However, CVA patients' data is vertically distributed across hospitals and rehabilitation clinics. Centralizing distributed medical data in a central repository leads to difficulty concerning data privacy and data ownership. Vertical federated learning has been introduced as a solution, but it is not secure. We introduce our secure vertical federated learning (SVFL) protocol that prevents label and data leakage through encrypted active-party backpropagation. We use this to produce the first CVA outcome model using hospital and rehabilitation data in a vertically federated setting. Data from 825 CVA patients admitted to the St. Antonius Hospital, the Netherlands was collected, including their rehabilitation trajectory in three clinics, to predict functional status (dichotomized mRS score) after 3 months. Our results show that a model trained on the vertically integrated hospital and rehabilitation data performs better than a model trained on either of these sets alone. Training using SVFL yields a slightly lower predictive performance compared to training on a fully centralized data set. No difference in predictive performance between secure and unsecured VFL was observed, although secure VFL is computationally more expensive. This highlights that SVFL is a promising alternative for situations where it is not possible (or desired) to centralize vertically partitioned data.