Vertical Federated Learning for Cerebrovascular Accident Outcome Prediction

Accurate and early predictions of CVA outcomes are critical for informing patients and healthcare professionals and help with creating personalized rehabilitation plans that aid in guiding treatment decisions and improving patient outcomes. Artificial Intelligence (AI) can be used to provide personalized predictions of stroke outcomes, but CVA patients typically have their data distributed across multiple care institutions during treatment and recovery. To avoid the privacy, security, and data ownership issues that come with centralizing medical data, federated learning (FL) can offer an alternative. In FL, the model is trained cooperatively, by bringing models to the data instead of data to the models, allowing the data of each party to remain at the source. However, in case of CVA outcome prediction, or in any other scenario where patient data are distributed in a chain of care (also known as vertically partitioned data), FL has additional challenges. The aim of this thesis is to provide personalized outcome predictions for the rehabilitation of patients who suffered from a CVA, by utilizing federated learning on vertically partitioned data in an accurate, secure, and clinically implementable manner. With regard to accuracy, we see that developments of more complex neural networks and frameworks, such as multimodal networks and vertically federated learning, allow better use of data and greater accuracy in prediction of stroke outcomes. Vertical federated learning might lead to a slight drop in predictive performance compared to centralized learning, but this loss is generally minimal. We show that secure vertical federated learning is a solution for dealing with vertically partitioned stroke outcome data. We introduce our SVFL framework to create the first CVA outcome model using hospital and rehabilitation data in a vertically federated setting. Our framework prevents label and data leakage through encrypted active-party backpropagation and outperforms a model built on data from a single care institute only. To achieve clinical implementability of the prediction model, the most important factors to consider are good reliability and clear communication of relevance, as these will be essential for the wide adoption of such a prediction model. In all, this thesis explores the development and implementation of a clinically relevant outcome prediction algorithm for situations where data are vertically partitioned and a centralized data repository is not desired or not possible.