Vertical Split Learning - an exploration of predictive performance in medical and other use cases

In healthcare and other fields, data of an individual is often vertically partitioned across multiple organizations. Creating a centralized data store for AI algorithm development is cumbersome in such cases because of concerns like privacy and data ownership. Methods of distributed learning over vertically partitioned data could offer a solution here. While several studies have evaluated the feasibility, privacy and efficiency of such methods, an extensive evaluation of their impact on predictive performance compared to a centralized approach is missing. Vertical Split Learning (VSL) aims to provide vertical distributed learning through distributed neural network architectures. Our study adapts and applies VSL to 8 datasets, both in medicine and beyond, evaluating the impact of different network and (vertical) feature distributions on predictive performance. In most configurations VSL yields comparable predictive performance to its centralized counterparts. However, certain data and network distributions give an unexpected and severe loss of performance. Based on our findings we give some initial recommendations under which conditions VSL can be applied as a suitable alternative for data centralization.