While the proliferation of data-driven machine learning approaches has resulted in new opportunities for precision healthcare, there are a number of challenges associated with fully utilizing medical data, for example partly due to the heterogeneity of data modalities in electronic health records. Moreover, medical data often sits in data silos due to various regulatory, privacy, ethical, and legal considerations, which complicates efforts to fully utilize machine learning. Motivated by these challenges, we focus on clinical care—length of stay prediction and propose a Multimodal Federated Learning approach. The latter is designed to leverage both privacy-preserving federated learning and multimodal data to facilitate length of stay prediction. By applying this approach to a real-world medical dataset, we demonstrate the predictive power of our approach as well as how it can address the earlier discussed challenges. The findings also suggest the potential of the proposed multimodal federated learning approach for other similar healthcare settings.
Wang, Tongnian; Guo, Yuanxiong; and Choo, Kim-Kwang Raymond, "Enabling Privacy-Preserving Prediction for Length of Stay in ICU - A Multimodal Federated-Learning-based Approach" (2023). ECIS 2023 Research Papers. 238.