Title:A Distributed Privacy Preserving Model for the Detection of Alzheimer's Disease

Abstract:BACKGROUND: Segmentation of medical data, concerns about personal health information (PHI) breaches, and the direct and indirect costs of consolidating and managing such segmented date should motivate diagnostic machine learning (DML) researchers to identify privacy-preserving machine learning algorithms that can train on distributed or decentralized datasets of different modalities. Federated learning models provide such a decentralized machine learning framework in which multiple investigators in possession of disparate datasets and working on different devices or servers can train collaboratively a global machine learning models without ever having to exchange local data and thus can meet statutory PHI protections. To this end, a vertical federate learning model is devised and tested for efficacy in the detection of Alzheimer's Disease (AD).
METHODS: The second version of Open Access Series of Imaging Studies -- with its panoply of demographic, imaging, and clinical assessment datasets -- was used to test a multimodal vertical federated learning (VFL) model for AD detection.
RESULTS: By training and validating this VFL model on the demographic, clinical, and MRI data in OASIS-2, an 82.9\% accuracy rate is achieved, consistent with previously reported results.
CONCLUSIONS: The VFL architecture proposed herein offers a novel distributed architecture, enabling collaborative learning across diverse sources of medical data while respecting statutory privacy constraints. By leveraging multiple modalities of data, the robustness and accuracy of AD detection can be enhanced. This model not only contributes to the advancement of federated learning techniques but also holds promise for overcoming the hurdles posed by data segmentation in medical research.