Federated Learning for Predicting Mild Cognitive Impairment to Dementia Conversion

TL;DR

This study tackles the privacy challenge of predicting progression from Mild Cognitive Impairment to dementia by applying Federated Learning to cognitive and socio-demographic data from the ADNI dataset. It compares centralized ML, site-specific training, and two FL architectures—client-server with FedAvg and P2P with FedGD—under horizontal FL with 10-fold cross-validation across 3, 4, and 5 year follow-ups. The results show FL achieves performance comparable to centralized ML and superior generalization to site-specific models, with FedGD offering benefits in collaborative normalization and FedAvg delivering computational efficiency. The work demonstrates that privacy-preserving FL can enable scalable, non-invasive dementia prognosis across multiple sites without sharing sensitive data.

Abstract

Dementia is a progressive condition that impairs an individual's cognitive health and daily functioning, with mild cognitive impairment (MCI) often serving as its precursor. The prediction of MCI to dementia conversion has been well studied, but previous studies have almost always focused on traditional Machine Learning (ML) based methods that require sharing sensitive clinical information to train predictive models. This study proposes a privacy-enhancing solution using Federated Learning (FL) to train predictive models for MCI to dementia conversion without sharing sensitive data, leveraging socio demographic and cognitive measures. We simulated and compared two network architectures, Peer to Peer (P2P) and client-server, to enable collaborative learning. Our results demonstrated that FL had comparable predictive performance to centralized ML, and each clinical site showed similar performance without sharing local data. Moreover, the predictive performance of FL models was superior to site specific models trained without collaboration. This work highlights that FL can eliminate the need for data sharing without compromising model efficacy.

Figures (5)

• Figure 1: Empirical graphs visualizing the FL architectures proposed to maintain data privacy: P2P (left panel) and client-server (right panel). Nodes represent the ADNI sites, and edges enable the connectivity among nodes. In client-server architecture, the star sign represents the server or computational node.
• Figure 2: Data distribution at ADNI sites based on the clinical status of an individual, where sMCI (label 0) and pMCI (label 1) are visualized with green and orange bars, respectively. From left, 3, 4, and 5 years of follow-up.
• Figure 3: Effect of collaborative parameters on site-specific models for Case 1: predicting cognitive status 3 years from baseline. The left panel illustrates the performance of the P2P architecture using the FedGD algorithm, where the parameter $\alpha \in [0, 1]$ regulates collaboration between local and neighboring nodes. The right panel shows the client-server architecture with the FedAvg algorithm, where the parameter $\eta$ controls gradient regularization. A higher value of $\alpha$ and a smaller value of $\eta$ constrain local updates to remain closer to those of neighboring nodes or the server, respectively.
• Figure 4: Performance metrics comparison of traditional ML (centralized and site-specific) and federated methods (client-server FedAvg and P2P FedGD) for 3-year advanced prediction of MCI to dementia conversion from baseline. The left panel presents AUC ROC scores and depicts the similarity of centralized ML to FL and the performance improvement over site-specific results after collaboration, regardless of participant count and site class distribution. The bar plots in the background represents the participant count where greens represent the number of participants with label 0: sMCI and orange depicts label 1: pMCI. The right panel visualizes the scores for sensitivity and specificity values. Metrics are averaged across all folds, with parameter values $\alpha = 1$ and $\eta = 0.1$ in the FedGD and FedAvg algorithms.
• Figure 5: Effect of federated and site-specific data normalization at $\alpha = 1$ in the FedGD P2P architecture. The scores are represented as the mean of all folds.