Federated simplex-constrained network for fine-grained clock drawing test scoring and cognitive pattern representation.

Abstract

The clock drawing test is a core tool for the early screening of cognitive impairment. Traditional manual methods are time-consuming and prone to significant subjective bias, while most existing automated methods are 'black-box' approaches that directly map images to scores or category labels. Not only do these methods lack interpretability in their decision-making processes, but they also fail to provide information beyond the raw scores. Furthermore, most of these methods employ a centralized training paradigm. However, due to strict privacy regulations that restrict medical data sharing, achieving multi-center collaboration remains difficult, which hinders clinical deployment. Therefore, this paper proposes the federated simplex-constrained network, which enables collaborative training through federated learning without raw data leaving the local clients; simultaneously, by embedding a simplex-constrained module in the model, a module that serves both as a latent-feature transformation layer and as a cognitive-pattern modeling module, it represents individual cognitive patterns of samples as convex combinations of multiple cognitive prototypes, thereby constructing a task-relevant library of cognitive prototypes. This approach simultaneously achieves fine-grained scoring, interpretability of the decision-making process, and cognitive pattern analysis. Performance experiments were conducted on a dataset of 1370 cross-sectional clock-drawing images collected from a single institution, using the Dirichlet partitioning method to simulate multi-federated clients with highly skewed label distributions. The main experimental results show that this model achieved a macro-average accuracy of 0.847 (95% CI: 0.841-0.854) and a macro-averageF1 score of 0.470 (95% CI: 0.405-0.536) on the scoring task.