Federated Prototype Learning for Cardiovascular Diseases Recognition Under Non-Independent and Non-Identically Distributed Heart Sounds

Artificial intelligence (AI)-powered heart sound auscultation offers a noninvasive and accessible approach for diagnosing cardiovascular diseases (CVDs). However, training robust diagnostic models requires large-scale data, which is hindered by strict privacy regulations across isolated medical institutions. While federated learning (FL) addresses privacy concerns, real-world medical data exhibits severe non-independent and non-identically (Non-IID) characteristics, including label heterogeneity, intraclass representation diversity, and interclass imbalance. Existing FL methods primarily focus on personalized client models and struggle to construct a comprehensive global model capable of recognizing all disease subtypes under such extreme data heterogeneity. To overcome these challenges, we propose a novel two-stage decoupled federated prototype learning framework. In the first stage, we introduce a multicluster trainable prototype aggregation mechanism equipped with comprehensive local feature space calibration—incorporating L_{2} normalization, learnable prototype radii, multilevel alignment, and adversarial domain adaptation—to dynamically handle label skew and intraclass diversity. In the second stage, we design a decoupled joint training paradigm that reconstructs a unified global classifier through privacy-preserving one-shot feature sharing and prior-based logit adjustment. Comprehensive experiments on the Yaseen valvular heart diseases (VHDs) and ZCHSound congenital heart diseases (CHDs) datasets demonstrate the advantages of our method. Specifically, our framework effectively narrows extreme performance disparities among individual client models and successfully constructs a robust, reusable global model that outperforms traditional centralized and federated baselines under severe class imbalance, all while ensuring minimal communication overhead.