Adaptive federated learning with differential privacy for multi-class respiratory disease recognition from lung sound recordings

Abstract

Introduction The classification of respiratory sounds remains a challenging task in clinical research, limited by severe dataset imbalance and strict privacy regulations. Traditional machine learning methods are centralized and jeopardize data privacy and have difficulty in identifying minority classes, which makes them unsuitable for real-world clinical use. Methods In this paper, a novel federated learning framework for multi-class respiratory disease classification is proposed based on the fusion of a Hybrid Support Vector Machine-K-Nearest Neighbor (SVMKNN) classifier with Adaptive Federated Aggregation (AFA), Synthetic Minority Oversampling Technique (SMOTE) balancing, and a Gaussian differential privacy (DP) mechanism. The suggested framework runs using five non-independently and identically distributed (non-IID) client nodes partitioned using Dirichlet distribution (α = 1.0) and able to perform privacy-preserving collaborative training without sharing raw data. We extract a rich multi-modal acoustic feature representation from respiratory recordings using Mel-Frequency Cepstral Coefficients (MFCCs), first and second order delta-MFCCs, Mel spectrograms, chroma features, spectral contrast, and a set of seven scalar acoustic descriptors. The resulting high-dimensional feature vector is compressed using Principal Component Analysis (PCA) to retain 95% cumulative variance. The hybrid classifier combines the discriminative boundary learning of SVM with the local density estimation of KNN through a weighted probability ensemble (α = 0.6), resulting in better generalization in federated, class-imbalanced settings. A major methodological innovation is the restriction of the differential privacy noise to the aggregation phase, so that all reported metrics are based on clean, noise-free model outputs. Result The suggested framework obtains an accuracy of 97.55, a weighted F1-score of 97.64, Matthews Correlation Coefficient (MCC) of 90.89, and Cohen's Kappa of 90.77. Discussion Ablation studies confirm that federated hybrid learning with AFA and SMOTE outperforms standalone SVM, standalone KNN, and centralized hybrid baselines for all evaluation metrics, proving the clinical viability of the proposed framework for distributed hospital networks and telemedicine platforms.