Personalized federated adaptive transformer ensemble (FATE) model for COPD detection utilizing spatial and temporal features in respiratory sound analysis

作者：Ayesha Jabbar, Jianjun Huang, Muhammad Kashif Jabbar, Tariq Mahmood · 发表于：Results in Engineering · 年份：2025 · DOI：10.1016/j.rineng.2025.107203 · 被引用次数：5 · 研究领域：Phonocardiography and Auscultation Techniques、Noise Effects and Management、Music and Audio Processing

Chronic Obstructive Pulmonary Disease (COPD) is an illness that endangers the lives of many people due to a lack of proper and early diagnosis. The issues with traditional centralized machine learning models include concerns over data privacy, data centralization, and limited generalizability. To overcome these difficulties, we introduce the Federated Adaptive Transformer Ensemble (FATE) approach, a privacy-preserving, decentralized system that integrates federated learning with Transformer-based models and ensemble techniques. The FATE model is a combination of Convolutional Neural Networks (CNNs) and Adaptive SpectroTemporal Transformers used to retrieve spatial and temporal statistics of signals or sounds of the respiratory system. There are some advanced pre-processing algorithms used to guarantee high-quality input representations: band-pass filtering, adaptive noise cancellation, wavelet-based segmentation, and feature extraction (Mel-spectrogram, MFCCs, ZCR). In keeping with data privacy mandated by GDPR and HIPAA compliance, federated training on 10 simulated healthcare institutes maintains privacy and keeps information safe by not sharing it with a third party. Expert assessment on a respiratory sound dataset of 920 samples collected on 126 patients shows that the FATE framework is extremely accurate in identifying COPD with an accuracy of 98.5 percent and higher than 20 state-of-the-art models. Ensemble learning strategy and SHAP-based interpretability solution find...