A heterogeneous ensemble learning framework for detecting acute viral respiratory infections via multisource data fusion

Abstract

Acute viral respiratory infections (AVRIs), including COVID-19 and SARS, present ongoing public health challenges. Traditional diagnostic methods often rely on limited inputs—typically cough or speech sounds—rarely incorporate self-reported symptoms, and still require clinical visits or medical imaging, leading to higher costs and increased exposure risks. To support affordable, contact-free self-testing, we introduce the heterogeneous ensemble framework with multisource data fusion mechanism. This framework integrates respiratory, cough, and speech audio data with self-reported information. Within this system, deep learning models are employed to extract and analyze audio features for estimating infection probability, while conventional machine learning classifiers handle the self-reported data. Finally, predictions from both data types are combined using a meta-learning strategy to enhance overall detection performance. Experiments on COVID-19 detection validate the proposed framework, and the interpretability analysis reveals the critical predictive features and classifiers. This multisource fusion strategy also provides a transferable foundation for detecting other AVRIs.