Cough sounds are accessible, noninvasive biomarkers that can be recorded using consumer-grade smartphones and used to assess respiratory disease. Prior approaches frequently focus only on cough detection or depend on server-based deep learning for disease classification, which limits deployability and introduces privacy concerns. In addition, available cough datasets are often small and class-imbalanced, undermining model generalizability and the reliability of automated classification for conditions such as asthma and COVID-19.
The study aims to develop a multilayer, smartphone-compatible AI framework that performs automated cough detection and respiratory disease classification on-device. A central goal is to introduce a generative data augmentation strategy that mitigates severe data scarcity and class imbalance by using a suite of Variational Autoencoder (VAE) variants and a probabilistic cough-level fusion mechanism, while preserving acoustic verifiability of synthetic samples.
The proposed system integrates three AI modules to support on-device operation and clinical interpretability:
The CDM performs real-time segmentation of cough events from continuous recordings. It uses compact, lightweight models tailored for on-device inference to ensure low computational and memory demands. The module is described as providing reliable segmentation across heterogeneous recording conditions, enabling downstream classification and augmentation steps to operate on properly segmented cough events without requiring server-side processing.
The DAM classifies segmented cough events into asthma, COVID-19, or healthy categories. Classification relies on compact cepstral and spectral feature representations rather than large-scale deep models, enabling efficient execution on commodity hardware. Parallel Support Vector Machine classifiers are used for event-level discrimination. To aggregate evidence across multiple coughs, the system employs a probabilistic cough-level fusion strategy, which combines per-event model outputs into a subject-level or session-level inference.
The GAM is composed of five distinct Variational Autoencoder variants. Its novelty lies in operating across the time-frequency domain for latent feature optimization while reconstructing samples back into the time domain. This cross-domain approach is intended to optimize latent representations where discriminative features are often clearer, while producing time-domain waveforms that can be acoustically reviewed for clinical verification. The GAM addresses class imbalance by synthesizing additional samples for underrepresented classes and complements conventional audio augmentation techniques that may be insufficient under severe data scarcity.
All components of the framework—CDM, DAM, and GAM—are implemented to run in real time on commodity Android devices. The design choices emphasize lightweight models and compact feature sets to enable on-device inference, which reduces dependency on network connectivity and mitigates privacy risks associated with uploading raw audio to remote servers.
According to the reported findings, the CDM achieved reliable segmentation across heterogeneous recording conditions. The DAM demonstrated strong discriminability between asthma, COVID-19, and healthy coughs when using compact cepstral and spectral features together with parallel SVM classifiers and probabilistic fusion. The multi-variant GAM framework was reported to be effective in alleviating class imbalance under conditions of severe data scarcity. The cross-domain reconstruction approach allowed for the clinical verification of synthetic cough biomarkers, since reconstructions were produced in the time domain for acoustic review. The authors emphasize that all system components operate in real time on commodity Android hardware.
The framework is designed to support clinical verification of generated cough biomarkers by reconstructing synthetic samples in the time domain, facilitating acoustic inspection by clinicians or domain experts. On-device execution addresses deployability and privacy concerns by avoiding continuous transfer of raw audio to remote servers. The proposed approach aims to improve model generalizability in low-data settings while maintaining interpretability of the augmentation process.
The integrated, smartphone-compatible framework addresses key limitations in dataset availability, model generalizability, and deployability for cough-based respiratory disease analytics. By combining lightweight on-device cough detection, SVM-based disease classification with probabilistic fusion, and a multi-variant VAE generative augmentation strategy that preserves time-domain verifiability, the study demonstrates the feasibility of scalable, privacy-preserving smartphone-based acoustic sensing for respiratory health monitoring. Details on quantitative performance metrics and dataset composition were not reported in the abstract and would require consultation of the full text for further specifics.
The study highlights methods and concepts including Variational Autoencoders, Support Vector Machines, on-device machine learning, generative data augmentation, cough sound analysis, and smartphone-based respiratory health monitoring. The authors declare no competing interests in the conflict of interest statement.