Abstract Rationale In the last decade, app-enabled home spirometry has become widely used for monitoring respiratory patients. This adaptation introduced a need to ensure ATS/ERS quality requirements are met without clinical supervision. Reliability of spirometry depends on correct technique and patient effort. In clinical settings, clinicians guide patients through successful test execution. In contrast, home-based spirometry tests must be performed independently, with companion mobile applications identifying spirometry errors and indicating whether tests satisfy acceptability criteria. Unlike other errors that follow rule-based definitions (e.g., hesitation per ATS 2019), cough lacks a standardized approach. Clinicians typically detect cough through direct observation or flow-volume curve inspection, which are inapplicable in unsupervised home testing. Method To address this challenge, we developed a Convolutional Neural Network (CNN) with spatial attention mechanism. Our approach generates location-specific weights, helping the model focus on cough-indicative regions in flow-volume curves. We utilized a subset of the NHANES database (2007-2012), using 13,198 spirometry tests: 9,998 acceptable quality tests and 3,200 cough-contaminated tests. To capture spatial and temporal information, we used a novel color-coding scheme: red for the FEV1 region (0-1 second) and blue for the remaining exhalation period. This encoding enables the model to distinguish between clinically significant coughs affecting FEV1 measurements and those occurring later that may not invalidate tests. We split the dataset into 70% training, 15% validation, and 15% test sets. Results Our model achieved 97.17% accuracy, 95.21% sensitivity, 97.80% specificity, 93.27% precision, and 94.23% F1 score for cough detection. Following clinician review of misclassified samples, several cases were identified as mislabeled data, and after correction, the model’s true performance improved to 99.1% accuracy. The spatial attention module successfully identified cough patterns in the FEV1 region by assigning higher weights to disrupted flow regions, as visualized in attention heat-maps. The figure demonstrates this through input image, attention heat-map, and overlay. Figure: Example showing (a) input flow-volume spirometry curve with red/blue temporal encoding, (b) attention heat-map highlighting cough-indicative region, and (c) overlay visualization Conclusion This automated approach provides a mechanism to detect cough errors, crucial for spirometry acceptability. By ensuring reliable quality control, this technology could reduce invalid tests and improve remote respiratory monitoring outcomes. While developed primarily for home use, the model has potential applications in clinical settings, particularly in primary care where spirometry expertise may be limited. This could standardize cough detection across different healthcare settings and support less experienced operators in obtaining quality results. This abstract is funded by: Clario
Yasar et al. (Fri,) studied this question.