Abstract TH849: Machine Learning-Based Magnetic Respiratory Sensing Technology and Hypergraph Network Modeling of Respiratory and Cardiovascular Comorbidity

Abstract

Background: Cardiovascular diseases (CVDs) are a leading cause of death globally, and coexisting respiratory diseases (RDs) amplify clinical risk and care complexity. A noninvasive approach that both classifies respiratory disease from bedside signals and characterizes CVD-RD comorbidity could improve early diagnosis, triage, and longitudinal management. Hypothesis: We hypothesize that features derived from breath signals can accurately distinguish healthy individuals from multiple RDs using machine learning (ML), and that high-order comorbidity network modeling can reveal age-stratified patterns linking RDs with CVDs. Methods: Magnetic Respiratory Sensing Technology (MRST) recordings of normal breathing, breath-holding, and deep breathing were obtained from 306 participants (122 healthy, 32 with COVID-19, 152 with other RDs (e.g., influenza/pneumonia and tuberculosis)). A total of 225 time/frequency/morphology-based features were extracted from the breath signals. A logistic regression (LR) model was trained on our dataset to detect RDs using five-fold cross-validation. A binary LR model was also trained to classify healthy versus non-healthy. In addition, age-stratified (≤45, 45–65, >65 years) comorbidity hypergraph networks were constructed to capture higher-order co-occurrence among RDs and CVDs. Results: The multiclass LR model achieved a mean accuracy of 87.8 ± 2.4% across five folds, demonstrating effective discrimination of five RDs from breath signals alone; the binary LR model achieved a superior mean accuracy of 97.0 ± 0.6%. Comorbidity network analysis showed prominent age-related patterns. In the youngest group (≤45 years), networks were sparse and centered on influenza/pneumonia with weak links to chronic lower RDs, tuberculosis, and CVDs. In middle age (45–65 years), network density increased, with hypertensive diseases, influenza/pneumonia, and chronic lower RDs forming strong connections that marked emerging CVD–RD overlap. In older adults (>65 years), networks were highly interconnected; hypertensive and ischemic heart diseases functioned as hubs linking multiple RDs, consistent with escalating comorbidity severity. Conclusions: A combined framework integrating MRST-derived features, ML classification, and comorbidity network modeling can diagnose RDs categories noninvasively and capture well the progression of CVD-RD comorbidity. This approach can support early screening, comorbidity monitoring, and treatment personalization.