Deep learning framework for cardiorespiratory disease detection using smartphone IMU sensors

Respiratory and cardiovascular diseases represent a significant global health burden, underscoring the need for innovative, accessible, and cost-effective screening solutions. This study introduces a clinically grounded framework for the early detection of cardiorespiratory conditions using commodity smartphones equipped with inertial measurement unit sensors. The proposed method leverages accelerometer and gyroscope data collected under a standardized protocol from five distinct thoracoabdominal regions, enabling the acquisition of respiratory kinematics through non-invasive, low-cost technology suitable for remote health monitoring-particularly in resource-limited settings or during pandemic outbreaks. A dedicated preprocessing pipeline segments the time series into individual breathing cycles, which are then analyzed using a bidirectional recurrent neural network to perform binary classification between healthy individuals and patients with cardiovascular disease. The non-healthy cohort comprised preoperative patients diagnosed with conditions including valvular insufficiency, coronary artery disease, and aortic aneurysm. The model was trained and validated using leave-one-out cross-validation with Bayesian hyperparameter optimization. Experimental results demonstrated robust classification performance, with an average sensitivity of 0.81±0.02, specificity of 0.82±0.05, F1 score of 0.81±0.02, and accuracy of 80.2%±3.9. On an independent set of unseen healthy individuals, the model achieved a true negative rate of 74.8%±4.5, confirming its generalization capability. The proposed framework offers a promising avenue for improving public health, enabling remote monitoring, and supporting clinicians in early diagnosis. Future work should focus on expanding the dataset, refining the methodology for long-term monitoring, and assessing its applicability across diverse clinical and at-home settings.