A Sensor‐to‐Algorithm Framework for Respiration Monitoring: Stretchable Strain Sensors Coupled With Machine Learning Feature Optimization

ABSTRACT

Reliable classification of respiration states is essential for wearable technologies and clinical monitoring, yet robust methods for distinguishing subtle waveform differences across breathing conditions remain limited. This study aims to develop a machine learning (ML)–optimized framework for accurately classifying shallow, normal, and deep respiration patterns from a stretchable strain sensor. Respiration waveforms were transformed into feature representations derived from both handcrafted ML features and convolutional neural network (CNN)‐based latent embeddings. Multiple classifier families were evaluated using ten‐fold stratified cross‐validation. Random forest achieved the strongest overall performance, yielding 95.0% accuracy for shallow versus normal breathing (59/60 and 55/60 correctly classified), 99.2% accuracy for shallow versus deep (60/60 and 59/60 correct), and 91.7% accuracy for normal versus deep (56/60 and 54/60 correct). Shapley Additive exPlanations (SHAP) analysis and hierarchical clustering confirmed that high‐ranking features consistently reflected amplitude variability and signal‐energy differences across breathing states. These results demonstrate that combining strain sensor waveforms with ML models enables preliminary and interpretable classification of respiratory patterns. Unlike prior respiration‐classification approaches, our framework explicitly links breathing‐induced strain‐sensor deformation to interpretable waveform features through SHAP‐guided model optimization, enabling a physically grounded and explainable pathway from material response to classifier decision.