LUNAR: Periodicity-aware time-series analysis framework for LUNg Auscultation Respiratory detection

Accurate detection of respiratory cycles in lung sound auscultation is crucial for diagnosing respiratory diseases such as asthma, chronic obstructive pulmonary disease, and pneumonia. Traditional auscultation methods rely heavily on clinician expertise, creating diagnostic inconsistencies particularly in resource-constrained settings where access to specialists varies. While recent artificial intelligence approaches have advanced abnormal breath sound classification and respiratory disease diagnosis, most systems depend on expert manual annotation to identify respiratory cycle boundaries. This dependency, coupled with moderate performance of existing detection models, creates a significant bottleneck in automated respiratory sound analysis. Here we present LUNg Auscultation Respiratory detector (LUNAR), a novel periodicity-aware deep learning framework that directly processes raw lung auscultation signals to automatically detect respiratory cycles. LUNAR integrates a novel Respiratory Periodicity Awareness Module (RPAM) with a convolutional neural network to explicitly model the repetitive nature of respiratory cycles. The RPAM consistently enhanced performance across six architectural variants, with improvements of 12.0% in AP and 6.3% in F1-score. Trained on HF_Lung_V1 (n=279) and validated across ICBHI (n=126) and SNUCH_Lung (n=203) datasets, the model achieved exceptional performance (AP: 0.880, F1: 0.894), maintaining robust results across both adult and pediatric populations with direct processing of raw time-series signals. These findings establish LUNAR as a robust tool for clinical lung sound analysis, with potential for broader application through multi-institutional data acquisition and extension to abnormal breath sound detection. Our approach reduces dependency on specialist expertise for respiratory assessments, potentially improving healthcare accessibility in resource-limited settings.