Multi-Cycle Temporal Attention Modeling for Automated Lung Sound Analysis

Lung sound analysis is a vital role that is involved in the early detection of respiratory pathology since it captures the airflow induced acoustic patterns, which occur during respiratory cycles. The signals carry clinically relevant information about airway obstruction, inflammation, and aberrant pulmonary mechanics but their interpretation is inherently complex due to low signal-to-noise ratio, pathological signatures overlap, inter-subject variability and non-stationary time-dependent variations across breaths. Traditional methods of auscultation, as well as existing automated methods, are usually based on isolated cycle analysis or short-term time analysis and thus limit their ability to capture progressive acoustical development and decay strength in realistic clinical conditions. In response to the drawbacks, the current study suggests a sequence-sensitive learning paradigm that combines both localized acoustic feature and large temporal dependencies between a number of respiratory cycles. The proposed Cycle-Aware Spatio-Temporal Attention Network (CSTANet) uses 2-channel acoustic encoding on time domain and time frequency representation that are then subjected to transformer-based temporal reasoning. Such a structure allows dynamic prioritization of cycles and enables sequence level decision making. The suggested approach was strictly tested in terms of numerous experiments using the ICBHI 2017 Respiratory Sound Database, which included four clinically relevant sound classes and diverse acoustic settings. Quantitative assessment demonstrated that CSTANet can be used to achieve a classification accuracy of 95.62%, precision of 94.28%, recall of 93.74%, specificity of 96.05%, F1-score of 94.01% and Matthews correlation coefficient of 0.934, which is always higher than the performance of CNN, CNN-LSTM, CNN-GRU, ACNN, and LCNN baselines. Based on these results, it is established that the multi-cycle temporal attention significantly increases the discriminative reliability, interpretability, and generalization of automated lung sound analysis.