PPG-Derived Respiration Signal Predictor-Guided Simple Rule-Based Detection of COPD

Chronic obstructive pulmonary disease (COPD) is one of the common respiratory diseases that can cause death if not intervened in time. Popular lung sound sensing-based COPD detection techniques mostly employ deep learning techniques, which suffer from the disadvantages of purely data-driven models, like low robustness against data variability across diversified datasets. This article introduces a finger photoplethysmography (PPG)-based COPD detection technique using a low-complexity, simple rule-based classifier. The primary feature extraction was done using dynamic mode decomposition (DMD), which requires low memory. A novel respiration signal predictor (RSP) based on bidirectional reservoir computing (B-RC) model is proposed to identify respiration indicative DMD mode. Evaluation was done with a total of 4970 PPG frames, each of size 2750 samples collected from three public datasets and volunteers, resulting in 95.13% accuracy and 100% sensitivity of the RSP. Detection accuracy and sensitivity of acute COPD, moderate COPD, healthy subjects, and artificially obstructed breathing (AOB) class (a synthetic one derived from healthy subjects) were 99.7% and 99.5%, respectively. The average latency and peak memory requirement on single-core ARM v6 controller (512-MB RAM and 1-GHz frequency) were 616 ms and 458.44 kB, respectively, to process one PPG frame. The present method can be utilized for quick screening of COPD as a point-of-care (POC) technology.