Machine Learning for the Diagnosis of Chronic Obstructive Pulmonary Disease and Photoplethysmography Signal – Based Minimum Diagnosis Time Detection

Abstract

The main pathological characteristic of Chronic obstructive pulmonary disease (COPD) is chronic respiratory obstruction. COPD is a permanent, progressive disease and is caused by harmful particles and gases entering the lungs. The difficulty of using the spirometer apparatus and the difficulties in having access to the hospital, especially for young children, disabled or patients in advanced stages, requires to make the diagnosis process easier and shorter. In order to avoid these problems, and to make the diagnosis of COPD faster and then easier to track the disease, it is considered that the use of the Photoplethysmography (PPG) signal would be beneficial. PPG is a biological signal that can be measured anywhere on the body from the skin surface. The PPG signal, that is created with each heartbeat, is an easy measurable signal. The literature contains lots of information on the PPG signal of the body. In this study, a system design was made to use PPG signal in COPD diagnosis. The aim of the study is to determine: “Can COPD be diagnosed with PPG?” and “If it is possible, with the help of minimum how many seconds of signals this process can be completed?” In the line with this purpose, in average 7–8 h of PPG records was obtained from 14 individuals (8 COPD, 6 Healthy ones) for the study. The obtained records are divided into sequences of 2, 4, 8, 16, 32, 64, 128, 256, 512 and 1024 s. Studies were made for each group of seconds and it was tried to determine which second signals could diagnose with higher performance. The 8-h records of the sick individuals are divided into sequences of 2-s, and each sequence was given a patient tag. When the same procedure was done for the healthy individual, all parts were labeled as Healthy. Each signal group was firstly cleaned by 0.1–20 Hz numerical filtering method. Later on, 25 items feature extractions were made in time domains. Finally the formed data set (2, 4, 8 s) were classified by decision tree machine learning methods. According to the obtained results, the highest performance values were achieved with a 2-s data group, and 0.99 sensitivity, 0.99 specificity and 98.99% accuracy rate was obtained. Consistent with these results, it is assumed that COPD diagnosis could be made based on machine learning with PPG signal and biomedical signal processing techniques.