Convenient prediction of TcPO2 level using machine learning techniques with the PPG signal’s characteristic features

Abstract

BackgroundTranscutaneous oxygen tension (TcPO2) is a noninvasive method for assessing the oxygen supply of the skin in clinics. However, it is difficult to deploy broadly in primary care settings due to its high cost and time-consuming. This study proposed a fast and inexpensive technique for predicting TcPO2 using the characteristic features of the photoplethysmography (PPG) signals and various machine learning algorithms. MethodsThe toe PPG signals of 149 subjects were collected, and 12 characteristic features were extracted according to contour analysis, followed by TcPO2 measurements. TcPO2 measurements were performed on the dorsum of the foot by PeriFlux 6000 TcPO2 system. In addition to PPG features, physiological parameters were also taken into account. A variety of machine learning regression approaches, including extreme gradient boost (XGBoost), random forest regression (RFR), support vector machine regression (SVR) and Ridge regression (Ridge), were used to predict TcPO2 levels. To evaluate the performance of machine learning algorithms, different performance indicators such as mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) and correlation coefficient were employed. Correlation analysis and Bland-Altman plot analysis were performed to visualize the data. ResultsXGBoost outperformed the other three regression models in predicting TcPO2 (R = 0.835, MAE = 5.23 mmHg, RMSE = 6.87 mmHg, MAPE = 9.99%). Reflection index, diabetes, duration of diabetes, heart rate, crest time, the time between the systolic and diastolic peaks (ΔT) were the top six contributions to TcPO2 prediction in the XGBoost model. The PPG characteristic features play a dominant role in TcPO2 estimation. Diabetes and duration of diabetes were also significant factors in TcPO2 prediction. ConclusionsWe compared four machine learning methods for TcPO2 prediction. The XGBoost model had the best predictive performance of all the algorithms investigated and could be used to forecast TcPO2 values. The TcPO2 prediction model based on PPG signal reduces the cost and shortens the measurement time of TcPO2, providing significant benefits for primary care.