Filter Validation for Detecting Outliers of Photoplethysmograph Data

Outliers are abnormal data in a distribution of data. Outlier data handling is critical in machine learning systems because outlier data can affect the results of predictions made by machine learning. So far, the easiest way to handle outliers is by trimming or removing the data. This study applies outlier data handling to photoplethysmograph (PPG) signals. PPG signal is one method used to determine the condition of the cardiovascular system by measuring changes in blood volume in skin tissue. The wrong sampling procedure, improper tool placement on fingers, and problems with light from the surrounding environment can cause outlier problems in the PPG signal. This paper proposes an outliers detection method for PPG signals to reduce prediction bias and increase accuracy. The method used is filter one with Moving Average Filter (MAF), filter two with Slope detection, and filter three with Z-score. Applied to three machine learning which are Linear regression, SVR, and Random forest. The filter method significantly improves the accuracy of all machine learning, but Random forest has the highest accuracy of the others. The results of the accuracy before the implementation of noise detection was 75.43%. A combined filter one and two accuracy increases to 89.65%, and filter three best accuracies becomes 93.99%.