Generalization of Data Reliability Metric (DReM) Mechanism for Pulsatile Bio-signals

Abstract

Due to rapid development of wearable technologies used for health monitoring, a robust data reliability assessment technique is required. Choosing the right sets of Data reliability metrics (DReM) can improve the performance of data reliability assessment and thus, it can help in reliable data acquisitions of bio-signals. Traditional reliability assessment techniques rely significantly on the signal specific peak detection algorithms. It impedes the endeavor of generalizing signal independent data reliability assessment techniques. In this work, we explored nine signal independent statistical candidates for DReM and finalized five top features as our DReMs which can identify acceptable signal segments from unacceptable segments with 0.83 precision, 0.84 recall and 0.83 F-1 score on accurately identifying acceptable pulses (ECG, PPG and respiratory signal). Additionally, the five DReMs are capable of detecting unacceptable signal segments with 0.92 precision, 0.92 recall and 0.92 F-1 score. We proposed optimal Random Forest classifier model with excellent Receiver Operating Characteristics (ROC) with significant Area Under the Curve (AUC) value of 0.994.