An Approximate Binary Classifier for Data Integrity Assessment in IoT Sensors

This paper presents a machine learning-based technique for quality assessment of electrocardiogram (ECG) signals in wearable Internet of Things (IoT) sensors. Quality assessment at the network edge aids in the elimination of corrupted data prior to storage or transmission. In this work, we used a k-nearest neighbor (k-NN) binary classifier for identifying whether the acquired sensor data is of acceptable quality for further processing and transmission. Feature vectors used for classification were derived from the raw signal using skewness and kurtosis based signal quality indicators (SQIs), and these SQIs do not require any prior processing or knowledge of fiducial points in the ECG signal. The proposed approach achieved a classification accuracy of 97.18% with an estimated complexity that corresponds to 12.72 fJs Energy-Delay-product (EDP) in terms of multiplications used. To further reduce computational complexity and power consumption, an approximate multiplier was used, and this method exhibited an accuracy of 96.48%. The EDP, while using an approximate multiplier for classifying a single record was found to be 34.5% lower at 8.333 fJs, and is within the power budget of a typical IoT device.