A Low-Power ECG Readout Circuit Integrated with Machine Learning Based ECG Heartbeat Classifier

Dry electrodes have been a popular mechanism for ECG acquisition due to their ease of use and ability to secure longer signals. However, for wearable ECG monitoring to be effective, high-fidelity signal processing is imperative. Lack of low-cost, compact, and effective processing options at the subject end could result in noisy ECG data transmission. To overcome this constraint this paper presents a four-layered Printed Circuit Board (PCB) based readout circuit using commercial off-the-shelf components (COTS), having a low power consumption of 1 mu W and an area of 2.1 x 1.8 cm(2). The readout circuit components are an amplifier with a gain of 40 dB, a 60 Hz notch filter, and a low pass filter with a cut-off frequency of 100 Hz. For a test signal, the SNR improved from -9 dB to 29.3 dB using the readout circuit. Voluminous ECG data acquired from patients over a long period requires deep analysis, feature extraction, and annotation for pattern recognition, which is a significantly time-consuming task. The development of machine learning programs can analyze hard-to-read ECG data and identify unique characteristics within the data in a short period. In this paper three machine learning models, namely Decision Trees (DT), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM) are built to classify the MIT-BIH Arrhythmia Database. The results show that the Weighted KNN model has the highest training accuracy of 88.7% and the Medium Decision tree is the fastest with 23.9 seconds.