MACHINE LEARNING-BASED CLASSIFICATION AND REDUCTION OF MOTION ARTIFACT NOISE IN ECG SIGNALS FOR WEARABLE VITAL SIGN MONITORING DEVICES
Abstract
Wearable vital signs monitoring devices, while widely adopted, face challenges in clinical diagnosis due to artifact noises. This thesis introduces a novel application of machine learning—specifically, the K-Nearest Neighbor (KNN) model—to recognize artifacts in ECG signals using 3-axis accelerometer data. The proposed fine KNN model, leveraging wavelet scattering coefficients, effectively classifies accelerometer samples likely to contain artifact noises in ECG signals, achieving a remarkable positive predictive value of 94.7%. To enhance accuracy and reliability, a robust technique is proposed, recognizing artifacts using wavelet scattering coefficients from both accelerometer and ECG signals. The KNN model achieves a test accuracy of 98.8%, making it suitable for integration into wearable ECG monitoring devices. Following artifact noise identification, a two-stage technique reduces noise levels and reconstructs the signal. Wavelet denoising preserves crucial signal information, and a denoising autoencoder enhances Signal-to-Noise Ratio (SNR) and reduces Root Mean Square Error (RMSE). Importantly, the proposed models can be implemented in wearable ECG monitoring devices without additional sensors. This opens possibilities for applications in various vital signs monitoring scenarios, such as photoplethysmography (PPG), electroencephalography (EEG), or other monitoring devices. The research significantly contributes to advancing noise reduction techniques in wearable health monitoring, ensuring accurate vital sign measurements for improved clinical applications.