PPG signal quality is often degraded by motion artifacts. This study presents an adaptive cluster-based filtering technique that leverages accelerometer data to suppress motion-induced artifacts while preserving the morphological integrity of the PPG signal.
Fifty patients wearing wristbands were recorded over 24 hours, each providing three PPG channels and tri-axial accelerometer data. Seven statistical features were extracted from the norm of the acceleration vector, segmented into 10-second windows. Standard deviation and skewness emerged as the most informative pair for clustering, selected based on a scoring system combining Silhouette Mean, Davies-Bouldin Index, Dunn Index, and a correlation penalty. K-means clustering with k=3 identified three levels of motion-related artifacts, which were mapped to synchronized PPG segments.
Each segment was filtered using a five-level Discrete Wavelet Transform (DWT) with a Symlet 4 basis. The filtered signals were then reconstructed using inverse DWT and smoothed with a 30% overlapping window blending to reduce boundary artifacts. The performance of the proposed method was compared to two alternatives: a 2nd-order Butterworth Bandpass Filter (0.5–3 Hz) and a five-level wavelet filter using the Daubechies 4 mother wavelet. Evaluation was based on the Signal-to-Noise Ratio (SNR), measuring the reduction of both noise and signal components.
Clustering was performed across 10 independent sessions, achieving a Silhouette Index of 0.8987 ± 0.0124, a Davies-Bouldin Index of 0.4442 ± 0.0317, and a Dunn Index of 0.00379 ± 0.00022. The proposed method reduced noise energy by an average of 30.90 dB while preserving signal morphology. In comparison, the Butterworth filter reduced noise by 34.88 dB but introduced waveform distortion, and Daubechies 4 filtering achieved 22.58 dB of noise reduction with visible artifacts.
This study proposes a cluster-based method that filters motion artifacts using accelerometer data, adapting to motion intensity and preserving signal quality, suitable for real-time wearable health monitoring applications.