Improving Signal Quality in Non-Contact Electrocardiography: Novel Strategy for Motion Artifact Reduction

Figure A1. Configuration screen of the gamification interface showing configurable experimental parameters. The interface allows selection of the IMU serial port and definition of motion protocol stages, each specified by oscillation frequency (Hz), duration (s), and target amplitude (in pixels, px).

Figure A2. “Get ready” screen of the gamification interface. The progress bar fills as both conditions are continuously satisfied, ensuring consistent initial posture across all stages and subjects. This enforced rest period also provides a clean baseline segment in the recorded data, facilitating reliable segmentation between motion and rest intervals.

Figure A3. Screenshot of the gamification interface for controlled motion generation. The red circle indicates the target position, moving sinusoidally along the vertical axis, while the blue circle represents the subject’s cursor controlled by steering wheel rotation measured via integrated IMU. Real-time feedback displayed includes current stage parameters (frequency f = 0.20 Hz, amplitude A = 260 px, gain = 12 px/deg, pixels per degree), tracking performance metrics (MAE = 164.9 px, RMSE = 183.0 px), IMU data (angle, angular velocity ω, sampling interval dt, gyroscope bias estimate), and connection status (69 Hz update rate). Subjects are instructed to minimize the distance between target and cursor by rotating the steering wheel, producing standardized, repeatable motion patterns at prescribed frequencies across experimental stages.

Figure 1. Block diagram of the experimental measurement system comprising two galvanically isolated, battery-powered acquisition units. Non-contact electrodes are positioned in the coplanar configuration, as discussed in the following subsection.

Figure 2. 3D model of non-contact electrodes used in the research, front (left) and back (right) view. More details are available in our prior work [12,14].

Figure 3. Illustration of the evaluated electrode configurations. (Upper left)—Top View of coplanar configuration; (Upper right)—Front view of coplanar configuration; (Lower left)—Top view of stacked configuration; (Lower right)—Front view of stacked configuration.

Figure 4. Complete set of reference signals derived from secondary pair of noise electrodes for multi-reference adaptive filtering (Subject 2, coplanar configuration, Environment A). (a) Raw signal dominated by 50 Hz mains coupling with visible motion-induced amplitude modulation. (b,c) Band-pass-derived references: The 0.5–10 Hz band-passed signal (BP) captures low-frequency baseline shifts, while its derivative (dBP) emphasizes rapid baseline transients. (d,e) Envelope-derived references: The mean envelope magnitude (Env) tracks amplitude modulation of the mains interference, and its derivative (dEnv) captures modulation dynamics. (f,g) Asymmetry-derived references: The envelope asymmetry signal (Asymm) reflects non-linear coupling differences between positive and negative interference half-cycles, with its derivative (dAsymm) highlighting non-linear transient components. All reference signals are normalized and zero-meaned for input to the NLMS adaptive filter.

Figure 5. Derivation of envelope-based reference signals from secondary pair of electrodes (Subject 2, coplanar configuration, Environment A). (a) Raw signal from a non-contact electrode showing dominant 50 Hz mains coupling, with upper (red) and lower (blue) envelopes extracted via peak detection. (b) Magnitudes of positive and negative half-cycle envelopes, revealing distinct amplitude modulation patterns induced by motion affecting electrode–skin coupling. (c) Envelope asymmetry signal computed as the difference between upper and lower envelope magnitudes, capturing motion-induced variations in the capacitive coupling that manifest differently for positive and negative interference half-cycles.

Figure 6. Power spectral density (PSD) of the six reference signals derived from the secondary pair of electrodes (Subject 2, coplanar configuration, Environment A), peak-normalized to facilitate spectral shape comparison (dB re peak: dB relative to peak). The shaded region indicates the QRS complex band (5–15 Hz); the 12 Hz correction signal low-pass filter cutoff (vertical red dashed line) is chosen to enable artifact cancelation while preserving R-peak morphology.

Figure 7. Signal processing pipeline for deriving six reference signals from the floating noise electrode pair. All references are direct-current-removed (DC-removed) and σ-normalized before concatenation into a 6 × M tap input vector for the NLMS adaptive filter, where M denotes filter order.

Figure 8. Block diagram of the multi-reference NLMS adaptive filtering system. The primary signal d[n], containing both ECG and artifacts, is preprocessed while the concatenated reference vector x[n] comprising six derived references (BP, dBP, Env, dEnv, Asymm, dAsymm) with M = 384 taps each feeds the adaptive filter (L = 6M = 2304 total taps). The filter output ŷ(n) is scaled by α = 0.9 to prevent overcorrection, low-pass filtered at 12 Hz for QRS preservation, and gated by the motion detection mask before subtraction from the primary signal. Motion detection operates on d[n] using dual-band monitoring (LF < 2 Hz, HF > 25 Hz) with hysteresis thresholds and minimum duration enforcement. Filter weights are reset at motion onset and updated via leaky NLMS (γ = 0.001) using the raw error signal computed before scaling, filtering, and gating, ensuring adaptation based on true residual error during motion segments only.

Figure 9. Example of data interpretation from Stage 1 of the gamification protocol for standardized motion artifact generation using IMU-based real-time feedback. (Top row): Raw accelerometer and gyroscope data from the apparatus-mounted IMU during controlled axial rotation, showing dominant motion in the z-axis (rotation about the vertical axis). (Middle row): Integrated rotation angle derived from gyroscope data, demonstrating consistent ±20° oscillation amplitude, alongside comparison of processed angular velocity with raw gyroscope output. (Bottom row): Target tracking performance from the PyGame (version 2.6.1; Pygame Project) visual interface, where subjects follow a sinusoidally moving target (blue) by rotating the apparatus; tracking error (right) shows a mean deviation of 42.7 pixels, indicating adequate adherence to the prescribed motion pattern.