Real-Time ECG Artifact Removal for Adaptive Deep Brain Stimulation: A Comparative Study

Figures

• 

  Figure A1. Time-domain signal examples showing clean reference, contaminated, and cleaned signals across contamination levels (0, +10, +20 dB) for all four ECG removal methods. Each row represents a different contamination level. First column: Clean reference (black) overlaid with contaminated signal (red) showing progressive ECG dominance. Subsequent columns: Cleaned signals overlaid with clean reference for SVD, TS, PR, and eSVD methods, demonstrating varying degrees of artifact suppression and signal preservation.

• 

  Figure A2. Power spectral density comparison at 0 dB contamination. Clean reference (black) vs. contaminated signal (red) shows 8–12 dB elevation. TS achieves superior cardiac harmonic suppression (<1 dB) with accurate beta peak recovery. SVD recovers beta peak at 22 Hz with residual low-frequency artifacts. eSVD exhibits over-aggressive removal with 60% reduction in beta peak prominence. PR shows incomplete artifact removal. Vertical dashed lines indicate beta band boundaries (13–35 Hz).

• 

  Figure A3. Power spectral density comparison at +10 dB contamination. Severe contamination (15–20 dB elevation) completely obscures neural features. TS maintains robust performance with consistent beta peak recovery (RMS deviation 0.9 dB). SVD shows progressive degradation compared to 0 dB. PR exhibits spurious spectral peaks at 25 and 32 Hz indicating interpolation artifacts. eSVD performance improves at higher contamination.

• 

  Figure A4. Power spectral density comparison at +20 dB severe contamination. Extreme contamination (20–30 dB elevation) obliterates neural signal. eSVD achieves best quality: beta peak 4.3 dB (4% error), accurate frequency (22.1 vs. 22.0 Hz), near-perfect power law (α = −1.08 vs. −1.12 clean), but exceeds computational budget (P99 = 186 ms). TS maintains clinically acceptable recovery within real-time constraints. SVD shows substantial degradation; PR catastrophic failure (82% beta power loss).

• 

  Figure A5. Beta burst detection examples at 0 dB contamination. (Top) Clean reference showing characteristic burst pattern with 75th percentile threshold (dashed line) and detected bursts (green shading). (Second) Contaminated signal with prominent QRS artifacts obscuring burst envelope. (Third-Sixth) Cleaned signals for SVD, TS, PR, and eSVD methods. All methods successfully recover burst timing and amplitude characteristics. TS (BCP = 0.94) and SVD demonstrate comparable performance with modest under-detection, while PR (BCP = 0.98) and eSVD (BCP = 0.97) show near-perfect burst preservation closely matching clean reference.

• 

  Figure A6. Beta burst detection examples at +10 dB contamination. (Second) Severe QRS contamination completely dominates signal, with peak amplitudes exceeding ±100 μV obscuring all neural activity. (Third) SVD-cleaned signal shows reduced burst amplitude and fewer detected events compared to clean reference. (Fourth) TS-cleaned signal exhibits over-detection (BCP = 1.26) with spurious burst identifications (e.g., t = 1 s, t = 5 s) likely driven by residual artifact-induced envelope fluctuations, though this difference does not reach statistical significance. (Fifth) PR shows near-complete burst suppression with minimal detected events. (Sixth) eSVD maintains accurate burst detection with timing and amplitude closely matching clean reference.

• 

  Figure A7. Beta burst detection examples at +20 dB severe contamination. (Second) Extreme QRS contamination with peak amplitudes reaching ±200 μV, rendering neural activity completely invisible. (Third) SVD-cleaned signal shows substantial burst under-detection (BCP = 0.77) with missing events (e.g., t = 2 s, t = 7 s) and reduced amplitudes. (Fourth) TS exhibits marked over-detection (BCP = 1.69) with numerous spurious burst identifications, likely driven by residual QRS transients triggering false-positive detections. Despite this over-detection, TS preserves more burst events than SVD, which shows substantial under-detection. (Fifth) PR exhibits catastrophic failure with near-complete absence of detected bursts. (Sixth) eSVD achieves excellent burst recovery with detection pattern closely matching clean reference, confirming superior performance at severe contamination despite computational cost (P99 = 186 ms).

• 

  Figure A8. Real data P99 latency distributions for medication-OFF state across both recording channels. Box plots show median (orange line), interquartile range (box), and whiskers extending to 1.5×IQR. (Left) Channel 1: SVD median = 11.0 ms (IQR: 10.9–13.3 ms), TS median = 10.7 ms (IQR: 10.6–12.9 ms). (Right) Channel 2: SVD median = 10.4 ms (IQR: 9.9–11.2 ms), TS median = 10.5 ms (IQR: 10.4–10.9 ms). Both methods demonstrate consistent computational feasibility with all measurements well below the 50 ms threshold. Inter-channel variability is minimal, confirming robust performance across diverse signal characteristics.

• 

  Figure A9. Real data P99 latency distributions for medication-ON state. (Left) Channel 1: SVD median = 6.5 ms (IQR: 6.3–8.3 ms), TS median = 7.4 ms (IQR: 6.5–7.6 ms). (Right) Channel 2: SVD median = 6.9 ms (IQR: 6.8–7.1 ms), TS median = 7.3 ms (IQR: 7.0–7.9 ms). Medication-ON state shows 11–19% reduction in P99 latencies compared to OFF state (both p < 0.003), likely reflecting reduced ECG artifact severity under dopaminergic medication. Latency distributions remain tightly clustered with minimal outliers, demonstrating stable computational performance across patients and recording sessions.

• 

  Figure 1. Study design overview. (A) LFP signal acquisition and pre-processing pipeline. Signals were acquired via BrainSense Streaming at 250 Hz, bandpass filtered and normalized, and processed using four ECG artifact removal algorithms: Template Subtraction (TS), Singular Value Decomposition (SVD), Extended SVD (eSVD), and the Perceive toolbox (PR). (B) Performance assessment framework. Algorithm quality was evaluated on simulated data with known ground truth using four metrics (Artifact Removal Efficiency, Signal Preservation Ratio, Beta Power Preservation, Burst Count Preservation) and validated on real clinical recordings using reference-free metrics (ECG Suppression Ratio, Beta Peak Recovery, Template Correlation Metric, Power Law Score). (C) Real-time feasibility framework. Buffer-by-buffer processing was implemented to simulate device-level operation: each 100 ms buffer (25 samples) was processed within a 5 s sliding context window (1250 samples), with 50 ms allocated to ECG removal and 50 ms reserved for power estimation, threshold comparison, control decisions, and system overhead.

• 

  Figure 2. Comprehensive ECG removal performance across contamination levels for simulated data. (Upper left) Artifact Removal Efficiency (ARE): convergence at high contamination, with TS superior at 0 dB. (Upper right) Signal Preservation Ratio (SPR): TS, SVD, and PR maintain near-unity preservation; eSVD shows progressive attenuation. (Lower left) Beta Power Preservation (BPP): TS maintains stable performance; SVD degrades progressively; PR consistently underestimates. (Lower right) Burst Count Preservation (BCP): TS shows slight over-detection; SVD degrades at severe contamination. Dashed horizontal line indicates perfect preservation (value = 1.0).

• 

  Figure 3. Histograms show burst duration profiles at (a) 0 dB, (b) +10 dB, and (c) +20 dB contamination. Clean reference (gray bars) and contaminated signal (red bars) shown for comparison. At 0 dB, all methods preserve the physiological right-skewed distribution (modal duration 100–150 ms, tail extending to 600+ ms), with PR (BCP = 0.98) and eSVD (BCP = 0.97) most closely matching the reference. At +10 dB, TS exhibits inflation of short-duration bursts and depletion of mid-duration bursts (200–400 ms, BCP = 1.26), suggesting artifact-induced false detections. SVD and eSVD show moderate under-representation (BCP = 0.75 and 0.73), while PR maintains best adherence to the reference distribution. At +20 dB, TS shows marked over-detection across all bins (BCP = 1.69), while SVD, eSVD, and PR exhibit substantial to catastrophic under-detection (BCP = 0.77, 0.56, and 0.69 respectively), with complete absence of bursts >500 ms for SVD.