All research

Physiological signal processing · 2024–2025

Cascaded 1D-CycleGANs for improved cardiovascular monitoring from wearable wrist PPG

A two-stage 1D-CycleGAN pipeline that restores noisy wrist PPG, translates it to finger PPG and improves agreement with ECG-derived heart rate and variability.

Qatar National Research Fund NPRP12S-0227-190164 · Prince Sattam Bin Abdulaziz University project PSAU/2023/R/1444

Two-part framework diagram. In the first part, separate CycleGANs learn from good and bad wrist PPG and its first and second derivatives, then restore wrist signals over repeated passes. In the second part, a three-channel CycleGAN translates the restored wrist signals to finger PPG for heart rate and variability analysis.
Figure 1 from Mahmud et al. (2024). TTR-GAN first restores wrist PPG and its derivatives (a), then translates the restored signals to finger PPG (b).View full-size figureSource articleCC BY 4.0

Approach

Can noisy wrist PPG from a wearable be restored and then translated into finger PPG that follows the heart more closely?

Wrist-worn sensors make long-term pulse monitoring practical, but movement, changing skin contact and noise distort their signals, and wrist PPG carries less detail than clinical finger PPG. Translation Through Restoration GAN (TTR-GAN) handles these as two separate problems. The first 1D-CycleGAN stage removes artefacts from the wrist signal. The second converts the restored signal into finger PPG. Heart rate and variability from ECG then provide an independent check.

  1. Restore wrist PPG
  2. Translate to finger PPG
  3. Validate against ECG

Stage 1: blind restoration of wrist PPG

Unpaired 1D-CycleGANs

The study used MAUS, with about 12 hours 50 minutes of recordings from 22 healthy young adults. A wristwatch recorded wrist PPG while finger PPG and ECG were captured at the same time. Signals were resampled to 125 Hz and split into 512-sample segments. Wrist segments were sorted by hand into 1,876 good, 2,732 mediocre and 4,204 bad, and 164 unusable segments were removed.

A corrupted wrist signal has no clean recording of the same moment, so the CycleGANs learned from unpaired good and bad examples, without being told the type of artefact. Separate restorers handled wrist PPG and its first and second derivatives (VPG and APG). Each generator was a nine-block residual network built from Super-ONN layers and tested with five-fold cross-validation on separate people.

With no clean target to compare against, restoration was measured by the fall in signal entropy, a measure of irregularity. Outputs were passed through the generator up to five times. In most cases performance peaked around the third pass, so third-pass signals moved on to translation. Good-quality signals stayed almost unchanged.

Wrist PPG approximate entropy after three passes
0.40 → 0.25
Wrist PPG sample entropy after three passes
0.39 → 0.22
Restoration passes selected for translation
3
Mean entropy before restoration and after the third pass, read from Figure 5 of Mahmud et al. (2024). Lower values indicate a more regular signal.
Approximate entropySample entropyFuzzy entropy
Wrist PPG0.40 → 0.250.39 → 0.220.13 → 0.12
First derivative (VPG)0.39 → 0.220.36 → 0.170.18 → 0.16
Second derivative (APG)0.48 → 0.370.50 → 0.400.27 → 0.25
Two examples, each with rows for PPG, VPG and APG. From left to right, a noisy unrestored wrist signal becomes smoother over five restoration passes, and the reported approximate entropy falls.
Figure 7 from Mahmud et al. (2024), cropped to examples a and b. Wrist PPG and its derivatives over five restoration passes.View full-size figureSource articleCC BY 4.0
Three line charts for wrist PPG, VPG and APG. Approximate, sample and fuzzy entropy are plotted from the unrestored signal through five passes. Approximate and sample entropy fall most at the first pass and then level off.
Figure 5 from Mahmud et al. (2024). Mean entropy of (a) wrist PPG, (b) VPG and (c) APG before restoration and after each pass.View full-size figureSource articleCC BY 4.0

Stage 2: translating restored wrist PPG to finger PPG

Three-channel 1D-CycleGAN

Even a clean wrist signal differs from finger PPG in shape and timing, and it carries less of the detail used in clinical analysis. The second CycleGAN therefore learned from 8,812 time-matched pairs of restored wrist and reference finger segments. Its Super-ONN UNet generator received restored wrist PPG, VPG and APG as three input channels.

Ordinary correlation penalised the natural time shift between wrist and finger, so the study introduced PRTX, which combines temporal cross-correlation with signal power ratio. Translation raised PRTX from 0.5867 to 0.9455 against the reference finger PPG. Restoring the wrist signal first gave a better result than translating it directly, and using all three channels added a further gain.

PRTX of translated finger PPG
0.9455
PRTX gain over restored wrist PPG
+35.88 points
Five-fold cross-validated PRTX against reference finger PPG, from Table 1 of Mahmud et al. (2024). Higher is better.
Single channelThree channels
Restored wrist PPG, before translation0.58670.5867
Translated without restoration0.89530.9212
Translated after restoration0.93570.9455
Three example rows. Each shows a noisy wrist PPG segment, the restored wrist PPG, and the translated finger PPG, next to the reference finger PPG and ECG recorded at the same time.
Figure 8 from Mahmud et al. (2024), cropped to examples a–c. Unrestored wrist PPG, restored wrist PPG and translated finger PPG, beside the reference finger PPG and ECG.View full-size figureSource articleCC BY 4.0

Checking pulse rate and variability against ECG

Evaluation of the full pipeline

ECG was not used for training. It served only as an independent reference for heart rate and heart rate variability (SDNN). Each stage improved agreement with ECG: restoration gave a first gain and translation a larger one. Measured finger PPG remained the closest match, so the translated signal narrows the gap without replacing a finger sensor.

Pulse rate correlation with ECG heart rate
0.797 → 0.900
Pulse rate variability correlation with ECG
0.699 → 0.846
Pearson correlation with ECG-derived values for the three-channel pipeline, from Table 2 of Mahmud et al. (2024).
RateVariability (SDNN)
Raw wrist PPG0.79680.6991
Restored wrist PPG0.83040.7438
Translated finger PPG0.90030.8457
Measured finger PPG0.92420.9254
Two scatter plots of pulse rate from reference finger PPG. The left plot compares it with wrist PPG and includes outliers above 150 beats per minute, with a correlation of 0.87. The right plot compares it with translated finger PPG, with a correlation of 0.91.
Figure 6e–f from Mahmud et al. (2024), with the two panels placed closer together. Pulse rate from reference finger PPG against (e) wrist PPG and (f) translated finger PPG.View full-size figureSource articleCC BY 4.0

Comparing CNN, Self-ONN and Super-ONN generators

Performance and computational cost

The same restoration and translation pipeline was built with three neuron types. Super-ONN gave the largest gains but also needed the most parameters and the longest inference time, so a wearable design would have to weigh accuracy against cost.

Three-channel results from Table 4 of Mahmud et al. (2024). Gains are in percentage points. Parameters and inference time are given for one generator.
PRTX gainRate gainVariability gainRestorerTranslator
CNN32.89.612.60.240 M · 12.9 ms0.853 M · 6.61 ms
Self-ONN33.79.812.90.715 M · 18.2 ms2.176 M · 15.7 ms
Super-ONN35.910.414.71.048 M · 26.7 ms2.184 M · 23.0 ms
Architecture diagrams of the restoration generator with nine residual blocks, the UNet translation generator built from Super-ONN layers, and the Self-ONN PatchGAN discriminator.
Figure 4 from Mahmud et al. (2024). (a) Restoration generator, (b) translation generator and (c) discriminator.View full-size figureSource articleCC BY 4.0

Follow-on study: atrial fibrillation detection

Biomedical Signal Processing and Control · 2025

The atrial fibrillation study applied 1D-CycleGAN restoration to wrist PPG before classifying atrial fibrillation. Restoration raised classification accuracy by 2.94 and 5.08 percentage points in two test splits, and the Self-AFNet classifier reached 96.41% and 97.09%. This study used restoration without the finger PPG translation stage.

My contribution

I led the development and evaluation of TTR-GAN, including the Super-ONN-based 1D-CycleGAN methodology, the experiments, the quantitative and qualitative analysis and the manuscript. For the atrial fibrillation study, I developed the 1D-CycleGAN implementation used for wrist PPG restoration and contributed to conceptualisation, formal analysis and writing.

Study scope

TTR-GAN was developed on one dataset of 22 healthy young adults and was not tested on an independent cohort. Restoration was judged by entropy because no clean reference exists for a corrupted wrist segment. Rare arrhythmic segments from one participant were harder to reproduce, and on-device performance was not reported. Patient data and testing on wearable hardware are the next requirements.