Fusion recovers one waveform, not a heart rate
The method, called Tri-Spectral PPG, does not output beats per minute. It takes synchronized green, red, and infrared reflective PPG signals and produces a single reconstructed PPG waveform that downstream processing can treat like an ordinary clean recording. Heart rate in this paper is one downstream use of that waveform, chosen to test whether the reconstruction is good enough to be useful.
That distinction matters for reading the headline number. The 4.5 bpm error is not the network's direct prediction; it comes from running a peak-detection pipeline on the fused signal and comparing the resulting interbeat intervals against a reference ECG. Any improvement therefore reflects both the recovered morphology and the detector applied to it.
Wavelengths suffer artifacts to different degrees
Reflective PPG infers blood volume pulses from how much light arterial blood absorbs as pulses propagate. Motion and ambient light changes corrupt that measurement, and the paper's premise is that these artifacts do not hit every wavelength equally. Green is the usual choice for heart rate and heart rate variability, while red and infrared are more commonly used for pulse oximetry — so a device already carrying all three channels holds information that single-channel processing discards.
Conventional fixes have limits that motivate fusion. Band-pass filtering struggles when artifact frequencies overlap the pulse spectrum. Averaging multiple traces reduces single-channel artifacts but still leaves the output affected, particularly when artifact amplitude far exceeds the true signal. Switching between channels every two seconds, as prior work did, throws away information from channels not currently judged best, which hurts when that quality judgment is imperfect. The paper positions fusion as keeping all channels in play while removing artifacts rather than blending them.
Training targets are synthesized from ECG timing
Supervised training needs a clean PPG to aim at, which the recordings do not provide. The authors instead built reference signals from the data itself. Using R-peaks detected on the reference ECG, they split the PPG into beat-length sections, down-sampled each to 100 samples, and standardized them. Sections that correlated sufficiently with a leaning triangle wave (r > .8) were averaged into a template, which excluded distorted and noisy beats.
Templates were computed in 5-minute windows to track morphology drift, and sections corresponding to heart rates below 40 or above 185 bpm were disregarded. The final reference signal concatenates one pattern per R-R interval, using a weighted average of the two nearest templates with weights inversely proportional to interval distance, then resamples each pattern to match the interval. The network itself is a U-Net with four down-sampling and four up-sampling layers, 32 to 256 kernels, three-channel input, and a single fused output, trained with Adam and selected by lowest L1 validation loss.
Fusion beats every single wavelength on heart rate
Evaluation used 4,000 evenly spaced 8-second sections per participant, totaling 8.9 hours of measurements, with five-fold cross-validation across the ten subjects and two held-out test participants per fold. On heart rate, the fused signal's mean error was 4.5 bpm and median error 1.8 bpm. Green alone reached 5.9 bpm mean and 1.9 bpm median; red reached 13.4 bpm mean and 9.2 bpm median; infrared reached 11.1 bpm mean and 6.9 bpm median.
The mean-median split is worth pausing on. Green's median error of 1.9 bpm is close to the fused signal's 1.8 bpm, while its mean is 1.4 bpm higher — indicating that green's disadvantage comes mainly from a tail of badly corrupted sections rather than uniformly worse tracking. The paper notes green was the best individual wavelength, so the 23% reduction is measured against the strongest single-channel baseline, not an average of the three.
Waveform correlation shows a wider margin than error
Morphology results separate the methods more sharply than heart rate error does. Against the synthesized reference, the fused signal averaged a correlation of 0.831 across participants, compared with 0.645 for green, 0.487 for red, and 0.608 for infrared. Mean absolute error was 0.410 and root mean square error 0.368 for the fused signal.
These two result sets answer different questions and should not be substituted for one another. Correlation and MAE describe how closely the recovered waveform matches the reference shape; bpm error describes how well peak timing survives into a heart-rate estimate. A signal can track morphology well yet still yield imperfect intervals, and the paper reports both because the method's stated purpose is morphology recovery for arbitrary downstream measures, with heart rate as one demonstration.
The evidence covers ten lightly pigmented participants
The dataset is small and narrow: 10 participants, all rating 1–3 on the Fitzpatrick scale, with none in the 4–6 range of darker skin tones. The authors flag this directly, noting that PPG measurement is more challenging for people with darker skin tones and that broader data collection is needed to test how wavelength choice and body location affect results. The 4.5 bpm figure therefore describes this group during everyday and motion-afflicted activities outside the laboratory, not wearables generally.
Two further boundaries sit inside the pipeline. Sections with inconsistent or noisy ECG were removed before evaluation, and the interbeat-interval filter was deliberately lenient — retaining intervals where the ratio of minimum to maximum IBI exceeded 0.51 across at least five consecutive beats — so that every segment retained a signal for evaluation. That choice keeps difficult sections in the comparison, which is consistent with the paper's aim of testing robustness, but it means the reported errors come from a specific filtering regime rather than an unprocessed stream.
The gain concentrates in artifact-heavy segments
The most decision-relevant reading of this paper is not that fusion lowers average error by 1.4 bpm. It is where that reduction comes from. Green's near-parity with fusion on median error, alongside its higher mean, points to corrupted segments as the source of the gap — exactly the conditions the method was designed for, since averaging and channel-switching both degrade when artifacts dominate or quality estimates fail.
For a device that already carries green, red, and infrared channels, the paper's evidence supports treating fusion as a way to salvage the hard sections rather than to lift typical-case accuracy. The per-participant spread reinforces this: fused mean errors ranged from 2.0 to 10.0 bpm across the ten participants, so the aggregate 4.5 bpm hides substantial individual variation. Whether that trade is worth a learned fusion stage depends on how much a given application cares about the artifact-heavy tail — and on evidence from skin-tone groups this study did not include.
Sources used in this answer
Tri-Spectral PPG: Robust Reflective Photoplethysmography by Fusing Multiple Wavelengths for Cardiac Monitoring
A U-Net fusing green, red, and infrared reflective PPG into one recovered waveform yielded heart-rate mean error of 4.5 bpm versus 5.9 bpm from green alone (23% lower) across 10 participants and 13 hours of free-living recordings.