Can multi-site PPG sensors make heart rate detection more robust?

Motion artifacts differ by body location

Photoplethysmography (PPG) estimates heart rate optically: a light source and photodiode measure blood-volume pulses, and the interval between systolic peaks gives the beat-to-beat timing from which heart rate is computed. The paper's premise is that motion affects these readings differently depending on where a device sits on the body. A wrist sensor swinging during a walk and a forehead sensor undergoing the same walk do not see the same artifact at the same moment.

That premise is what makes fusion plausible. If every sensor failed at once, combining them would add little. If their failures are partly independent, a method that favors whichever sensor is currently clean can produce a better signal than any single device. The authors build their method on this location-specific view of motion artifacts rather than on adding accelerometers or extra light sources inside one device.

Quality scores decide which sensor dominates

The method first detects systolic peaks in each PPG stream after a 0.6–3.3 Hz bandpass filter, discarding peaks whose intervals would imply a heart rate above 185 bpm. It keeps interbeat intervals using a deliberately lenient filter (at least five consecutive intervals with minIBI/maxIBI > 0.51), so that less accurate stretches survive to be weighed later instead of being deleted. Heart rate is then estimated over a 30-second moving window.

Signal quality is scored per beat by comparing each segment against a template built separately for each sensor location, since waveform shape differs by body site. Segments are down-sampled to 40 samples and z-scored, the template is formed from segments correlating above r > .8 with a leaning triangle wave, and the 500 best-correlating segments are kept. A segment's quality is its correlation with that template.

Fusion then aligns peaks across sensors within ±150 ms to account for pulse travel time through the arterial tree, interpolates the 30-second quality estimates to avoid abrupt jumps, and combines signals with weights raised to the sixth power. That exponent amplifies small correlation differences, so a clearly better sensor dominates the output; normalization keeps output amplitude independent of how many sensors contribute.

The head was the strongest single site

Before fusion, the authors measured each location alone against Lead I ECG recorded at the sternum. Mean error was 4.38 bpm at the head, 5.57 bpm at the sternum, 7.63 bpm at the ankle and 8.22 bpm at the wrist. The ordering plausibly tracks how much motion each site experienced during the outdoor activities, though the paper presents this as an interpretation rather than a measured motion comparison.

This baseline matters for reading the headline number. The 46% reduction is relative to the head, the best single device, not to the wrist. Against the wrist, the paper reports a 71% lower mean error and 91% lower median error. Quoting the wrist comparison alone would overstate what fusion adds over the strongest available single site.

Fusion beat ICA on mean error in every setting

The comparison baseline is independent component analysis (ICA), a standard approach that separates mixed signals into components and takes the best output. The authors tested both methods with two, three and four devices. ICA never produced a lower mean error than the head sensor alone: 5.04 bpm with two devices, 5.72 with three and 5.01 with all four, against 4.38 for the head. Its median errors were lower than any single PPG sensor, which the authors read as ICA helping in most cases but failing badly in some.

The proposed fusion produced 3.52 bpm mean error with two devices, 3.01 with three and 2.37 with all four. Median errors were 1.47, 0.99 and 0.53 bpm respectively. Every configuration beat every single sensor on both mean and median. The two-device result is the practically relevant one for consumer hardware: a wrist-plus-head pairing, which the authors describe as a smartwatch plus smartglasses scenario, still cut mean error 20% below the head alone and median error 26% below.

Gains appear at low error percentiles too

Sorting 30-second segments by error shows where the improvement comes from. The reduction is visible not only in the high-error percentiles, where individual sensors struggle, but also in the low-error percentiles, where the contributing PPG signals were already accurate. The authors use this to argue the method helps across the signal-quality range rather than only rescuing bad moments.

The paper also shows qualitative examples where individual sensors had high heart-rate error and the fused output recovered a plausible PPG morphology. These examples illustrate the mechanism but are not a separate quantitative result; the error table and percentile plot carry the statistical claim.

One day, fourteen people, real activities

The evaluation dataset comes from 14 participants who each wore four standalone PPG devices at the forehead, sternum, ankle and wrist for about 13 hours, plus a sternum ECG for ground truth. The protocol included a minivan ride, cable car and train travel to 3460 m, museum and outdoor walking, stairs, lunch and rest, totaling roughly 182 hours of synchronized signals at 128 Hz. Devices were synchronized offline to 33 ms accuracy.

This design is the paper's main strength over lab-based PPG tests: the signals include genuine daily activity and changing altitude and context. It is also the boundary. Fourteen participants is a small sample, all data come from a single day-long excursion, and the four fixed body sites are research devices rather than consumer products. The reported errors describe this cohort and protocol; the paper does not establish performance across broader populations, skin tones, clinical conditions or longer wear.

Fusion pays off most when one site is weak

The paper's practical priority is a comparison, not a blanket recommendation. Fusion's advantage over the best single sensor shrinks as the sensor set improves: 46% mean-error reduction with all four sites, 31% with three, and 20% with the wrist-plus-head pair. The head sensor alone remains a strong option when only one site is available, and the wrist alone is the weakest choice in this dataset.

The decision the evidence supports is about which sites to combine, not whether more sensors are always better. Adding sites lowered error in every configuration tested, and the method is described as agnostic to the number of sources and body sites, but the measured benefit of the second and third sensors is smaller than the first fusion step away from a single weak site. For a reader weighing a multi-site design, the wrist-plus-head result is the most relevant threshold: it is the configuration closest to existing consumer form factors, and it still delivered a lower mean and median error than the head sensor by itself.

Sources used in this answer