Vascular aging makes large arteries stiffer and alters the shape of the pulse wave. The gold standard for diagnosis is pulse wave velocity between the carotid and femoral arteries, but it requires complex measurements. Photoplethysmography (PPG) from wearable devices is simpler, but it is usually recorded at the wrist, where the signal is distorted. Nearby in the archive are studies that look for the pulse in different places: a ring on the finger, a thread on the wrist, heart rhythm topology, neural network explanations, a blood pressure chip. All of them try to squeeze more information out of the pulse, but almost none compare measurement sites directly. Our work sits off to the side: the nearest neighbor in meaning is at a distance of 0.516, and there is nothing right next to it.
They took three datasets: 1) short-wave infrared spectroscopy (fNIRS) with sensors at the temple (13 people), 2) a dedicated PPG sensor at the temple (126 healthy adults aged 20–56), 3) wrist PPG from a MAX30101 device (240 healthy adults aged 30–60). From each pulse window they extracted 89 waveform features: time intervals, amplitudes, areas, derivatives, and their ratios. Features were classified as cardiac, vascular, or mixed using synthetic data from a model in which cardiac and vascular parameters were varied independently. Then for each feature they assessed the association with age via bootstrap correlations (1000 iterations with age stratification) and selected robust correlates.
At the temple they found 14 robust age correlates, while at the wrist only 3. The strongest feature at the temple is (Ad-Ab)/Aa, the normalized contrast between early and late systolic curvature of the second derivative. Its correlation with age is r = −0.679 with a selection probability of 70%. At the wrist, the strongest feature Ad/Aa had r = −0.413 and a probability of 10.1%. The difference in association strength between sites was large for 8 features (Cohen's d > 1.5). Principal component analysis showed that explaining 85.3% of the variance at the temple requires 5 components, while at the wrist 3 components suffice for 100%. The mean pairwise correlation among the 89 features was slightly lower at the temple (0.372 vs. 0.385), indicating less redundancy.
The work shows that the PPG measurement site is not neutral: it determines which physiological processes are visible in the waveform. At the temple, age manifests through several independent axes (curvature, area, timings), whereas at the wrist it does so through fewer and weaker ones. This differs from neighboring studies that either improve models at a single site or seek explanations for neural networks without comparing anatomical locations. Here, for the first time, the temple and wrist are systematically compared using the same feature set, and it is shown that a proximal site (closer to the heart) yields a richer and more interpretable picture.
The topic could move toward head-worn devices: glasses, headphones, headbands. If the temples really do provide a cleaner signal about vascular aging, this opens a path to continuous monitoring of arterial stiffness without a cuff. But first, paired measurements in the same people with identical sensors are needed to rule out the influence of different cohorts and sampling rates.
Will influence the development of wearable sensors for cardiovascular monitoring, the choice of measurement site in clinical studies, and the interpretation of smartwatch data.
Conduct a within-subject study with simultaneous PPG recording at the temple and wrist, compare with reference pulse wave velocity and direct measurements of pressure and flow.
Connection to an unsolved problem in circulatory physiology: how local properties of the vascular bed and reflected waves are encoded in the shape of the peripheral pulse. The work shows that the same feature can have different physiological origins depending on the measurement site, which complicates the interpretation of biomarkers and calls for multi-site models.
The short-wave fNIRS channels that are usually discarded as noise unexpectedly yielded age correlations up to r = −0.852 at a wavelength of 830 nm.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.