The human body is a noisy machine. Every time your heart beats, it triggers a cascade of subtle physical events: your skin changes color slightly as blood oxygenates, your head bobs from the force of the carotid pulse, and your chest vibrates. For researchers, the challenge isn't finding these signals—it's separating them from the "noise" of daily life. A person breathing, walking, or even sitting under a flickering fluorescent light creates signals orders of magnitude stronger than a heartbeat.
To solve this, a broad research direction has emerged that uses physiological priors—hard-coded biological constraints—to "mask" the noise and isolate the signal. By anchoring algorithms in the known limits of human biology, researchers can extract clinical-grade data from everyday sensors like webcams, smartwatches, and even radio waves.
The Frequency Bottleneck
The most fundamental constraint is the human heart rate itself. A resting heart doesn't beat at 10 Hz, nor does it stop for minutes at a time. Researchers like Xiaobai Li and Guha Balakrishnan utilize this by applying strict bandpass filters to raw data.
In Detecting Pulse from Head Motions in Video, Balakrishnan demonstrates that while a head might move for many reasons, only the movement within the 0.75 to 5 Hz range (roughly 45 to 300 beats per minute) is likely to be cardiac. By restricting the search space to this "biological window," the system can ignore low-frequency sways from posture and high-frequency jitter from the camera sensor. Li further refined this in Remote Heart Rate Measurement from Face Videos under Realistic Situations, using these bounds to make remote photoplethysmography (rPPG) work in "wild" environments where lighting and movement would otherwise overwhelm the sensor.
Harmonic Separation and the Decay Effect
A more sophisticated constraint involves the relationship between different bodily rhythms. Respiration is the "enemy" of cardiac sensing because the chest movement from a breath is much larger than the vibration of a heartbeat. However, Zhi Lu and Bin-Bin Zhang identified a physical property that allows for their separation: differential harmonic decay.
In Monitoring Long-term Cardiac Activity with Contactless Radio Frequency Signals, the researchers show that while respiratory motion is large, its signal strength decays rapidly as you look at higher-order harmonics in the frequency spectrum. Heartbeat signals, though smaller, are sharper and more impulsive, meaning their harmonics persist much longer. By looking at the 10th harmonic and beyond, the system effectively "filters out" the breathing noise because the breathing signal has already faded to zero, leaving the heartbeat isolated and clear.
Biological Feasibility as a Filter
Beyond frequency, researchers use the temporal logic of biology to reject artifacts. A human heart rate cannot instantaneously jump from 60 to 180 BPM in a single beat.
Umang Yadav and Sherif N. Abbas apply "physiological outlier rejection" in Evaluation of PPG Biometrics for Authentication in Different States. If a sensor detects a pulse interval that is biologically impossible, the system doesn't just average it—it discards it entirely. Similarly, Yuntong Zhang uses a multi-stage pipeline that applies "temporal smoothing" based on the maximum possible change in heart rate per second. This ensures that a sudden glitch in a smartwatch sensor isn't misinterpreted as a cardiac event.
Contextual Gating
Finally, researchers are increasingly using behavioral constraints to improve signal quality. The most effective way to reduce noise is to wait for the body to be quiet.
Arash Alavi and colleagues at Stanford demonstrated this in their Real-time Alerting System for COVID-19. By monitoring resting heart rate (RHR) exclusively during sleep, they removed the "noise" of exercise, caffeine, and stress. This established a stable physiological baseline, allowing them to detect the subtle RHR elevations associated with early infection days before symptoms appeared.
Why It Works
This direction works because it shifts the burden from the sensor to the model. Instead of trying to build a "perfect" camera or radar that can see through anything, researchers like Zhaodong Sun and Xiaobai Li build models that "know" what a human looks like. In Contrast-Phys, they use an unsupervised loss function that penalizes any signal that doesn't fit a biologically plausible heart rate interval. By treating the human body's known constraints as a mathematical guide, the field has moved from simple pulse counting to capturing high-fidelity waveforms that were once thought to require a hospital-grade ECG.