Mobile Health Sensors: Heart Rate, SpO2 and Glucose Monitoring Technology

Photoplethysmography has become the default optical interface between human physiology and the smartphone. Yet heart rate is the easy part. SpO2 demands dual-wavelength precision, and non-invasive glucose remains the most difficult sensing problem in consumer electronics.

1. The PPG Foundation: Light, Tissue, and Blood

Every mobile health sensor for cardiovascular metrics starts with photoplethysmography (PPG). An LED illuminates capillary beds — typically on the wrist, fingertip, or ear — and a photodiode measures back-scattered light modulated by volumetric changes in blood. During systole, increased blood volume absorbs more light; during diastole, less. The resulting AC signal rides on a large DC baseline from tissue, bone, and venous blood.

In smartphones and watches, reflective PPG dominates. Unlike transmissive pulse oximeters that clamp a finger between LED and detector, reflective designs place LEDs and photodiodes on the same plane, separated by a light barrier to prevent direct crosstalk. Green light (~525-530 nm) is preferred for heart rate because hemoglobin absorbs strongly at this wavelength, yielding high pulsatile amplitude even at low perfusion. Red (~660 nm) and infrared (~940 nm) penetrate deeper and are mandatory for oximetry and emerging glucose research.

The signal chain is deterministic. LEDs are driven with pulsed current (20-100 mA, duty-cycled to <5%) to limit power and heat. The photodiode feeds a transimpedance amplifier, then a high-resolution ADC (typically 22-24-bit) inside an integrated AFE such as the Maxim MAX86140, TI AFE4404, or custom silicon in Apple/Samsung wearables. Sampling at 50-200 Hz provides enough bandwidth for heart rate while keeping power under 1 mW per channel. Ambient light cancellation — sampling between LED pulses and subtracting — is non-negotiable under sunlight that can saturate the detector 1000x over the PPG signal.

Reflective vs Transmissive Coupling

Smartphones present a unique coupling challenge. When a user presses a fingertip over the rear camera/flash module — a trick used by Samsung's Heart Rate Sensor (S5/S9 era) and many third-party apps — the camera acts as the photodiode and the LED flash as the illuminator. This is inexpensive but suffers from auto-exposure hunting, inconsistent pressure, and poor light sealing. Dedicated modules under the display or on the side frame, now common in flagship watches and some phones, use molded optics and black silicone light barriers to achieve stable coupling independent of camera ISP behavior.

For deeper context on how camera ISP tuning affects optical measurements, see our analysis of AI image processing in mobile cameras, where auto-white balance and tone mapping introduce similar calibration pitfalls.

2. Heart Rate: From Raw PPG to BPM

Heart rate appears simple: detect peaks, measure intervals. In practice, motion artifact is the dominant error source. Walking introduces 0.5-3 Hz acceleration that overlaps the cardiac band (0.8-3.5 Hz, or 48-210 BPM). Without compensation, error can exceed 15 BPM during activity.

Flagship algorithms fuse PPG with the IMU. A 3-axis accelerometer provides a noise reference for adaptive filtering — normalized least-mean-squares (NLMS) or recursive least squares — effectively subtracting the motion-correlated component. More recent designs run a small neural network (10-50 kB quantized) on the sensor hub to separate cardiac and motion spectra, trained on diverse skin tones and activity types. The sensor hub, not the AP, runs this continuously at <2 mW, waking the main processor only for UI updates.

Pressure and perfusion variability matter. Light pressure blanches capillaries, reducing AC amplitude; excessive pressure collapses them. The best wearables use a pressure-insensitive optical stack and guide the user to a "snug but not tight" fit. Green PPG achieves ±3 BPM accuracy (mean absolute error) at rest on the wrist, degrading to ±5-7 BPM during running. Fingertip measurement is more accurate but less continuous, which is why continuous monitoring has migrated to the watch.

Audio and haptics share the same hub scheduling constraints. Similar to smartphone speaker volume and quality testing, health sensor validation requires anechoic-like control of external stimuli — in this case, motion tables and optical phantoms rather than audio chambers.

3. SpO2: Why Two Wavelengths Changes Everything

Peripheral oxygen saturation relies on the differential absorption of oxyhemoglobin (HbO2) and deoxyhemoglobin (Hb) at red and infrared wavelengths. Hb absorbs more red (~660 nm), HbO2 absorbs more IR (~940 nm). The classic ratio-of-ratios method computes:

This appears trivial, but R is exquisitely sensitive to error. A 1% error in R translates to ~2-2.5% SpO2 error due to the steep calibration curve. That error budget is consumed quickly: LED wavelength drift with temperature (~0.2 nm/°C for AlGaAs red LEDs), photodiode dark current, quantization noise, and critically, optical path difference between red and IR. Red and IR photons sample different tissue depths; motion or pressure changes alter the path mismatch, shifting R without any physiological change.

Mobile implementation therefore requires per-wavelength automatic gain control and temperature compensation. Dual-PD designs — one broadband, one with IR-cut filter — help isolate wavelengths without time-division penalty. Calibration is performed against arterial blood gas or FDA-cleared transmissive oximeters across 70-100% SaO2 induced via breath-hold protocols, generating the k-coefficients per device model. Regulatory-cleared wearables (Apple Watch Series 9, Masimo W1, Samsung Galaxy Watch) claim ±2-3% Arms accuracy in the 70-100% range under motion-free conditions, but validation under motion shows degradation to ±4-5%, and low perfusion (<0.2% modulation) remains a failure mode.

Mobile SpO2 is not continuous by default for power and regulatory reasons. Most devices offer spot-checks (15-30 seconds fingertip or wrist) plus background sampling during sleep when motion is minimal. The display implications are also non-trivial: prolonged OLED PWM dimming at low brightness can interact with optical sensor sampling. Our testing in display PWM dimming health impact shows why sampling windows are synchronized to display blanking intervals on some integrated designs.

4. Glucose Monitoring: The Non-Invasive Frontier

Glucose is orders of magnitude harder. Blood glucose modulates tissue scattering and absorption in the near-infrared (NIR), but the signature is weak — glucose absorption peaks around 1550-1700 nm and 2100-2300 nm overlap heavily with water, lipids, and proteins. At 940 nm, the glucose absorption coefficient is ~0.0001 mm⁻¹, dwarfed by water (~0.05 mm⁻¹). Detecting a 20 mg/dL change requires resolving <0.05% optical change against motion and hydration variables 100x larger.

Three optical approaches are being pursued for mobile form factors:

4.1 Multi-Wavelength NIR Spectroscopy

Extending PPG to 3-5 NIR wavelengths (940, 1300, 1450, 1550 nm) and applying multivariate regression (PLS, or tiny neural nets) to extract glucose-correlated scattering. Rockley Photonics' abandoned Bioptx platform and current efforts by Apple and Samsung prototypes use silicon photonics + InGaAs detectors to reach beyond 1300 nm, where silicon PDs lose sensitivity. The challenge is path length uncertainty: without knowing photon path through the dermis, absolute concentration cannot be derived, only trends.

4.2 RF / Impedance Spectroscopy

At 1-60 GHz, glucose alters dielectric permittivity of blood/interstitial fluid. RF sensors integrated into watch backs measure S11/S21 shift. Less sensitive to skin tone than optics, but highly sensitive to hydration, temperature, and contact pressure. Current lab sensitivity is ~5-10 mg/dL under controlled conditions, insufficient for insulin dosing.

4.3 Raman and Photoacoustic Concepts

Raman scattering provides molecular specificity but needs laser excitation and cooled detectors — impractical for phones today. Photoacoustic (pulsed laser + acoustic detection) shows promise but requires >1W peak optical power.

For fine-grained signal measurement methodology that also applies to optical stacks, see smartphone audio quality measurement — the same principles of THD+N and SNR budgeting govern AFE validation.

5. Integration: AFE, Power, and Thermal Reality

Health sensors live on the power and thermal margins. LEDs are the dominant consumer; a typical HR measurement at 100 mA pulses, 25% duty, consumes more than the entire sensor hub. The AFE's LED driver efficiency (often <85% at low current) and boost converter ripple directly couple into optical noise if not synchronized to sampling. Layout isolation — separating analog ground from digital switching noise, and shielding the TIA input — is critical; a 10 µV ripple at the TIA input can eclipse the 50 µV cardiac AC signal.

Thermal constraints are regulatory as well as ergonomic. IEC 60601-1 and FDA guidance limit prolonged skin contact temperature to 41°C. At 100 mA LED drive, localized heating can reach 2-3°C above ambient in 30 seconds, biasing SpO2 wavelength coefficients. Good firmware throttles drive current when skin temperature (measured via NTC near the optics) exceeds 38°C, trading SNR for safety.

Coexistence with other subsystems matters. The optical window must survive oleophobic coatings and scratch resistance without altering NIR transmission. On phones with under-display sensors, the OLED stack introduces additional attenuation and angle-dependent color shift that must be calibrated per display lot. Under-display PPG prototypes therefore use 850 nm IR that better penetrates OLED layers, at the cost of melanin sensitivity.

6. Accuracy, Regulation, and the Skin Tone Problem

PPG bias with skin tone is physics, not just algorithm tuning. Melanin absorbs green and red light strongly, reducing DC throughput and AC amplitude by 2-5x on Fitzpatrick V-VI skin vs. I-II at green wavelengths. This lowers SNR and, if AFE gain is not adapted, drives quantization noise up. Worst-case error in early devices exceeded 8 BPM and 4% SpO2 on darker skin during motion.

Mitigations now standard in Tier-1 vendors include: automatic LED current scaling based on DC level, adaptive integration time, and training datasets balanced across Fitzpatrick scales with spectrophotometer-verified skin typing. FDA's 2024 draft guidance explicitly requires subgroup accuracy reporting by pigmentation, and 510(k) submissions for oximeters must include at least 25% subjects with dark pigmentation. For integrators, per-unit factory calibration of LED wavelength and PD responsivity reduces unit-to-unit R error from ~2% to <0.5%.

From a software architecture perspective, treat health data as SensorManager's most sensitive tier: permission-gated, on-device processed, with HR/SpO2 history encrypted at rest. The OS should never expose raw AFE counts to third-party apps — only derived, confidence-qualified values — to prevent unvalidated diagnostic claims and to preserve the calibration chain.

Bottom line: heart rate is solved for wellness use, SpO2 is solved under constrained conditions but not yet robust continuous, and non-invasive glucose remains research-grade. The most credible near-term path pairs minimally invasive continuous glucose monitors (microneedle patch) with the phone as gateway and trend analyzer, while optics mature to provide non-invasive trend support rather than bolus dosing decisions.