2025 rPPG Equity Report: Accuracy and Performance Across Diverse Populations

In December 2020, Sjoding et al. published a study in the New England Journal of Medicine that sent shockwaves through healthcare technology: pulse oximeters — devices used billions of times per year in clinical settings — were significantly less accurate on patients with darker skin tones. Black patients were three times more likely to have occult hypoxemia (dangerously low oxygen levels) missed by their pulse oximeter readings. The finding wasn't new to researchers who had documented the bias for years, but its publication in the NEJM forced a broader reckoning.

That reckoning extends directly to rPPG. Camera-based vital sign measurement relies on the same fundamental optical principles as pulse oximetry — detecting changes in light absorption and reflection caused by blood flow beneath the skin. Melanin, the pigment that determines skin color, absorbs light across visible wavelengths and modulates the signal that rPPG algorithms use. If the field doesn't address this head-on, it risks replicating the same disparities that contact-based devices have only recently been forced to confront.

"Racial bias in pulse oximetry has been hiding in plain sight for decades. As new technologies like camera-based vital signs emerge, we have an obligation — and an opportunity — to build equity into the technology from the start." — Sjoding et al., New England Journal of Medicine (2020)

The Optical Challenge of Melanin

Understanding why skin tone affects camera-based measurement requires understanding the physics. The rPPG signal originates from subtle changes in light absorption caused by blood volume fluctuations with each heartbeat. Melanin, concentrated in the epidermis, absorbs light broadly across the visible spectrum — particularly at shorter wavelengths (blue and green). This absorption has two effects on rPPG:

Signal attenuation: Higher melanin content absorbs more of the incident light before it reaches the blood-containing dermal layers, reducing the amplitude of the pulsatile signal. Verkruysse et al. (2008) noted this effect in their foundational rPPG paper.

Reduced signal-to-noise ratio: With less pulsatile signal reaching the camera, the ratio of cardiac signal to noise (from motion, lighting, camera sensor noise) decreases, making accurate extraction more challenging.

These are not theoretical concerns. Multiple research groups have quantified the performance gap:

Performance Across Skin Tones: What the Research Shows

Study	Technology	Metric	Lighter Skin Performance	Darker Skin Performance	Gap
Nowara et al. (2020)	rPPG (multiple algorithms)	HR MAE	2-4 BPM	5-12 BPM	2-4x worse
Ba et al. (2023)	rPPG deep learning	HR MAE	2-3 BPM	3-6 BPM	1.5-2x worse
Sjoding et al. (2020)	Pulse oximetry	Occult hypoxemia missed	3.6%	11.7%	3.2x worse
Bent et al. (2020)	Wearable PPG (smartwatch)	HR MAE	2-3 BPM	4-8 BPM	2x worse
Fallow et al. (2013)	rPPG (early algorithms)	SNR	Higher	40-60% lower	Significant
Wang et al. (2017)	rPPG POS algorithm	HR correlation	0.95+	0.88-0.92	Narrower gap

Sources: As cited, published in IEEE, NEJM, Nature Digital Medicine, and related journals.

Two patterns emerge. First, the bias is real and measurable across both contact and contactless optical devices — this isn't an rPPG-specific problem but a physics-level challenge affecting all optical physiological sensing. Second, the gap is narrowing as newer algorithms are specifically designed for cross-skin-tone robustness.

Factors Beyond Skin Tone

While melanin content receives the most attention, rPPG performance varies across other demographic and physiological factors:

Age: Older adults have thinner skin but may have reduced peripheral perfusion, and age-related vascular changes affect pulse wave characteristics. Mcduff et al. (2023) noted that algorithm performance in elderly populations requires specific validation.

Gender: Differences in facial fat distribution, skin thickness, and hormonal effects on vasomotion can influence signal quality. Published research shows small but measurable gender-related performance differences.

BMI and facial structure: Higher BMI and different facial structures affect the region of interest selection and signal quality. Algorithms trained primarily on one facial structure type may underperform on others.

Medical conditions: Anemia reduces hemoglobin (weakening the pulsatile signal), peripheral vascular disease reduces skin perfusion, and conditions causing edema alter optical properties. These clinical confounders disproportionately affect certain populations.

Algorithmic Approaches to Improving Equity

Researchers are actively developing methods to reduce performance disparities:

Diverse training data: The most straightforward approach — and arguably the most impactful — is training algorithms on datasets that proportionally represent diverse skin tones. Early rPPG datasets (MAHNOB-HCI, UBFC-rPPG) were predominantly light-skinned. Newer datasets like MMPD (Multi-Modal Physiological Dataset) and efforts by researchers at UCLA and TU Eindhoven are actively addressing this gap.

Skin-tone-adaptive algorithms: Wang et al. (2017) POS algorithm projects color signals onto a plane orthogonal to the skin color vector, inherently reducing melanin-related bias. Nowara et al. (2020) demonstrated that algorithms could be explicitly conditioned on estimated skin tone to adjust processing parameters.

Deep learning generalization: Ba et al. (2023) showed that modern deep learning rPPG models, when trained on diverse data, naturally learn to compensate for melanin-related signal differences — achieving substantially narrower performance gaps than traditional signal processing approaches.

Multi-region and multi-wavelength processing: Analyzing signals from multiple facial regions (forehead, cheeks, periorbital area) and leveraging all color channels provides redundancy that helps maintain accuracy when individual channels are degraded by melanin absorption.

Synthetic data augmentation: Researchers have explored using color-space transformations and generative models to augment training data with simulated darker skin tones, partially mitigating the data scarcity problem — though this approach has limitations compared to real diverse data.

Lessons from Pulse Oximetry's Reckoning

The pulse oximetry bias story offers important lessons for rPPG:

Decades of inaction: The Sjoding et al. (2020) finding confirmed what researchers like Bickler et al. (2005) and Feiner et al. (2007) had documented years earlier. The FDA has since issued guidance recommending diverse clinical validation, but the slow response cost patient safety.

Regulatory change: In November 2022, the FDA convened an advisory committee specifically on pulse oximetry accuracy across skin tones, signaling that future device clearances will likely require diverse population validation — a standard that rPPG devices should proactively meet.

The transparency imperative: Sjoding's paper changed practice because it made the bias visible and quantified. rPPG developers have an opportunity to build transparency into their validation from the start — reporting performance by skin tone, age, and gender rather than aggregate accuracy alone.

What Equitable rPPG Validation Should Look Like

Based on the pulse oximetry experience and current research, equitable rPPG validation should include:

• Balanced representation across Fitzpatrick skin types I-VI in validation cohorts
• Disaggregated reporting of accuracy metrics by skin tone, age, gender, and clinically relevant subgroups
• Real-world conditions including varied lighting environments, which interact with skin tone effects
• Clinical population inclusion — patients with anemia, peripheral vascular disease, and other conditions that affect optical signals
• Longitudinal validation across different times, conditions, and settings
• Transparent benchmarking against established contact-based devices across the same diverse cohorts

The Road Ahead

The rPPG field has an opportunity that pulse oximetry missed: building equity into the technology before widespread clinical deployment, rather than discovering bias after decades of use. The research community is responding — newer algorithms show meaningfully narrower performance gaps, diverse datasets are being built, and the conversation about equitable validation is happening early.

Companies like Circadify are developing rPPG technology with equity as a core design principle, prioritizing diverse population validation and transparent performance reporting. The goal isn't just camera-based vital signs that work — it's camera-based vital signs that work equitably for everyone.

Frequently Asked Questions

Does skin tone affect rPPG accuracy?

Published research shows that earlier rPPG algorithms showed measurable performance differences across Fitzpatrick skin types, with accuracy typically decreasing for darker skin tones. Newer algorithms trained on diverse datasets have significantly narrowed this gap, though ongoing validation remains essential.

How does rPPG compare to pulse oximeters for bias?

FDA-cleared pulse oximeters have documented accuracy disparities across skin tones, with Sjoding et al. (2020) in NEJM showing that Black patients were three times more likely to have occult hypoxemia missed. rPPG faces similar optical challenges but benefits from newer algorithmic approaches and larger diverse training datasets.

What is being done to improve rPPG equity?

Researchers are building larger, more diverse training datasets, developing skin-tone-adaptive algorithms, and establishing validation standards that require performance reporting across demographic subgroups.

Related Articles

• What is rPPG Technology? — A complete overview of remote photoplethysmography and the science behind camera-based vital sign measurement.
• Contactless Heart Rate Monitoring — Heart rate detection performance across skin tones is the most studied equity dimension of rPPG.
• Contactless SpO2 Monitoring — SpO2 estimation faces the most significant skin-tone-related accuracy challenges of any rPPG measurement.