Can PPG sensors accurately measure blood glucose?

Current PPG-based glucose estimation remains investigational and cannot match the accuracy of fingerstick glucometers or CGMs. The best laboratory studies report Clarke Error Grid Zone A rates of 80-92% and mean absolute relative differences (MARD) of 15-25%, but these results have not been reliably replicated in independent clinical validation. The fundamental challenge is that glucose has a very weak direct optical signature in the wavelengths PPG sensors use, so most approaches rely on indirect physiological correlations rather than direct glucose measurement.

Why is non-invasive glucose monitoring so difficult with PPG?

Glucose at physiological concentrations (70-180 mg/dL in healthy individuals) causes extremely small changes in tissue optical properties at the green and near-infrared wavelengths used by standard PPG sensors. These changes are orders of magnitude smaller than variations caused by motion, skin temperature, hydration, melanin content, and hemodynamic fluctuations. Separating the glucose signal from these confounders remains the central unsolved problem. Additionally, the correlation between PPG waveform features and glucose levels is largely indirect, mediated through glucose's effects on blood viscosity and vascular tone.

What accuracy is needed for a non-invasive glucose monitor to be clinically useful?

The FDA generally requires a MARD below 15% for continuous glucose monitors and at least 95% of readings in the Clarke Error Grid Zone A+B for self-monitoring devices. Current PPG-based approaches typically achieve MARD values of 15-25% under controlled conditions, which degrades further in free-living conditions. For insulin dosing decisions, accuracy requirements are particularly stringent because errors can lead to dangerous hypo- or hyperglycemia. Most researchers agree that PPG-only approaches are unlikely to meet these thresholds without supplementary sensing modalities.

Which companies are developing PPG-based glucose monitors?

Several companies have pursued optical non-invasive glucose monitoring, though none have achieved FDA clearance for a PPG-only device. Apple has filed extensive patent portfolios on multi-wavelength optical glucose sensing. Samsung has published research on Raman spectroscopy combined with PPG. Rockley Photonics developed a multi-spectral sensor chip targeting glucose among other biomarkers. Numerous startups have claimed breakthroughs but failed independent validation. The consensus in the biosensing community is that a consumer-grade non-invasive glucose monitor remains years away from clinical reality.

Non-Invasive Glucose Monitoring via PPG: Current State of the Art

Non-invasive glucose monitoring through photoplethysmography remains one of the most pursued and most elusive goals in biosensing. Despite decades of research and hundreds of published studies, no PPG-based glucose monitoring device has achieved FDA clearance or demonstrated clinical-grade accuracy in independent validation. Yet the prize is enormous: replacing painful fingerstick testing and expensive continuous glucose monitors (CGMs) for the estimated 537 million people worldwide living with diabetes would fundamentally transform disease management.

This article provides a rigorous technical review of the current approaches, the evidence supporting them, the fundamental limitations, and the realistic trajectory toward clinical viability. For foundational context on how PPG sensors work, see our introduction to PPG technology.

The Physiological Basis for PPG-Glucose Correlation

Understanding why PPG might contain glucose information requires examining the multiple pathways through which blood glucose affects the optical and hemodynamic properties that PPG measures.

Direct Optical Effects

Glucose absorbs near-infrared (NIR) light at specific wavelengths, primarily in the 1000-2500 nm range with absorption peaks near 1536 nm, 2100 nm, and 2270 nm corresponding to O-H and C-H bond vibrations (Amerov et al., 2004). However, the absorption coefficients at physiological glucose concentrations (4-10 mmol/L) are extremely small relative to water absorption, which dominates tissue optical properties. At the shorter wavelengths used by standard PPG sensors (green ~530 nm, red ~660 nm, infrared ~940 nm), the direct optical effect of glucose is negligible, on the order of 10^-5 absorbance units, which is well below the noise floor of consumer-grade photodetectors.

This physical reality means that standard dual-wavelength PPG sensors (used for SpO2 measurement) cannot directly measure glucose through absorption spectroscopy. Any glucose-related information in conventional PPG signals must therefore come from indirect physiological effects.

Indirect Hemodynamic Effects

Glucose influences several hemodynamic parameters that do modulate the PPG waveform. Elevated blood glucose increases blood viscosity, which affects pulse wave velocity and waveform morphology (Cho et al., 2008). Hyperglycemia induces endothelial dysfunction and alters vascular reactivity, changing the augmentation index and reflection wave characteristics visible in the PPG pulse contour. Insulin secretion in response to glucose loads causes vasodilation that alters peripheral perfusion and PPG amplitude.

These indirect effects form the basis of most PPG-glucose correlation studies. Monte et al. (2021) demonstrated that the PPG second derivative (acceleration plethysmogram) features, including the b/a ratio and augmentation index, showed statistically significant correlations (r = 0.35-0.52) with blood glucose during oral glucose tolerance tests in 42 non-diabetic subjects. However, these correlations are modest and confounded by numerous other factors that affect vascular tone.

Scattering Effects

Glucose alters the refractive index of the interstitial fluid surrounding cells. As glucose concentration increases, the refractive index mismatch between intracellular and extracellular fluid decreases, reducing light scattering (Tuchin et al., 1997; DOI: 10.1117/12.2285026). This "optical clearing" effect has been demonstrated in vitro and in tissue phantoms, but the magnitude at physiological glucose concentrations is small, typically producing less than 1% change in detected light intensity across the 70-400 mg/dL glucose range.

Machine Learning Approaches

The weakness of direct physical correlations has driven researchers toward data-driven machine learning approaches that attempt to extract glucose information from complex patterns across multiple PPG features.

Feature Engineering Methods

Traditional machine learning pipelines extract handcrafted features from the PPG waveform and use regression models to predict glucose. Commonly extracted features include:

Temporal features: systolic peak amplitude, diastolic peak amplitude, pulse interval, systolic time, diastolic time, pulse width at half maximum
Morphological features: augmentation index, inflection point area ratio, crest time ratio, large artery stiffness index
Frequency features: spectral power in cardiac harmonics, spectral centroid, harmonic amplitude ratios
Derivative features: first and second derivative peak amplitudes and ratios (a, b, c, d, e waves of the acceleration plethysmogram)

Hasan et al. (2022) used 42 PPG waveform features with a gradient boosted regression model trained on data from 35 subjects undergoing oral glucose tolerance tests. They reported a mean absolute error (MAE) of 17.3 mg/dL and MARD of 14.8% on held-out test data, with 85.4% of predictions in Clarke Error Grid Zone A. However, the study used a within-subject data split, meaning training and test data came from the same individuals, which likely inflates accuracy compared to cross-subject generalization.

Elgendi et al. (2023) conducted a systematic review of 45 studies using PPG for glucose estimation and found reported MARDs ranging from 7.8% to 35.6%, with a median of 19.2%. Critically, they identified that studies using leave-one-subject-out cross-validation, which better approximates real-world deployment, had substantially worse performance (median MARD 26.1%) compared to within-subject evaluations (median MARD 15.8%).

Deep Learning Approaches

Deep learning models can automatically learn feature representations from raw PPG signals, potentially capturing patterns that handcrafted features miss.

Convolutional neural networks (CNNs) have been applied to raw PPG waveforms and their spectrograms. Zhang et al. (2023) used a 1D-CNN with residual connections trained on 3-second PPG windows from 250 subjects and reported a MARD of 18.2% in leave-one-subject-out evaluation. Their model architecture used 6 convolutional layers with batch normalization and achieved 81.3% Zone A on the Clarke Error Grid.

Recurrent architectures (LSTM, GRU) have been employed to capture temporal dynamics in PPG glucose prediction. Srinivasan et al. (2022) combined a bidirectional LSTM with attention mechanisms operating on sequences of 30 consecutive PPG pulses, reporting improved performance (MARD 16.4%) compared to single-pulse analysis. The attention mechanism revealed that the model focused primarily on diastolic phase features and inter-beat interval variations.

Transformer-based architectures represent the most recent approach. Li et al. (2024) applied a vision transformer to spectrogram representations of PPG signals, achieving a MARD of 15.9% on a dataset of 180 subjects (DOI: 10.1109/JBHI.2024.3367892). However, the computational requirements of transformer models raise questions about deployment feasibility on resource-constrained wearable hardware.

Multi-Wavelength and Hybrid Approaches

Recognizing the limitations of standard dual-wavelength PPG, several research groups have explored expanded spectral approaches.

Multi-Spectral PPG

By using 5 or more LED wavelengths spanning from visible to near-infrared, multi-spectral PPG systems aim to capture wavelength-dependent glucose absorption features that are invisible at standard wavelengths. Caduff et al. (2009) at Solianis (now Biovotion) developed a multi-sensor patch combining impedance spectroscopy, multi-wavelength optical sensing, and temperature measurement. Their system achieved a MARD of 17.5% over 48 hours of continuous monitoring in 20 diabetic subjects, but the sensor was too bulky for a consumer wearable form factor.

Rockley Photonics developed a silicon photonics-based sensing chip integrating multiple wavelengths in the 900-1700 nm range on a single die. Their approach targeted the glucose absorption features in the short-wave infrared (SWIR) that are inaccessible to conventional LED-based systems. Early publications reported encouraging in vitro results, but in vivo clinical validation data has not been published in peer-reviewed literature.

Raman Spectroscopy Augmentation

Samsung Research reported combining PPG with Raman spectroscopy for glucose estimation (Jeon et al., 2023). Raman spectroscopy provides molecular-specific fingerprints and has well-characterized glucose spectral features near 1125 cm^-1. Their prototype achieved a MARD of 12.6% in a 50-subject study, though the Raman excitation laser and spectrometer components present significant miniaturization and power consumption challenges for wearable integration.

PPG Combined with Bioimpedance

Combining PPG with bioelectrical impedance spectroscopy (BIS) provides complementary information about tissue composition and fluid distribution. Glucose affects interstitial fluid osmolarity and cellular hydration, producing measurable impedance changes. Zanon et al. (2022) demonstrated that a PPG+BIS fusion model reduced MARD from 22.1% (PPG alone) to 16.8% (fused) in a 40-subject study, suggesting that multi-modal approaches may be necessary to achieve clinically useful accuracy.

Critical Methodological Concerns

The PPG-glucose literature is plagued by methodological issues that call many reported results into question. Understanding these concerns is essential for evaluating any claimed breakthrough.

The Meal Correlation Problem

Blood glucose follows a predictable temporal pattern after meals: rising for 30-60 minutes, peaking, then declining over 2-3 hours. Many confounders, including skin temperature, heart rate, blood pressure, and autonomic tone, follow similar postprandial patterns driven by digestion and insulin secretion rather than glucose itself. A machine learning model trained on data collected around meals can achieve apparently good glucose prediction by learning these temporal patterns without actually sensing glucose. Caduff et al. (2018) demonstrated this problem explicitly, showing that a model using only timestamp and heart rate could predict glucose with a MARD of 18.3%, comparable to many PPG-glucose studies.

Subject-Specific Calibration

Many studies calibrate their models on a per-subject basis using the first portion of each subject's data, then evaluate on the remaining data. This inflates accuracy because the model learns the individual's glucose-PPG relationship, which varies substantially across people due to differences in skin pigmentation, subcutaneous fat thickness, vascular anatomy, and metabolic response patterns. For a practical consumer device, either frequent recalibration (defeating the purpose of non-invasive monitoring) or robust cross-subject generalization is required. Cross-subject performance is consistently 40-60% worse than within-subject performance across published studies.

Small Sample Sizes and Limited Demographics

The median sample size in PPG-glucose studies is approximately 30-50 subjects (Elgendi et al., 2023). Most studies are conducted in controlled laboratory settings with homogeneous populations, typically young, healthy adults with limited representation of darker skin tones, which significantly affects PPG signal quality and feature extraction. Generalization to the diverse population of diabetes patients, who tend to be older with comorbidities affecting peripheral perfusion, remains undemonstrated.

Regulatory and Clinical Pathway

The FDA classifies non-invasive glucose monitors as Class III medical devices requiring premarket approval (PMA) through clinical trials. The required accuracy thresholds, derived from the SpO2 estimation and CGM regulatory frameworks, are substantially higher than what current PPG-based systems achieve.

For a self-monitoring blood glucose (SMBG) replacement, the FDA expects at least 95% of paired readings within +/-15% of reference values for glucose above 75 mg/dL and within +/-15 mg/dL for glucose below 75 mg/dL. For CGM-class devices, MARD below 10% is the current standard set by devices like the Dexterity G7 (MARD 8.2%) and Abbott FreeStyle Libre 3 (MARD 7.9%).

No PPG-based approach has demonstrated MARD below 15% in a properly designed, sufficiently powered clinical study with independent validation. The gap between current performance and regulatory requirements remains significant.

The Path Forward

Several developments may incrementally improve PPG-based glucose estimation toward clinical utility.

Advanced Photonic Integration

Continued advances in silicon photonics may enable integration of SWIR wavelengths (1000-1700 nm) into wearable-compatible form factors. These wavelengths offer 10-100x stronger glucose absorption than visible/NIR bands. The key engineering challenges are miniaturizing the optical components while maintaining sufficient signal-to-noise ratio and managing the power consumption of longer-wavelength emitters.

Personalized Foundation Models

Large-scale pre-training on diverse PPG datasets followed by few-shot personalization could improve cross-subject generalization. Transfer learning approaches that adapt a general glucose estimation model to an individual using only a few calibration points show promise in preliminary studies (Park et al., 2024), achieving 15-20% relative MARD improvement compared to pure cross-subject models.

Trend Detection Rather Than Absolute Values

Rather than targeting absolute glucose values, some researchers argue that detecting glucose trends (rising, falling, stable) and threshold crossings (hypoglycemia alerts) may be achievable with current technology and clinically useful even without precise numerical accuracy. For trend detection, the relevant metric is sensitivity and specificity for directional changes rather than numerical MARD. Preliminary data suggests 75-85% accuracy for trend direction detection, which, if validated, could provide value as a screening tool complementing rather than replacing conventional monitoring.

Multi-Modal Sensor Fusion

The convergence of PPG with additional sensing modalities, including bioimpedance, skin conductance, temperature, and accelerometry, within modern wearable platforms provides richer input data. Multi-modal approaches consistently outperform PPG-only methods by 20-35% in MARD reduction. As wearable sensor integration advances, the combination of modalities may eventually close the accuracy gap, though this remains speculative.

Conclusion

Non-invasive glucose monitoring via PPG is a problem where the clinical need is immense but the physics are unfavorable. The direct optical signature of glucose at standard PPG wavelengths is too weak for reliable measurement, forcing reliance on indirect correlations that are confounded by numerous physiological variables. Machine learning approaches show promise in controlled settings but have not demonstrated the cross-subject generalization and accuracy required for clinical deployment.

The honest assessment of the field is that PPG alone is unlikely to achieve diagnostic-grade glucose monitoring. Multi-wavelength optical approaches extending into the SWIR range, combined with complementary sensing modalities, represent the most plausible path toward a wearable non-invasive glucose monitor. Researchers and engineers should approach claims of PPG-glucose accuracy with healthy skepticism, paying particular attention to study design, validation methodology, and independent replication.

For related work on extracting clinical biomarkers from PPG signals, see our guides on PPG signal processing algorithms and cardiovascular health assessment.

ChatPPG

Non-Invasive Glucose Monitoring via PPG: Current State of the Art (2026)