Anemia is one of the most common medical conditions on earth, affecting roughly 1.9 billion people, yet the test that confirms it has barely changed in a century. A needle, a vial, a centrifuge, and a wait. For a pregnant patient in a rural clinic, a dialysis recipient, or a parent watching a pale toddler, that friction is the difference between catching a problem early and missing it entirely. A growing body of research now asks a provocative question: can a camera, pointed at skin, eyelid, or fingernail, estimate hemoglobin accurately enough to matter, without drawing any blood at all?
"Smartphone-based, non-invasive hemoglobin estimation reached a mean absolute error of approximately 0.72 g/dL in real-world testing, tightening to about 0.50 g/dL above 10 g/dL, which approaches the variability seen between repeat laboratory draws." - Dr. Wilbur Lam, Emory University and Georgia Institute of Technology (2025)
Can AI predict hemoglobin with no blood draw and contactless capture?
The premise behind the idea that AI can predict hemoglobin with no blood draw in a contactless manner is grounded in optics, not magic. Hemoglobin is the dominant light-absorbing molecule in blood, and its concentration changes how skin and mucosal tissue reflect and absorb specific wavelengths of light. A camera that captures these subtle color and pulsatile signals is, in effect, sampling the same optical property that a laboratory analyzer measures directly. The machine learning task is to map that noisy, indirect signal back to a number expressed in grams per deciliter.
There are two broad signal chains under active study. The first is color-based: a model analyzes the pallor of the palpebral conjunctiva (the inner eyelid), the fingernail bed, the palm, or the retina, where blood-rich tissue is visible close to the surface. The second is pulsatile: remote photoplethysmography (rPPG) and multi-wavelength PPG extract the tiny periodic changes in reflected light caused by each heartbeat, then infer hemoglobin from the shape and absorption characteristics of that waveform.
Neither approach measures hemoglobin. Both infer it. That distinction defines every strength and every limitation discussed below.
How contactless hemoglobin estimation compares to a blood draw
The table below summarizes how the leading non-invasive approaches stack up against the conventional venous blood draw, drawing on reported accuracy figures from 2024 to 2025 literature.
| Method | Signal captured | Reported accuracy | Main limitation |
|---|---|---|---|
| Venous blood draw (lab CBC) | Whole blood sample | Reference standard | Invasive, needs lab, delayed result |
| Conjunctiva imaging + deep learning | Inner eyelid color | RMSE around 0.68 g/dL, MAPE 3.4% | Requires consistent eyelid eversion and lighting |
| Fingernail selfie app | Nail bed color | MAE around 0.72 g/dL | Sensitive to nail polish, lighting, skin tone |
| Retinal image deep learning | Fundus vasculature | MAE around 0.58 g/dL | Needs fundus camera, clinical setting |
| Multi-wavelength PPG + ML | Pulsatile blood volume | Mean relative error under 2.46% | Needs specialized multi-wavelength sensor |
| Wearable single-channel PPG | Pulsatile blood volume | RMSE 0.90 to 1.20 g/dL | Motion artifact, calibration drift |
A few patterns stand out. Color-based methods that look at blood-rich tissue tend to perform best when image capture is standardized. Pulsatile methods can run continuously and passively but are more vulnerable to movement and require careful wavelength selection. None yet matches the precision of a laboratory complete blood count, but several land within the range that clinicians consider useful for triage and screening.
Clinical Applications
Maternal and prenatal screening
Anemia in pregnancy raises the risk of preterm birth, low birth weight, and maternal mortality. In many low and middle income settings, a hemoglobin check is supposed to happen at every antenatal visit but often does not, because the test requires supplies and trained staff. A contactless estimate captured on a clinician's phone, or even a patient's own device, could turn an irregular event into a routine one. The clinical bar here is not laboratory precision. It is reliable detection of who falls below a treatment threshold.
Chronic disease and home monitoring
Patients with chronic kidney disease, heart failure, or cancer frequently develop anemia and need repeated hemoglobin checks. Contactless estimation offers the possibility of trend monitoring between clinic visits, flagging a meaningful drop before symptoms escalate. Research by Dr. Wilbur Lam at Emory University and Georgia Institute of Technology (2025) emphasizes that personalized calibration, where the model learns an individual's baseline from one reference value, substantially improves accuracy for self-monitoring.
Blood donation and transfusion services
Blood-bank operators screen every prospective donor for hemoglobin to protect both the donor and the supply. A fast, reagent-free pre-screen could reduce deferrals caused by single borderline finger-prick readings and improve the donor experience. The same logic applies to transfusion decision support, where a non-invasive trend could complement, not replace, confirmatory testing.
Current research and evidence
The 2024 to 2026 literature has moved the field from proof of concept toward clinical evaluation. Researchers reporting on deep learning applied to palpebral conjunctival images (2024) achieved a root mean squared error around 0.68 g/dL and a mean absolute percentage error of 3.4 percent, with conjunctival pallor outperforming palm and nail-bed pallor for anemia detection. A separate group working on retinal fundus images (2024) reported a mean absolute error near 0.58 g/dL, framing the approach as a scalable option for resource-limited settings, though it depends on access to a fundus camera.
On the smartphone front, the NiADA application underwent clinical validation in 2024 for real-time hemoglobin estimation from lower-eyelid images. Building on years of work behind fingernail-based estimation, Dr. Wilbur Lam at Emory University and Georgia Institute of Technology (2025) and colleagues published clinical and real-world evaluation of a fingernail selfie app reporting a mean absolute error of roughly 0.72 g/dL, improving to about 0.50 g/dL above 10 g/dL when individually personalized. That personalization step, anchoring the model to a single past laboratory value, is one of the most important practical findings in the field.
Pulsatile approaches have advanced in parallel. Work on multi-wavelength PPG signal features combined with a multilayer perceptron (2024) reported a mean relative error under 2.46 percent, while research on an adaptive lightweight convolutional network for wearable PPG reported root mean squared errors of 0.90 and 1.20 g/dL across two datasets. Explainable AI methods are also entering the picture, with multichannel PPG studies attempting to show which signal features drive each prediction rather than treating the model as a black box.
The open machine learning problems are consistent across every method. Generalization across skin tones remains the most serious, since color-based models trained on narrow populations can encode bias. Image acquisition variability, ambient lighting, camera hardware differences, nail polish, and eyelid eversion technique all inject noise. Dataset size is still limiting, and external validation in diverse, prospective cohorts lags behind the impressive numbers reported on internal test sets. Distribution shift, where a model trained in one clinic degrades in another, is the recurring failure mode.
The Future of contactless hemoglobin estimation
The trajectory points toward hybrid systems rather than a single winning modality. Combining a color-based reading from blood-rich tissue with a pulsatile rPPG signal could let each method compensate for the other's weaknesses, and pairing both with a one-time personal calibration value appears to be the most realistic path to clinically actionable accuracy. Expect the near-term role to be screening and trend monitoring, identifying who needs a confirmatory test, rather than replacing the laboratory outright.
Equity will determine whether the technology earns trust. Models validated across the full range of skin tones, capture devices, and clinical conditions, with transparent reporting of where they fail, are the ones that will survive regulatory and clinical scrutiny. The most valuable contribution from industry over the next two years will be large, diverse, well-labeled benchmark datasets that let independent researchers test generalization honestly.
Circadify maintains prepublication hemoglobin benchmark data from its trials, available to researchers and partners on request at [email protected], to support exactly this kind of external validation.
Frequently asked questions
Can AI really measure hemoglobin without any blood at all? AI does not measure hemoglobin directly. It infers the value from optical signals captured by a camera, such as the color of the inner eyelid or fingernail or the pulsatile signal in skin. Studies from 2024 to 2025 report mean absolute errors in the range of roughly 0.5 to 1.2 g/dL, which is useful for screening and trend monitoring but not yet a replacement for a laboratory complete blood count.
How accurate is contactless hemoglobin estimation compared to a lab test? Reported accuracy varies by method. Conjunctiva and retinal imaging models have shown errors around 0.58 to 0.68 g/dL, and a personalized fingernail app reached about 0.50 g/dL above 10 g/dL. These figures come from research test sets, and accuracy can drop in new populations, lighting conditions, or devices, which is why external validation matters.
What are the biggest unsolved problems? Generalization across skin tones, consistent image capture, limited and non-diverse training data, and the need for personal calibration. Motion artifact and wavelength selection are additional challenges for pulsatile PPG methods. Independent validation in large prospective cohorts is still catching up to the published accuracy.
Who benefits most from this technology today? Maternal-health programs, chronic disease patients who need frequent monitoring, and blood-donation services are among the clearest use cases, particularly where laboratory access is limited. The realistic role for now is rapid screening that flags who needs a confirmatory blood test.