Every year, roughly 481,000 late-preterm and term newborns develop severe jaundice. About 114,000 of them die. Most of those deaths happen in low and middle-income countries where the standard diagnostic tools either don't exist or can't be maintained. A transcutaneous bilirubinometer costs several thousand dollars, needs regular calibration, and breaks in environments without reliable supply chains. A blood draw requires trained phlebotomists, lab infrastructure, and time that a deteriorating newborn may not have.
The smartphone in a community health worker's pocket costs a fraction of that. Over the past five years, researchers have been asking whether its camera is good enough to screen for the yellow skin discoloration that signals dangerous bilirubin levels. The answer, increasingly, is yes — with caveats worth understanding.
"Smartphone-based technologies have opened up new and innovative ways of delivering healthcare, which could be especially beneficial in low and middle-income countries where conventional devices are unavailable." — Aune et al., BMJ Paediatrics Open (2025)
How camera-based jaundice detection works
Bilirubin is a yellow pigment. When it accumulates in a newborn's blood, it deposits in the skin and the sclera (the white of the eye), producing visible color changes. This is the same principle behind the Kramer scale that clinicians have used for decades to visually grade jaundice by how far down the body the yellow discoloration extends — from the face (mild) to the soles of the feet (severe).
The problem with visual assessment is that humans are bad at it. Clinicians routinely misclassify jaundice severity, particularly in babies with darker skin pigmentation where yellow discoloration is harder to see. Aune et al. (2025) confirmed this in their multi-country study: the Picterus smartphone system outperformed visual assessment using the Kramer scale across all three study populations.
Camera-based approaches work by quantifying what the eye estimates poorly. A smartphone camera captures an image of the newborn's skin or sclera. Software then analyzes the color values in specific regions of interest, accounting for ambient lighting conditions, and maps those values to estimated bilirubin concentrations. Two broad technical approaches have emerged:
Skin color analysis captures images of the forehead, sternum, or other body regions and extracts color features from specific color channels. Almansour et al. (2021) demonstrated that deep transfer learning models trained on skin images could classify jaundiced versus non-jaundiced newborns with high accuracy. The challenge is that skin pigmentation, ambient lighting, and camera white balance all influence the color values captured.
Scleral analysis photographs the white of the eye, where bilirubin deposition produces yellowing against a relatively uniform background. The neoSCB app uses this approach, and validation studies published in Pediatrics (2022) showed its diagnostic accuracy was comparable to the JM-105 transcutaneous bilirubinometer and other skin-based smartphone apps like BiliCam.
Comparing jaundice screening methods
| Method | Contact required | Cost | Accuracy vs TSB | Skin tone sensitivity | Setting | Time to result |
|---|---|---|---|---|---|---|
| Total serum bilirubin (blood draw) | Yes (invasive) | $15-50 per test | Reference standard | None | Lab required | 30-60 minutes |
| Transcutaneous bilirubinometer (JM-105) | Yes (probe) | $3,000-8,000 device | r = 0.83-0.91 | Moderate (less reliable above Fitzpatrick V) | Clinical | Seconds |
| Visual assessment (Kramer scale) | No | Free | Poor (frequent misclassification) | High (unreliable in dark skin) | Any | Minutes |
| Picterus smartphone app | No (camera) | Smartphone + calibration card | r = 0.76 (multi-population) | Moderate (varies by population) | Any with smartphone | Under 1 minute |
| neoSCB scleral app | No (camera) | Smartphone only | Comparable to JM-105 | Lower (sclera less affected) | Any with smartphone | Under 1 minute |
| BiliCam smartphone app | No (camera) | Smartphone + color card | Within 2.2 mg/dL of TSB | Moderate | Any with smartphone | Under 1 minute |
Sources: Aune et al. (2025), Taylor et al., Pediatrics (2022), Almansour et al. (2021), published device validation studies.
The table tells you something that matters for implementation: no smartphone app matches a blood test. None of them are trying to. The clinical question isn't whether a phone camera can replace lab analysis. It's whether a phone camera can tell you which babies need the blood test. That's screening versus diagnosis, and it's an important distinction.
What the research shows
Aune, Gierman, Bergseng et al. (2025) at the Norwegian University of Science and Technology conducted cross-sectional studies in Mexico, Nepal, and the Philippines between 2018 and 2022. They enrolled 416 infants and compared the Picterus smartphone system against total serum bilirubin. The overall correlation was r=0.76, with Bland-Altman limits of agreement of plus or minus 89.2 micromoles per liter. An interesting finding: the system underestimated bilirubin in Mexico but overestimated it in Nepal and the Philippines, suggesting population-specific calibration may be necessary.
Taylor, Gangadharan, Engwa et al. (2022) published validation results for the neoSCB app in Pediatrics. Their sclera-based approach showed diagnostic accuracy comparable to the JM-105 transcutaneous bilirubinometer. By analyzing the white of the eye rather than skin, they partly sidestepped the skin pigmentation problem that complicates other camera-based methods.
Almansour, Alshehri, and Aljarbou (2021) at King Saud University tested deep transfer learning models (VGG-16, ResNet) trained on smartphone-captured images of neonatal skin and eyes. Their fused-feature approach — combining skin and eye image data — outperformed models trained on either feature alone. The transfer learning models achieved classification performance comparable to traditional machine learning methods like support vector machines.
A 2026 study published in Nature Scientific Reports used a vision transformer architecture (T2T-ViT) on neonatal skin images captured via smartphone, achieving 99% accuracy across multiple metrics. While these numbers are impressive, they come from a single center and need broader validation before clinical deployment.
Clinical applications where this matters most
Community screening in resource-limited settings
The biggest impact of smartphone jaundice screening won't be in hospitals that already have bilirubinometers. It will be in settings where newborns are discharged within hours of birth, where follow-up visits happen at community health posts without lab infrastructure, and where the nearest hospital is hours away. A community health worker with a smartphone and basic training could screen newborns during routine home visits and refer only those flagged for further testing. This triage function could catch cases that currently go undetected until brain damage has already occurred.
Early discharge monitoring
Even in high-income countries, early hospital discharge creates a window of risk. Bilirubin levels typically peak between days 3 and 5 of life, often after the newborn has gone home. Parents using a validated smartphone app could monitor for rising bilirubin levels and seek care before levels become dangerous. This doesn't replace clinical judgment, but it adds a data point where previously there was only parental observation of skin color.
Telemedicine triage
When a parent calls about a yellowish baby, the clinician on the other end has almost nothing to work with. A standardized smartphone image with estimated bilirubin levels would give the telehealth provider actual data. Combined with gestational age, hours of life, and feeding status, this could support better remote triage decisions.
The calibration problem
Camera-based jaundice screening has a physics problem that researchers are still working through. The color of a newborn's skin as captured by a smartphone camera depends on at least four things: the actual bilirubin concentration in the skin, the baby's baseline skin pigmentation, the ambient lighting conditions, and the camera's sensor characteristics and white balance algorithms.
The Picterus system addresses this with a physical calibration card placed next to the baby during image capture. The card provides reference colors that allow the software to compensate for lighting and camera variation. BiliCam uses a similar approach. These cards work, but they add a step and a physical consumable to the workflow.
The neoSCB approach of analyzing the sclera reduces some of these variables since the scleral baseline is more consistent across skin types. But photographing a newborn's eye is harder than photographing their forehead. Babies squirm. Their eyes are small. Getting a clear, well-lit image of the sclera with a phone camera takes practice.
Machine learning approaches, especially the deep learning models tested by Almansour et al. and the vision transformer work, attempt to learn these confounding factors from training data rather than correcting for them explicitly. The risk is that models trained predominantly on one population perform poorly on another — exactly the pattern Aune et al. observed with Picterus across Mexico, Nepal, and the Philippines.
Where this technology is heading
The trajectory is toward integration and validation. Individual apps measuring individual biomarkers will likely merge into platforms that capture multiple physiological signals from camera-based assessment. A smartphone pointed at a newborn could simultaneously estimate bilirubin levels, heart rate, respiratory rate, and perfusion — providing a more complete clinical picture than any single measurement.
Circadify has developed camera-based physiological measurement technology that captures multiple vital signs from video. As smartphone jaundice screening matures and regulatory pathways become clearer, integrating bilirubin estimation into multi-vital-sign camera-based platforms could make newborn assessment more comprehensive without adding hardware.
The regulatory landscape is evolving. The FDA has not yet cleared any smartphone-only jaundice screening app, though several are in various stages of clinical validation. CE marking has been obtained by Picterus in Europe. The path from research tool to clinical device is long, but the clinical need — particularly the 114,000 annual deaths from a treatable condition — makes this one of the more compelling applications of camera-based health assessment.
Frequently asked questions
Can a smartphone camera detect jaundice in newborns?
Yes. Several validated smartphone apps use the phone's camera to estimate bilirubin levels by analyzing skin or scleral color. The Picterus system showed a correlation of r=0.76 with serum bilirubin across populations in Mexico, Nepal, and the Philippines. These tools work best as screening devices to identify which newborns need confirmatory blood tests.
How accurate are smartphone jaundice apps compared to blood tests?
Smartphone apps are designed for screening, not diagnosis. The Picterus system demonstrated limits of agreement of plus or minus 89.2 micromoles per liter compared with total serum bilirubin. The neoSCB sclera-based app showed diagnostic accuracy comparable to commercial transcutaneous bilirubinometers. Both outperform visual assessment by clinicians.
Why is neonatal jaundice screening important in low-income countries?
Severe neonatal jaundice affects roughly 481,000 late-preterm and term newborns each year, causing an estimated 114,000 deaths, most in low and middle-income countries. Transcutaneous bilirubinometers cost thousands of dollars and require calibration, making them impractical in many settings. Smartphone-based screening could dramatically expand access to early detection.
Does skin color affect the accuracy of camera-based jaundice detection?
Skin pigmentation does affect accuracy. The Picterus system showed higher correlation with serum bilirubin in lighter-skinned populations in Mexico compared with darker-skinned populations in Nepal and the Philippines. Sclera-based approaches like neoSCB may be less affected by skin pigmentation since they analyze the white of the eye instead.
Related articles
- Contactless Vitals in Neonatal Intensive Care — How camera-based monitoring supports the most fragile patients without skin contact.
- Contactless Vitals in Pediatric Care Applications — Broader applications of contactless vital sign monitoring for children.
- rPPG Technology and Global Health in Sub-Saharan Africa — Camera-based health assessment in resource-limited settings where smartphone screening has the most impact.