Smartphone screening for neonatal jaundice via ambient-subtracted sclera chromaticity

Abstract

Jaundice is a major cause of mortality and morbidity in the newborn. Globally, early identification and home monitoring are significant challenges in reducing the incidence of jaundice-related neurological damage. Smartphone cameras are promising as colour-based screening tools as they are low-cost, objective and ubiquitous. We propose a novel smartphone method to screen for neonatal jaundice by imaging the sclera. It does not rely on colour calibration cards or accessories, which may facilitate its adoption at scale and in less economically developed regions. Our approach is to explicitly address three confounding factors in relating colour to jaundice: (1) skin pigmentation, (2) ambient light, and (3) camera spectral response. (1) The variation in skin pigmentation is avoided by imaging the sclera. (2) With the smartphone screen acting as an illuminating flash, a flash/ no-flash image pair is captured using the front-facing camera. The contribution of ambient light is subtracted. (3) In principle, this permits a device- and ambient-independent measure of sclera chromaticity following a one-time calibration. We introduce the concept of Scleral-Conjunctival Bilirubin (SCB), in analogy with Transcutaneous Bilirubin (TcB). The scleral chromaticity is mapped to an SCB value. A pilot study was conducted in the UCL Hospital Neonatal Care Unit (n = 37). Neonates were imaged using a specially developed app concurrently with having a blood test for total serum bilirubin (TSB). The better of two models for SCB based on ambient-subtracted sclera chromaticity achieved r = 0.75 (p<0.01) correlation with TSB. Ambient subtraction improved chromaticity estimates in proof-of-principle laboratory tests and screening performance within our study sample. Using an SCB decision threshold of 190μmol/L, the sensitivity was 100% (specificity 61%) in identifying newborns with TSB>250μmol/L (area under receiver operating characteristic curve, AUROC, 0.86), and 92% (specificity 67%) in identifying newborns with TSB>205μmol/L (AUROC 0.85). These results are comparable to modern transcutaneous bilirubinometers.

Fig 1. neoSCB app screen illumination sequence and typical flash/ no-flash sclera image pair.

(A) Screen illumination on for flash image capture. (B) Example of flash image captured by app. (C) Screen illumination off for no-flash image capture. (D) Corresponding no-flash image captured by app. Media consent was provided by the parents.

Fig 2. Processing pipeline to estimate scleral-conjunctival bilirubin from raw format flash/ no-flash image pairs.

In Step 1, a color descriptor is estimated (ambient-subtracted sclera chromaticity). In Step 2, this is mapped to a scleral-conjunctival bilirubin (SCB) value.

Fig 3. Visualising ambient subtraction in CIE 1931 chromaticity diagram for Jaundice Eye Colour Index (JECI) colours.

JECI values are 0.01, 0.03, 0.05, 0.07, 0.09. After ambient subtraction, values come into alignment with ground truth values. (A) Under warm ambient illumination, correlated colour temperature is 2700K. (B) Under cool ambient illumination, correlated colour temperature is 6500K.

Fig 4. Scatter plot of scleral-conjunctival bilirubin (SCB) against TSB with and without ambient subtraction, 37 subjects.

The use of ambient subtraction greatly improves the correlation. (A) Correlation without ambient subtraction, r = 0.56. (B) Correlation with ambient subtraction, r = 0.75.

Fig 5. Bland-Altman plot for TSB and SCB_xy derived from ambient-subtracted scleral chromaticity.

Fig 6. Receiver operating characteristic plots screening TSB at two thresholds using SCB_xy as the decision metric.

(A) Screening threshold of 250μmol/L. (B) Screening threshold of 205μmol/L.