Smartphone-based app highly sensitive for detecting neonatal jaundice

A smartphone-based application that integrates machine learning technology holds promise as a screening tool for neonatal jaundice, demonstrating good correlation and agreement with total serum bilirubin (TSB).

Developed by a team of researchers from Singapore, the app captures skin images from specific areas of the body such as the forehead, sternum, and abdomen, following the cephalocaudal progression of jaundice. A standardized colour calibration sticker card with a central hole is affixed to the skin to ensure accurate colour readings.

The app integrates a machine learning model that underwent iterative development and k-folds cross-validation. Yellowness-related features from the forehead, sternum, and abdomen based on the Kramer principle were included as predictors.

During the validation, the app achieved 100-percent sensitivity and 70-percent specificity for detecting neonatal jaundice based on a decision rule of using a threshold of ≥13 mg/dL to predict TSB levels of ≥17 mg/dL. The positive predictive value was 10 percent, negative predictive value was 100 percent, positive likelihood ratio was 3.3, and negative likelihood ratio was 0. [JAMA Netw Open 2024;7:e2450260]

The area under the receiver operating characteristic curve was 0.89 (95 percent confidence interval, 0.82–0.96), suggesting diagnostic accuracy, the researchers noted. “Its diagnostic accuracy was comparable with transcutaneous bilirubinometry,” and there was good correlation and statistical agreement with TSB.

The smartphone-based neonatal jaundice screening app was validated in a multiethnic cohort of 546 neonates (median gestational age 38.0 weeks, 52.4 percent male, 57.7 percent Chinese, 6.4 percent Indian, 31.0 percent Malay, 4.9 percent other ethnicities). Iterative development and cross-validation of the machine learning model was performed on 352 neonates.

“The app is an innovative neonatal jaundice screening tool that has demonstrated strong sensitivity and moderate specificity across a multiethnic population, improving accuracy by using multiple skin regions and colour calibration stickers,” the researchers said.

One of the advantages of the app is the potential to allow remote screening by means of home-based measurements and teleconsultations with healthcare professionals, they added. “This approach could reduce the need for frequent clinic visits, especially during epidemics, and has been positively received by parents for its convenience.” [JMIR MHealth UHealth 2023;11:e53291; J Med Internet Res 2022;24:e37843]

The researchers acknowledged that despite the app achieving 100-percent sensitivity, the specificity of only 70 percent may lead to potential false positives and unnecessary blood draws for TSB testing. This could, in turn, impose additional costs and heighten parental emotional stress.

“Nevertheless, [the app’s] specificity was still higher than that of transcutaneous bilirubinometry, which had a specificity of 51 percent. As the app is intended to be used as a screening tool, misclassifying negative cases as neonatal jaundice are less critical than missing actual jaundice cases,” they added.

In the future, the researchers plan to validate the app’s effectiveness across different smartphone models, different camera specifications, varying lighting conditions, as well as exploring various bilirubin thresholds for guiding the need for TSB assay. In the works are a pilot study to assess the app’s safety and acceptability and a parallel health economics study to evaluate the app’s cost-effectiveness and environmental impact.

“Ultimately, the goal is to integrate the screening tool with teleconsultation services in primary care clinics in Singapore, establishing a decentralized care model for neonatal jaundice screening,” they said.