Development and validation of neonatal jaundice detection using smartphone application

Neonatal jaundice is highly prevalent in newborns globally, requiring vigilant monitoring to prevent severe complications such as brain damage. Current detection methods, including invasive blood tests and expensive bilirubinometers, have limitations. While smartphone-based tools show promise, their accuracy and robustness are often suboptimal.

From 2016 to 2022, healthy newborns who were less than 28 days old and had a gestational age of at least 35 weeks were enrolled. Facial and chest images were captured using a smartphone with a custom-made color calibration card. These images underwent color normalization, skin region extraction and color feature engineering. Multiple regression models were trained on these features to predict total serum bilirubin (TsB) levels. Model performance was evaluated by TsB correlation, root mean square error (RMSE), and the ability to identify newborns with significant hyperbilirubinemia.

A total of 341 newborns were enrolled, including 166 male infants (48.7 %). In total, 168 infants (49.3 %) were categorized as having a high or high-intermediate risk of jaundice. The multilayer perceptron model demonstrated the highest predictive accuracy of TsB (R = 0.90, RMSE = 2.44 mg/dL), outperforming six other machine learning algorithms. It accurately identified 89.4 % of newborns with TsB>15 mg/dL, with an area under the curve of 0.963, a sensitivity of 90.1 % and a specificity of 89.0 %.

Smartphone imaging with machine learning algorithms can serve as an accessible, noninvasive and precise method for quantifying jaundice severity in newborns. Our proposed approach addresses limitations in existing tools and provides a pragmatic solution for convenient jaundice screening at home by parents. With further validation, this technique could aid early intervention and prevent complications from severe neonatal jaundice globally.