Development and validation of neonatal jaundice detection using smartphone application.

Area of Science:

• Medical Imaging
• Machine Learning
• Neonatal Health

Background:

• Neonatal jaundice is a common condition requiring monitoring to prevent brain damage.
• Current detection methods like blood tests and bilirubinometers have limitations.
• Existing smartphone tools for jaundice detection are often inaccurate.

Purpose of the Study:

• To develop and evaluate a smartphone-based imaging method for noninvasively quantifying neonatal jaundice.
• To assess the accuracy of machine learning models in predicting total serum bilirubin (TsB) levels from smartphone images.

Main Methods:

• Healthy newborns (gestational age ≥35 weeks) had facial and chest images captured using a smartphone and a color calibration card.
• Images underwent color normalization, skin region extraction, and feature engineering.
• Multiple regression models, including a multilayer perceptron, were trained to predict TsB levels.

Main Results:

• A multilayer perceptron model achieved high predictive accuracy for TsB (R=0.90, RMSE=2.44 mg/dL).
• The model accurately identified 89.4% of newborns with TsB > 15 mg/dL (AUC=0.963, sensitivity=90.1%, specificity=89.0%).
• This approach outperformed six other machine learning algorithms.

Conclusions:

• Smartphone imaging combined with machine learning offers a precise, accessible, and noninvasive method for assessing neonatal jaundice severity.
• This technique provides a practical solution for home-based jaundice screening by parents.
• Further validation could enable early intervention and prevent severe neonatal jaundice complications globally.