Development and Validation of a Smartphone Application for Neonatal Jaundice Screening

Area of Science:

• Medical Informatics
• Neonatology
• Machine Learning in Healthcare

Background:

• Neonatal jaundice (NNJ) affects a significant proportion of newborns, necessitating effective screening methods.
• Current screening methods can be invasive or subjective, highlighting the need for objective, non-invasive tools.

Purpose of the Study:

• To develop and validate a smartphone-based machine learning (ML) application for predicting serum bilirubin (SpB) levels in neonates.
• To assess the app's performance in screening for hyperbilirubinemia using skin color analysis.

Main Methods:

• A smartphone app integrated ML algorithms trained on skin images from neonates, guided by the Kramer principle.
• The app utilized a standardized color calibration card and analyzed multiple regions of interest.
• Model performance was evaluated using metrics like Pearson correlation, sensitivity, specificity, and agreement with total serum bilirubin (TSB).

Main Results:

• The ML app demonstrated strong correlation (Pearson r=0.84) and agreement with TSB levels.
• Achieved 100% sensitivity and 70% specificity for detecting TSB ≥ 17 mg/dL with SpB ≥ 13 mg/dL.
• 82% of predictions were within clinically acceptable limits (±3 mg/dL) of TSB.

Conclusions:

• The developed smartphone-based ML app shows significant potential as a non-invasive screening tool for neonatal jaundice.
• The tool exhibits high sensitivity for detecting high bilirubin levels, complementing TSB measurements.
• Further research is recommended to confirm generalizability and cost-effectiveness in diverse clinical settings.