Non-Invasive Jaundice Screening Using AI: Machine Learning Analysis of Sclera and Urine Images

Background: Jaundice commonly indicates liver dysfunction and is traditionally diagnosed via invasive blood tests. Objective: This study aimed to develop an AI-based program utilizing images of sclera and urine for non-invasive jaundice screening and compared its accuracy to that of standard blood tests. Methods: This retrospective study involved patients who underwent liver function and bilirubin tests. Scleral and urine images were collected and processed following a standardized protocol to ensure consistency; A4 paper was utilized for white balance correction. Various machine learning and deep learning algorithms, including Decision Tree, Random Forest, XGBoost, DeepSets, and ResNet, were applied to predict jaundice. A stratified five-fold cross-validation was employed, and jaundice was classified as present or absent based on the total bilirubin levels. Results: In total, 57 patients with liver disease and 31 controls were included in the analysis. Various machine learning models were applied to analyze scleral and urine images. The DeepSets model exhibited the highest predictive performance, achieving an R² of 0.782 in predicting bilirubin levels. For jaundice detection, the DeepSets model, using a threshold of 2.6 mg/dL, achieved an accuracy of 87.1% (AUC = 0.869), with a precision of 90.2% and a recall of 88.1%. In contrast, the Random Forest model, with a threshold of 2.9 mg/dL, achieved an accuracy of 84.2% (AUC = 0.833), with a precision of 86.0%, and a recall of 88.1%. Conclusions: This study demonstrates that scleral images, when used with simple A4 white paper for color correction, have the potential to screen for jaundice and predict bilirubin levels. (Clinical Research Information Service of Republic of Korea, KCT0009915).