Neonatal jaundice, characterized by elevated bilirubin levels, poses serious health risks for newborns. Early detection is key to preventing complications such as kernicterus, which can lead to permanent brain damage. A promising solution involves technology: researchers from Thailand have developed a texture-based machine learning approach for non-invasive jaundice screening, which could revolutionize infant care, especially in resource-constrained environments.
This study’s approach utilizes the Neonatal Jaundice Screening and Assessment Plate (NJSNAP), which collects clinical data and skin images from newborns. The research focused on 200 infants, whose images were taken from four body locations—forehead, sternum, abdomen, and lower leg. The datasets were divided, with 160 used for training machine learning models and 40 reserved for blind testing.
The key innovation lies in the integration of texture-based analysis with machine learning algorithms. A total of 92 features were extracted from the images, including three clinical details and 89 texture-based characteristics after image processing. By employing various machine learning models, the Support Vector Machine (SVM) emerged as the best performer, particularly effective when paired with feature selection methods like RRelief-F and Univariate Regression.
The findings exhibited strong predictive capabilities for bilirubin levels, showing minimal mean absolute error (MAE) of 1.675 mg/dL and root mean square error (RMSE) of 2.192 mg/dL during blind testing. Notably, there was moderate correlation (r = 0.644, p < 0.001) between predicted and actual bilirubin levels, indicating the reliability of the machine learning model.
The NJSNAP device plays a pivotal role by capturing accurate skin color data through gentle pressure application, improving color fidelity compared to traditional assessment methods. This not only enhances accuracy but aligns with the principles of physical examination traditionally used by healthcare professionals.
Building on previous smartphone-based jaundice detection methods, which often rely solely on camera imaging, the NJSNAP incorporates standardized reference colors to diminish discrepancies from varying lighting conditions and camera-specific processing. This could facilitate broader applicability across diverse populations.
The need for effective jaundice screening tools is urgent, particularly as existing traditional methods can be invasive and uncomfortable for infants. By establishing a non-invasive and easily accessible option, this study aims to offer healthcare providers, especially those operating with limited resources, the necessary tools for timely neonatal jaundice assessment.
Ethnic and demographic variations can impact physiological responses, including bilirubin metabolism, thereby affecting detection strategies. Previous research highlights discrepancies among ethnic groups, emphasizing the need for targeted validation of tools like AmberSNAP across diverse populations. The current study focuses on Southeast Asian infants, integrating their unique skin tones and bilirubin profiles.
While this study marks significant progress, the authors acknowledge several limitations including the small sample size during blind testing. Further research with larger cohorts is necessary to validate this model's applicability across different demographic backgrounds, ensuring robustness and reliability.
Overall, this feasibility study indicates the potential of leveraging machine learning for early neonatal jaundice screening. By utilizing the AmberSNAP system and NJSNAP device, healthcare practitioners can significantly improve early detection capabilities, paving the way for advancements in neonatal care and reducing the prevalence of severe complications associated with untreated jaundice.