Serum Bilirubin Prediction for Neonates Using Segmentation-Guided Neural Networks

Overview

Across the world, millions of newborns suffer from severe neonatal jaundice, a condition that can cause neurological damage and death. Approaches such as laboratory blood tests and transcutaneous bilirubinometers for assessing jaundice are financially inaccessible in developing countries, and current computer vision approaches suffer from ease-of-use issues. Furthermore, spectroscopy-based devices produce inaccurate total serum bilirubin (TSB) estimates for neonates with darker skin tones. This research develops novel multi-stage deep learning models to predict bilirubin levels from smartphone imagery of blood plasma test strips. The first task involved training a segmentation model to segment the region of extracted bilirubin from the test strips. In the second task, color and environmental features from the segments are calculated as inputs into a deep regression neural network to predict TSB. The machine learning models are integrated into an end-to-end mobile application for real-world clinical use. The results indicate the segmentation model can adapt to rotational, scale, and ambient lighting irregularities in blood-based bilirubin extraction test strips and successfully segment regions with high bilirubin concentration. The second neural network predicts TSB levels that strongly correlate with state-of-art laboratory measurements (cross-validated Pearson R = 0.83). The mobile application provides bilirubin predictions with an error of 2.38 mg/dL in under five seconds. This tool offers significant ease-of-use advancements with clinical relevance due to the use of inherently skin-tone agnostic blood plasma test strips. The inexpensive (<$1) system can enable widespread proliferation of neonatal jaundice screening in low-middle income countries and reduce fatalities.