An AI Approach: Photo-Based Jaundice Detection and Prediction in Newborns

Overview

An AI Approach: Photo-Based Jaundice Detection and Prediction in Newborns Ronika Pillarisetty J-15-17 I did not use AI, when writing this abstract, these are my own words and data analysis figures. Neonatal jaundice is caused by high levels of bilirubin in the blood, affecting 60-80% of newborns worldwide. If jaundice is left undetected, all severe cases can result in permanent neurological damage or death. Current diagnostic methods including TSB (total serum bilirubin) and TcB (transcutaneous bilirubin measurement), are either invasive or expensive for rural and low-resource areas to find, leaving millions of newborns unscreened. This project proposes a iOS-based AI model for non-invasive neonatal jaundice detection using a dual-head deep learning framework, and uses the MobileNetV3-Large CNN, pretrained on ImageNet (a massive dataset consisting of 1,000 categories). Training took place in 2 phases - warming up the conv head by keeping the backbone fully frozen, and fine-tuning only the last MobileNetV3 block and conv head. MobileNetV3-Large serves as a shared backbone between the 2 models. The classification model detects the presence of jaundice from a RAW-image. The regression model estimates the bilirubin in mg/dL, by combining calibrated skin image features with 5 clinical outputs (age, birth weight, gestational age, weight-age ratio, and gender). A color calibration pipeline was developed to ensure a device-independent skin color measurement across different iOS smartphones. The classification model achieved a 86.5% accuracy on 52 test patients, while the regression model achieved a MAE (mean absolute error) of 2.37 mg/dL on 75 test patients. This model runs with no internet connection required and is deployable on iOS devices, making early jaundice screening easily accessible to everyone.