From Data to Dialogue: A Communication-Centered Perspective on AI-Based Multimodal Prediction of Neonatal Jaundice

Abstract

Neonatal jaundice is a common issue in the newborn period, and timely, accurate monitoring and prediction  are  crucial  to  prevent  severe  complications  (e.g.  kernicterus).  However,  traditional diagnosis relies on invasive serum bilirubin tests or subjective judgment, which limits applicability in developing countries  and home  settings.  In recent years, machine learning techniques have been introduced for  non-invasive  neonatal jaundice  prediction:  convolutional  neural  networks  (CNNs) have been used for analyzing skin images of jaundice, and long short-term memory (LSTM) networks for time series prediction of bilirubin changes, achieving some success. Yet, CNNs and LSTMs have limitations such as local receptive fields and difficulty capturing long-term dependencies. This paper comprehensively reviews and reconstructs the machine learning framework for neonatal jaundice prediction by introducing the Transformer model as the core. We use a Vision Transformer (ViT) instead of CNN for processing  skin images,  and  a time-series Transformer  in place  of LSTM for modeling dynamic data like transcutaneous bilirubin readings, and we fuse multimodal information to improve prediction accuracy. We highlight the potential of this Transformer-based framework in the health communication domain: for example, using attention heatmap visualization to improve model  interpretability,  and  integrating  with  mobile  health  (mHealth)  applications  for  remote monitoring and interaction, thereby increasing public awareness and trust in new technology. We detail the methodological architecture, demonstrate its promising application in early prediction of neonatal jaundice, discuss current challenges, and envision future research directions. By leveraging a  Transformer-centric  multimodal  deep  learning  framework,  neonatal  jaundice  prediction  could achieve  higher  accuracy  and  interpretability,  better  serving  clinical  decision-making  and  family healthcare.