Multimodal learning for neonatal sepsis detection in the NICU using brain-inspired graph neural networks

• The research presents an intelligent methodology which combines Brain-inspired Graph Neural Networks (Bi-GNNs) to merge both electroencephalogram (EEG) signals and structured electronic clinical records (ECR) data for identifying neonatal sepsis. • The proposed model duplicates brain neural pathways though its implementation involves structure-based graph development and temporal pattern and physiological marker relation mapping. • The approach utilizes vital features from heart rate variability blood markers and temperature alongside neural activity measurements obtained from EEG signals that display sensitivity to septic conditions. • The architectural model implements multimodal interactions through graphs which keep track of both temporal relationships and geographical along with cross-modal associations. Attention techniques significantly improve the GNN framework, making it possible for the model to consider the critical sepsis-indicative characteristics. • The system surpasses conventional deep-learning models by achieving state-of-the-art performance in terms of sensitivity, specificity, and overall accuracy, as demonstrated through exhaustive trials on real-world new-born datasets. • Additionally, interpretable graph-based embedding can help gain better insights into the decision-making process and build confidence in AI-based diagnostics. • Multimodal GNNs can be used in critical health applications with a reliable and comprehensible solution for the early and accurate detection of neonatal sepsis in the Neonatal Intensive care unit (NICU). Neonatal sepsis is a life-threatening condition that requires early and reliable detection to reduce mortality and improve clinical outcomes in Neonatal Intensive Care Units (NICUs). Conventional diagnostic approaches based on biomarkers and clinical observations often fail to capture complex temporal and multimodal interactions present in heterogeneous clinical data. This study proposes a Brain-Inspired Graph Neural Network (Bi-GNN) framework that integrates EEG waveform signals and structured electronic clinical records (ECR) derived from the MIMIC-III waveform database for neonatal sepsis prediction. EEG signals are encoded into spike sequences and mapped onto anatomically grounded graph structures, while clinical features are modeled as complementary graph nodes, enabling effective multimodal fusion. A Forward-Forward Graph Convolutional Network with biologically inspired synaptic plasticity is employed to learn temporal and spatial dependencies without backpropagation through time. The proposed framework achieves an accuracy of 94.54%, precision of 94.58%, recall of 93.78%, and an F1-score of 94.00%, demonstrating robust and reliable performance. Comparative evaluation against baseline and state-of-the-art models confirms the effectiveness of the proposed approach. Furthermore, graph-based representations and attention mechanisms enhance interpretability by highlighting clinically relevant neural and physiological patterns, supporting trustworthy and explainable neonatal sepsis assessment in critical care environments.