[2Win5] Poster session 2

[2Win5-74] Integration of Graph Embedding Propagation and LSTM: Patient Outcome Prediction from Multimodal Medical Data

〇Yoshitaka Nozaki¹, Kenichiro Akagi¹, Takashi Ohnishi¹, Giampaolo Pileggi², Federico Errica², Marco Lepri² (1.NEC Corporation, 2.NEC Laboratories Europe)

Keywords:Graph Neural Network, Embedding Propagation, Patient Outcome Prediction, Multimodal AI, Medical Time-series Data

In healthcare, graph technologies are expected to model patient relationships and improve prediction accuracy through data integration. However, traditional graph-based methods struggle with time-series data processing, limiting their performance in complex tasks like patient outcome prediction. We propose a novel approach integrating Embedding Propagation (EP) and Long Short-Term Memory (LSTM) Encoders. Our method constructs a patient-node graph structure, using EP to capture inter-patient relationships while processing time-series data with LSTM Encoders. This allows effective representation of both static and dynamic features. We validated our approach using the PhysioNet Challenge Dataset by predicting whether patients' length of stay would exceed a threshold. Compared to conventional methods, our approach achieved superior performance with an AUROC of 78.6%. This research expands the potential of graph-based machine learning in healthcare and holds promise for future clinical applications.