Graph Neural Network-Based Modeling of Disease Progression Using Longitudinal Clinical Data

Abstract

Simulation of the clinical dynamics of disease progression based on longitudinal clinical data is critical to support early diagnosis, stratification of risks, and individual planning of therapeutic approaches in contemporary healthcare systems. But, more traditional machine learning and sequence-based models typically have difficulty jointly modeling intricate temporal and inter-feature relationships found in electronic health records. In an effort to manage this weakness, this paper introduces a Graph Neural Network (GNN)-based disease progression modeling of longitudinal clinical data. Patient records in the proposed approach are modeled as dynamic graphs, with the nodes being the clinical variables, and the edges containing physiological correlations and time-dependent interactions among the clinical variables on the same time steps. The model combines graph convolutional and temporal aggregation to acquire expressive disease evolution representations. The MIMIC-III clinical data, which consists of a wide range of patient trajectories, and a diverse set of features are experimentally evaluated. The suggested approach outperforms the baseline models, such as Logistic Regression, Random Forest, and Long Short-Term Memory (LSTM) networks with a difference of up to 57% in the prediction accuracy and ROC-AUC. Moreover, the framework offers a better understanding of results, revealing the clinically relevant interactions of features. The findings indicate that the GNN-based methods are effective at representing both structural and temporal aspects of healthcare data, which can become an effective solution to advanced clinical decision support systems.