Abstract
Predicting the emergence of multiple chronic conditions (MCC) is crucial forearly intervention and personalized healthcare, as MCC significantly impactspatient outcomes and healthcare costs. Graph neural networks (GNNs) areeffective methods for modeling complex graph data, such as those found in MCC.However, a significant challenge with GNNs is their reliance on an existinggraph structure, which is not readily available for MCC. To address thischallenge, we propose a novel generative framework for GNNs that constructs arepresentative underlying graph structure by utilizing the distribution of thedata to enhance predictive analytics for MCC. Our framework employs a graphvariational autoencoder (GVAE) to capture the complex relationships in patientdata. This allows for a comprehensive understanding of individual healthtrajectories and facilitates the creation of diverse patient stochasticsimilarity graphs while preserving the original feature set. These variationsof patient stochastic similarity graphs, generated from the GVAE decoder, arethen processed by a GNN using a novel Laplacian regularization technique torefine the graph structure over time and improves the prediction accuracy ofMCC. A contextual Bandit is designed to evaluate the stochastically generatedgraphs and identify the best-performing graph for the GNN model iterativelyuntil model convergence. We validate the performance of the proposed contextualBandit algorithm against $\varepsilon$-Greedy and multi-armed Bandit algorithmson a large cohort (n = 1,592) of patients with MCC. These advancementshighlight the potential of the proposed approach to transform predictivehealthcare analytics, enabling a more personalized and proactive approach toMCC management.