TITLE:
Heterogeneous Graph Neural Network Modelling of Mood Episode Recurrence in Bipolar Disorder: A Proof-of-Concept Simulation Study of Temporal, Pharmacological, and Social-Rhythm Relations
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Graph Neural Network, Heterogeneous Graph, Bipolar Disorder, Mood Episode Recurrence, Relational Graph Convolutional Network, Social Rhythm, Pharmacological Transition, Relation-Type Ablation, R-GCN, Relapse Prediction
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.8,
August
27,
2026
ABSTRACT: Mood episode recurrence in bipolar disorder is shaped by the structure of a patient’s episode history: the temporal succession of episodes, pharmacological transitions between episodes, and social-rhythm disruptions that may destabilise the circadian and social-zeitgeber system. Existing recurrence models generally flatten this history or merge all connections into one homogeneous graph, thereby discarding relation-type information. We therefore conducted a proof-of-concept simulation study in which each patient history was represented as a heterogeneous graph with episode nodes and three typed edge sets: temporal succession, pharmacological transition, and social-rhythm disruption. The prediction target was a simulated new syndromal depressive, manic, hypomanic, or mixed episode occurring within 12 months after the final observed episode; the binary graph-level label was assigned at the end of the observed episode sequence, using only information available up to that index episode. In N = 700 simulated patients (recurrence prevalence 31.4%), we compared a relation-specific R-GCN with homogeneous GCN, GraphSAGE, GAT, graph-transformer, and no-graph MLP baselines. The heterogeneous R-GCN achieved AUC = 0.822 (95% CI: 0.723 - 0.902), F1 = 0.644, sensitivity = 0.664, and specificity = 0.820 on a held-out simulated test set. Relation-type ablation produced the largest performance decrease after removal of pharmacological-transition edges, followed by social-rhythm-disruption and temporal-succession edges. Because recurrence labels were generated partly from the same pharmacological and social-rhythm mechanisms encoded in the graph, these results demonstrate recovery of simulation-imposed relational structure rather than clinical validity. They support the methodological hypothesis that preserving relation types can improve recurrence prediction when such relations genuinely carry signal, but require confirmation in real longitudinal bipolar-disorder cohorts. In this simulated proof-of-concept setting, relation-specific graph modelling outperformed matched homogeneous and non-graph comparators. The findings should be interpreted as evidence about model behaviour under an explicit generative mechanism, not as a clinically validated recurrence-prediction tool or as proof that any relation type is clinically dominant.