A multi-scale temporal hypergraph neural topic model for treatment pattern mining

The widespread adoption of Electronic Health Records (EHR) offers opportunities to model real-world treatment patterns, but existing methods often fail to capture both the combination of treatments within a stage and their evolution over time. We propose the Temporal Hypergraph Neural Topic Model (THNTM), a unified framework whose core innovation is a joint generative process that simultaneously models treatment topics, their structural co-occurrence, and temporal evolution, with a formal proof that it generalizes Latent Dirichlet Allocation (LDA). THNTM integrates neural topic modeling with multi-scale temporal hypergraph learning to achieve this goal THNTM first constructs patient-specific hypergraphs at multiple time scales, encoding both concurrent treatment combinations and their temporal order. It then jointly learns interpretable treatment patterns and models patient trajectories across them. On three public datasets comprising over 160,000 patient traces, THNTM consistently outperforms state-of-the-art baselines, achieving relative improvements of up to 19.3% in topic coherence and 3.6% to 5.7% in clinical relevance, while maintaining high topic diversity. Ablation studies confirm the contribution of each model component, with the full model improving NPMI by 53.2% over a standard topic model. THNTM provides a robust method for mining clinically meaningful, structured treatment pathways to support decision-making.