Abstract
BACKGROUND: Multimorbidity has become a major global public health challenge. However, existing research primarily emphasizes the identification of disease patterns at the population level and lacks the capacity to provide predictive insights into individual future pattern membership. Bridging this gap is crucial for personalized prevention and management. OBJECTIVE: This study aims to propose an innovative framework that integrates population-level multimorbidity pattern recognition with individual-level predictive modeling, thus advancing multimorbidity research from descriptive analysis to prospective multimorbidity pattern prediction. METHODS: Using longitudinal health follow-up data, we first applied latent transition analysis (LTA) to identify temporally stable multimorbidity patterns. These patterns were subsequently transformed into predictive labels to construct a novel deep learning model, CLA-Net (Cross-Lag Attention Network). CLA-Net is designed to predict individual future multimorbidity patterns by leveraging the complementary strengths of Gated Recurrent Units (GRU) and transformer architectures. It introduces a bitemporal directed cross-attention mechanism to simultaneously capture temporal dependencies and complex feature interactions. We compared CLA-Net against several advanced baselines and conducted ablation studies to validate its architectural components. RESULTS: In terms of pattern recognition, the LTA identified 5 clinically meaningful multimorbidity patterns: Cardiometabolic-Multisystem, Hypertension-Arthritis, Respiratory-Musculoskeletal, Metabolic Syndrome, and Gastritis-Arthritis. In terms of prediction, experimental results demonstrated that CLA-Net significantly outperformed all baseline models. CLA-Net achieved an accuracy of 0.8352 (SD 0.0048), a precision of 0.8326 (SD 0.0053), a recall of 0.8312 (SD 0.0056), and an F1-score of 0.8319 (SD 0.0051). Notably, it achieved an area under the curve of 0.9293, surpassing baseline models. Ablation studies confirmed the necessity of the dual-branch architecture and the directed cross-attention mechanism, as removing these components resulted in performance declines ranging from 0.93% to 2.50%. CONCLUSIONS: This study extends the scope of LTA beyond descriptive statistical modeling and establishes the scientific value of multimorbidity pattern prediction as an independent research task. By bridging population-level insights with individual-level prediction, the proposed framework provides a data-driven tool for the prospective prediction of future multimorbidity pattern membership conditional on survival, thereby supporting stratified disease management and care planning, rather than general risk stratification for acute or end-stage deterioration. This offers new methodological and practical value for precision medicine and public health policymaking.