Multidimensional feature fusion in longitudinal physical examination data: a machine learning framework for classification of type 2 diabetes.

Background: Machine learning-based risk classification of type 2 diabetes mellitus (T2DM) has become a prevailing research direction in intelligent healthcare. Nevertheless, current studies suffer from two key limitations. First, most existing models are built on cross-sectional data and cannot capture the dynamic progression characteristics of T2DM. Second, traditional longitudinal time-series models are commonly adopted without sufficiently exploring the deep interactive relationships among multi-temporal physical examination features, limiting model predictive performance. Methods: To tackle the above drawbacks, this study proposes a novel multi-dimensional feature extraction and fusion network (MFFNet) for T2DM early risk prediction based on sequential annual physical examination data. The proposed architecture comprehensively mines latent feature information from three complementary dimensions: the horizontal interaction of multiple examination indicators within a single period, the longitudinal dynamic evolution pattern of individual indicators across consecutive years, and the cross-temporal synergistic correlations among different indicators across multiple periods. Experiments were conducted on a real-world longitudinal dataset consisting of three consecutive years of physical examination records from the Health Management Center of Beijing Hospital, which presents an extremely imbalanced sample distribution with a positive-negative ratio of 1:8.34. Results: Comparative experiments with state-of-the-art time-series models (RNN, LSTM, GRU, and Transformer) demonstrate the superiority of MFFNet. The proposed model achieved a sensitivity of 0.8766, a specificity of 0.7082, a negative predictive value (NPV) of 0.9799, and a PR-AUC of 0.4322. Furthermore, SHAP interpretability analysis identified core predictive features for T2DM risk assessment, including age, creatinine (Cr), waist circumference (WC), alanine aminotransferase (ALT), triglycerides (TG), and systolic blood pressure (SBP). Conclusion: The proposed MFFNet can effectively implement accurate T2DM classification and early risk prediction using longitudinal physical examination sequences. This multi-dimensional feature fusion strategy substantially improves the mining capability of temporal healthcare data. The lightweight and cost-effective MFFNet provides a reliable artificial intelligence-assisted solution for large-scale early pre-screening and risk intervention of T2DM in primary clinical healthcare.