Tracing Health–to–Disease Trajectories from Annual Checkup Data
Satoshi NAGAIE *1, Satoshi MIZUNO1, Soichi OGISHIMA1, 2
1Tohoku Medical Megabank Organization, Tohoku University
2The Advanced Research Center for Innovations in Next-Generation Medicine (INGEM), Tohoku University
[Purpose]
Personalized prevention remains limited by the absence of a common scale for the pre–disease phase, the inadequacy of one–size–fits–all cutoffs and the clinical opacity of unsupervised clusters. To quantify and map the full health–to–disease continuum―from robust health, through Mibyo (pre–disease), to overt disease―by constructing a multimodal, longitudinal health state space from annual checkups and claims, and by modeling individual trajectories to support clinically interpretable, personalized prevention.
[Methods]
We analyzed the JMDC Claims Database, integrating laboratory test results and questionnaire responses, in a cohort of 0.4 million individuals.Missing values were handled using multiple imputation by chained equations (MICE). After preprocessing, 46 features (18 laboratory and 28 questionnaire variables) were modeled. To construct a low–dimensional “health state space,” we reduced the 46 features to 40 principal components with principal component analysis (PCA) and embedded the 40 component scores into a 3–dimensional representation using Uniform Manifold Approximation and Projection (UMAP). In the male sub-cohort, cluster structure was explored with Density–Based Spatial Clustering of Applications with Noise (DBSCAN) and Clustering Large Applications (CLARA). Finally, trajectories were grouped into clinically interpretable categories.
[Results and Discussion]
The PCA–UMAP embedding uncovered discrete structure in men. DBSCAN found 124 clusters: CLARA summarized to 33 groups. The 33 cluster groups collapsed into five profiles: Health–Preservation, Borderline–Risk, Metabolic–Risk, Lifestyle–Issue, and Secondary–Prevention. Trajectories split into stable (58%), single–transition (17%), and multi–transition (25%); These patterns support trajectory–aligned prevention.
[Conclusions]
In a ten year cohort of men undergoing annual health checkups, multimodal data were distilled into 33 clusters and further consolidated into five actionable health profiles―Health Preservation, Borderline Risk, Metabolic Risk, Lifestyle Issue, and Secondary Prevention of Complications. Individual trajectories within the health state space were successfully visualized, capturing transitions along the continuum from robust health through preclinical change to disease onset. Trajectory dynamics across the 33 cluster groups fell into three patterns: stable, single transition, and multi transition. These findings demonstrate the feasibility of quantifying and tracking health to disease pathways from routine checkups and provide a practical framework to align preventive strategies with trajectory type and risk profile.