A dynamic health-state manifold maps instability and heterogeneous aging progression from longitudinal clinical records
Longitudinal clinical records offer a direct view of how human health changes over time, yet most computational models compress this information into endpoint-specific risk estimates or single-time biological-age scores. Such summaries are useful, but they do not fully describe where an individual is located in health-state space. To address this gap, we developed VitaFlux, a dynamic state-space framework for modelling aging trajectories from irregular longitudinal clinical data. VitaFlux learns visit-level latent health states by forecasting next-visit clinical representations rather than by directly supervising the model with chronological age or predefined disease labels. This design uses temporal transition as the organizing signal, allowing the model to learn a reusable coordinate system in which health-state position, short-term instability, and longer-term progression can be analysed together.
We applied VitaFlux to the Zhejiang Population Health Cohort, a population-scale longitudinal health-examination resource containing 178,970 individuals and 4,439,390 aligned visits. The model integrated static baseline attributes, historical exposure and event records, repeated clinical measurements, and observation-structure descriptors. Routine health-examination variables were harmonized into a common patient-specific timeline. For each training sample, model inputs were restricted to information available at or before an anchor visit, while supervision was derived only from measurements observed at the subsequent visit. Continuous values, short categorical text, and longer clinical text were encoded through modality-specific representation pipelines and then projected into a shared latent token space. The VitaFlux architecture summarized baseline context, historical measurement profiles, exposure and event histories, and scalar descriptors before using a gated temporal transition module to predict the future latent state. Training optimized contrastive alignment between predicted and observed next-visit representations, together with auxiliary abnormality supervision and temporal smoothness regularization.
The learned latent space formed a continuous health-state manifold organized by chronological age, disease burden, and multisystem physiology. Chronological age varied smoothly across the manifold. Individuals with no chronic disease, a single chronic disease, and multimorbidity occupied progressively shifted regions, and hypertension and diabetes showed disease-specific enrichment patterns that only partly followed the age gradient. Community-level summaries further revealed coordinated clinical variation across metabolic, hepatic, renal, hematological, cardiovascular, and anthropometric domains. These results suggest that VitaFlux captured structured multisystem health variation without requiring direct age-label or disease-label supervision during representation learning.
We next evaluated movement on the manifold as a readout of dynamic vulnerability. Short-term health-state instability was defined as the Euclidean displacement between adjacent visit representations. Instability showed strong spatial heterogeneity across the manifold and was not explained solely by follow-up interval. High-instability regions were enriched for visits followed by greater organ-system involvement and higher subsequent mortality risk. In multivariable-adjusted models controlling for chronological age, sex, baseline organ-system burden, baseline cardiometabolic stage, follow-up interval, and measurement density, each 1-standard-deviation increase in instability was associated with higher odds of mortality at 1 year, 3 years, and 5 years. These associations were also evident in lower-risk subgroups, including individuals without baseline organ-system burden or with low baseline cardiometabolic stage. Thus, displacement in the learned health-state space provided information about future adverse outcomes beyond static baseline status, although causal interpretation remains outside the scope of this retrospective observational analysis.
Beyond adjacent-visit instability, VitaFlux enabled individual trajectories to be traced across repeated follow-up. Population age-bin centroids aligned along a broad aging axis, and individual movement along and away from this axis revealed heterogeneous trajectory morphologies. We summarized long-term progression using trajectory pace of aging (TPoA), a trajectory-derived score designed to capture adverse longitudinal progression on the manifold. Higher TPoA was associated with greater disease-burden gain, faster biological-age deviation drift, higher incident disease-family counts, and higher 5-year mortality. Visit-indexed TPoA estimates were temporally stable across repeated observations and remained informative across strata defined by follow-up duration, baseline burden, sex, and age. TPoA and instability provided complementary information: instability captured short-term state displacement, whereas TPoA summarized longer-term trajectory progression. Marker-level analyses linked TPoA-related variation to glucose metabolism, adiposity, renal function, liver function, hematological indices, and cardiovascular status, supporting the interpretation of TPoA as a multisystem trajectory summary rather than a single-biomarker signal.
We also evaluated whether VitaFlux representations supported downstream risk stratification. Adding latent-state representations to simple clinical anchors improved prediction across classifier families, and forecasted future states provided selective gains for outcomes in which near-term state change was informative. Improvements were most evident for selected cerebrovascular, neuropsychiatric, metabolic, ophthalmic, and mortality-related tasks, and the incremental value of the forecasted state was greatest among individuals with larger predicted transition magnitude. Exploratory branch-topology analysis further suggested that trajectories followed a common trunk before diverging into downstream branches with different burden-accumulation profiles. A localized branchpoint-adjacent region was associated with later entry into a higher-burden branch and was characterized by greater instability, higher prior burden, and higher TPoA. Cardiometabolic markers, including systolic blood pressure, body mass index, fasting glucose, triglycerides, creatinine, and HDL cholesterol, helped clinically annotate this region, although the branch analysis should be interpreted as hypothesis-generating rather than causal.
Finally, we assessed external transfer in two independent aging cohorts, CHARLS and CLHLS. Despite differences in sample size, variable availability, visit density, and cohort design, transferred representations preserved age-related ordering and burden-related anchoring after harmonization. Light adaptation improved quantitative future-state forecasting, and a source-derived trajectory proxy retained ordered relationships with laboratory burden, aging-axis position, instability, and adverse-event rates. These results support partial transfer of the learned geometric backbone and trajectory-related ordering, while also indicating that cohort-specific calibration remains necessary.
Overall, VitaFlux provides a dynamic representation framework for studying aging from longitudinal clinical records. Its main contribution is not a single disease-risk model or a new biological-age clock, but a reusable health-state coordinate system that links current position, short-term instability, and longer-term progression. By organizing irregular clinical histories into a shared state space, VitaFlux may support future studies of aging heterogeneity, disease-transition risk, trajectory-informed stratification, and interpretable population health modelling.