Interpretable dynamic graph learning for low-frequency cohorts: mapping pseudotime windows to CKM network dynamics
Z Qi, J L Liu, T G Gu, J L Liu, T L LiuAbstract
Background
Cardiovascular-Kidney-Metabolic (CKM) Syndrome progresses through evolving inter-organ coupling: biomarker interactions rewire as disease advances, yet clinical staging treats each risk factor in isolation. Dynamic network methods—proven in high-frequency ICU settings—have never been adapted to low-frequency population cohorts lacking real-time data. By projecting individuals onto a continuous biological trajectory (pseudotime), we construct the first population-scale atlas of how cardiometabolic coupling networks reconfigure across CKM progression.
Purpose
To quantify stage-dependent rewiring of biomarker coupling networks across the CKM continuum and derive a personalised Network Interaction Score (NIS) translating shifting inter-biomarker dependencies into actionable cardiovascular risk stratification.
Methods
In a Chinese national population-based cohort (n=5,386), individual cardiometabolic profiles were mapped onto a continuous pseudotime axis via tree-structured manifold learning to reconstruct the CKM disease trajectory, which was then partitioned into six progression windows (B1–B6). Consensus biomarker networks (18 markers, 6 domains) were estimated per window using graphical lasso (200-bootstrap, threshold ≥80%) to identify stable phenotype association. Reconfiguration was quantified by Frobenius distance (global network shifts), Jaccard edge-turnover (phenotypic rewiring), and hub-strength migration (transition of dominant risk markers). NIS—capturing personalised cumulative network reconfiguration—was derived from the top-decile most dynamic edges and validated in multivariable Cox models against composite CVD and stroke.
Results
A reproducible hub-migration signature (reflecting the transition of primary risk markers) emerged: lipid-cluster dominance (TC–TG–HDL–LDL strength 2.2–2.9) in early disease trajectories gave way to cardiorenal coupling as the prevailing phenotypic motif by advanced windows, with B2→B3 representing the sharpest topological inflection point (Frobenius=1.575). Advancement along the pseudotime trajectory conferred graded CVD hazard after full covariate adjustment: B4 HR=2.07 (95% CI 1.57–2.72), B5 HR=1.82 (1.39–2.40), B6 HR=2.65 (2.03–3.46), all P<0.001. NIS independently predicted incident stroke: highest-quintile participants carried 52% excess hazard versus the lowest (HR=1.52, 95% CI 1.09–2.11, P=0.014).
Conclusions
Pseudotime-anchored network analysis reveals that CKM progression systematically rewires biomarker coupling from lipid-dominated to cardiorenal-dominated interactions—a dynamic invisible to conventional staging. The personalised NIS captures this evolving dependency and independently stratifies stroke risk. Clinically, coupling signatures can alert physicians that one biomarker abnormality implies co-evolving risk in linked domains, enabling integration into health record systems as automated cross-domain surveillance prompts for coordinated intervention.