Panvascular-JEPA: Unseen-Structure Evaluation of Factorized Predictive States for Hemodynamic Networks
- 1Department of Cardiology, Zhongshan Hospital, Fudan University, Shanghai, China
Abstract
Predictive representations for vascular simulation should be tested on unseen network structures and compared with the governing equations before being interpreted as general-purpose world models. We constructed a new, dimensioned but uncalibrated nodal resistance-inertance-compliance benchmark containing all 38 connected, non-isomorphic six-node graphs with five to seven edges. Twenty-four structures supplied training, validation, and new-background testing; 14 structures were wholly withheld. Across 648 parameter backgrounds and two prescribed action programs, 1,296 trajectories contained 96 forecast steps over three cycles. Four graph predictors received equal six candidate optimization searches, followed by five fixed-seed fits. Selected configurations and analysis code were locked before test trajectories were generated. Comparators included known-equation numerical integrators, an inertia-free reduction, and a five-coefficient model fitted to training trajectories. Unseen-structure mean standardized RMSE was 8.46 x 109 for four-factor OPF, 329 for single factor OPF, and 1.83 x 1031 for supervised graph prediction, compared with 1.09 x 10-4 for foursubstep known-equation integration and 6.77 x 104 for the fitted mechanism. The primary OPF-4 minus OPF-1 difference was 8.46 x 109, with a paired stratified-graph 95% percentile interval of [2.23 x 105, 2.54 x 1010]. Neural errors were strongly dependent on training seed and network structure, with extreme forecasts dominating the means. All four neural means exceeded persistence; relative performance within this comparison did not establish a usable forecasting model. The study supplies a distinct numerical test of structure transfer, factor-count control, and recursive stability. Its conclusions concern the implemented models, optimization budget, and finite graph catalogue. It establishes neither patient-specific performance nor anatomical transfer, biological remodeling, device efficacy, or clinical readiness.
Related
- PanVasc Research for AI assisted evidence analysis in panvascular intervention
- Reward changes the use and structure of predictive spatial maps
- DCAF: Unifying Data Validity, Evaluation Validity and Predictive Modeling for Single-cell Perturbation Simulation
- Similar Predictive Fit but Different Latent Dynamics: Characterizing Learned Dynamical Structure in Personalized Models of Brain Disorders
- Mangrove taxon mapping from Sentinel-1/2 using a compact one-dimensional network with probability calibration, predictive entropy, and frozen transfer evaluation
BibTeX
@misc{you2026panvascularjepa,
title = {{Panvascular-JEPA: Unseen-Structure Evaluation of Factorized Predictive States for Hemodynamic Networks}},
author = {L. You and Z. Feng and C. Gao and Y. Wu and Z. Yin and L. Yang and L. Shen and J. Ge},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.12.758273},
url = {https://www.biorxiv.org/content/10.64898/2026.10.12.758273v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Panvascular-JEPA: Unseen-Structure Evaluation of Factorized Predictive States for Hemodynamic Networks
AU - L. You
AU - Z. Feng
AU - C. Gao
AU - Y. Wu
AU - Z. Yin
AU - L. Yang
AU - L. Shen
AU - J. Ge
PY - 2026
DA - 2026/10/12
PB - bioRxiv
DO - 10.64898/2026.10.12.758273
UR - https://www.biorxiv.org/content/10.64898/2026.10.12.758273v1
AB - Predictive representations for vascular simulation should be tested on unseen network structures and compared with the governing equations before being interpreted as general-purpose world models. We constructed a new, dimensioned but uncalibrated nodal resistance-inertance-compliance benchmark containing all 38 connected, non-isomorphic six-node graphs with five to seven edges. Twenty-four structures supplied training, validation, and new-background testing; 14 structures were wholly withheld. Across 648 parameter backgrounds and two prescribed action programs, 1,296 trajectories contained 96 forecast steps over three cycles. Four graph predictors received equal six candidate optimization searches, followed by five fixed-seed fits. Selected configurations and analysis code were locked before test trajectories were generated. Comparators included known-equation numerical integrators, an inertia-free reduction, and a five-coefficient model fitted to training trajectories. Unseen-structure mean standardized RMSE was 8.46 x 109 for four-factor OPF, 329 for single factor OPF, and 1.83 x 1031 for supervised graph prediction, compared with 1.09 x 10-4 for foursubstep known-equation integration and 6.77 x 104 for the fitted mechanism. The primary OPF-4 minus OPF-1 difference was 8.46 x 109, with a paired stratified-graph 95% percentile interval of [2.23 x 105, 2.54 x 1010]. Neural errors were strongly dependent on training seed and network structure, with extreme forecasts dominating the means. All four neural means exceeded persistence; relative performance within this comparison did not establish a usable forecasting model. The study supplies a distinct numerical test of structure transfer, factor-count control, and recursive stability. Its conclusions concern the implemented models, optimization budget, and finite graph catalogue. It establishes neither patient-specific performance nor anatomical transfer, biological remodeling, device efficacy, or clinical readiness.
ER -