VascFlux-GEO: Comparator selection and auditability in vascular transcriptome module prediction
- 1Department of Cardiology, Zhongshan Hospital, Fudan University, Shanghai Institute of Cardiovascular Diseases, Shanghai 200032, China.
Abstract
Background: Mathematical coupling and comparator selection are established concerns when a transcriptome response is defined as follow-up minus baseline. We assessed whether apparent within-cohort improvement survives matched constant-follow-up and scalar-shrinkage controls, and whether specified evaluation faults are detectable in an auditable vascular transcriptome benchmark. Results: Using eight frozen modules, 72 rosuvastatin pairs, 14 peripheral artery disease pairs, 20 repeated subject-held-out five-fold partitions, and 5,000 subject bootstrap draws of fixed losses, we compared 15 explicit specifications. Supporting post hoc analyses included 800 known-truth training/test simulations, annotation and input-ablation sensitivities, complete-tuning integrity checks, and an independent-source cross-sectional artery description. In GSE86216, nested follow-up ridge had lower observed repeated-CV mean absolute error than mean follow-up, 0.3280 versus 0.3405 (difference 0.0126; descriptive central 95% percentile range 0.0033-0.0224), while scalar shrinkage had lower observed error (0.3256). These ranges have no validated coverage guarantee. No full-matrix advantage was established; equivalence was not tested. In peripheral artery disease, nested ridge (0.3449) had slightly higher observed MAE than the median follow-up comparator (0.3409), and no advantage was established. Simulations illustrated how mean follow-up can outperform mean change without temporal signal, while the GSE281841 histological stenosis-grade description remained limited by anatomical imbalance, unconfirmed participant identity and an unresolved sample-count discrepancy. Conclusions: These retrospective, nonconfirmatory comparisons show modestly lower observed within-cohort error than constant-follow-up controls in one cohort, without established full-matrix advantage. They do not partition prediction into mathematical, technical and biological components. Independent longitudinal validation and clinical utility remain unestablished.
Related
- Comparator Event Rates, Baseline Activity, and Operational Selection Pressure in Relapsing Multiple Sclerosis Trials: A Meta‐Epidemiological Analysis
- An innovative framework for student adaptability prediction using discriminant-based classification, feature selection, and sensitivity analysis
- Efficacy and safety of duloxetine for osteoarthritis: a comparator-stratified systematic review and meta-analysis
- AI-enabled and comparator digital interventions for adolescent mental health (2020–2025): early detection, human–AI collaboration, and ethical governance—a PRISMA-ScR scoping review
- Acoustic feature selection and clinical information integration for machine learning based dialysis shunt function prediction
BibTeX
@misc{you2026vascfluxgeo,
title = {{VascFlux-GEO: Comparator selection and auditability in vascular transcriptome module prediction}},
author = {L. You and Z. Peng and Y. Guo and L. Shen and J. Ge},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.25.26366405},
url = {https://www.medrxiv.org/content/10.64898/2026.10.25.26366405v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - VascFlux-GEO: Comparator selection and auditability in vascular transcriptome module prediction
AU - L. You
AU - Z. Peng
AU - Y. Guo
AU - L. Shen
AU - J. Ge
PY - 2026
DA - 2026/10/25
PB - medRxiv
DO - 10.64898/2026.10.25.26366405
UR - https://www.medrxiv.org/content/10.64898/2026.10.25.26366405v1
AB - Background: Mathematical coupling and comparator selection are established concerns when a transcriptome response is defined as follow-up minus baseline. We assessed whether apparent within-cohort improvement survives matched constant-follow-up and scalar-shrinkage controls, and whether specified evaluation faults are detectable in an auditable vascular transcriptome benchmark. Results: Using eight frozen modules, 72 rosuvastatin pairs, 14 peripheral artery disease pairs, 20 repeated subject-held-out five-fold partitions, and 5,000 subject bootstrap draws of fixed losses, we compared 15 explicit specifications. Supporting post hoc analyses included 800 known-truth training/test simulations, annotation and input-ablation sensitivities, complete-tuning integrity checks, and an independent-source cross-sectional artery description. In GSE86216, nested follow-up ridge had lower observed repeated-CV mean absolute error than mean follow-up, 0.3280 versus 0.3405 (difference 0.0126; descriptive central 95% percentile range 0.0033-0.0224), while scalar shrinkage had lower observed error (0.3256). These ranges have no validated coverage guarantee. No full-matrix advantage was established; equivalence was not tested. In peripheral artery disease, nested ridge (0.3449) had slightly higher observed MAE than the median follow-up comparator (0.3409), and no advantage was established. Simulations illustrated how mean follow-up can outperform mean change without temporal signal, while the GSE281841 histological stenosis-grade description remained limited by anatomical imbalance, unconfirmed participant identity and an unresolved sample-count discrepancy. Conclusions: These retrospective, nonconfirmatory comparisons show modestly lower observed within-cohort error than constant-follow-up controls in one cohort, without established full-matrix advantage. They do not partition prediction into mathematical, technical and biological components. Independent longitudinal validation and clinical utility remain unestablished.
ER -