Integrity Review: UKB-MDRMF: a multi-disease risk and multimorbidity framework based on UK Biobank data

Academic fraud report · Geng Detector

Summary

This report flags the Nature Communications paper 'UKB-MDRMF' (DOI: 10.1038/s41467-025-58724-3) as 'highly suspicious' (高度可疑) following a text-only forensic review. The central concern is a direct contradiction between the authors' claim of 'superior predictive performance' and the data presented in Table 1, where their model's AUC/C-index is lower than comparator SOTA models in multiple diseases (CAD: 0.76 vs 0.88; AF: 0.69 vs 0.75; PC: 0.79 vs 0.86). Additionally, the authors performed Wilcoxon tests across numerous disease types without applying any multiple-comparison correction, a serious statistical omission in a study of this scale. Methodological concerns include subjective (non-algorithmic) feature selection and use of mean imputation for continuous variables. A 'modified publication 2026' note may indicate post-publication revisions. No image-level or code-level verification was performed; verdicts are limited to text-based evidence.

Verdict

🟠 Highly suspicious based on text-only review. The authors' headline claim of superior predictive performance is contradicted by their own Table 1 data across at least three diseases, and core statistical procedures omit multiple-comparison correction at a scale that is difficult to justify. Image manipulation and code reproducibility have not been verified.

Key findings

Self-contradictory performance claims. Table 1 numbers (AUC/C-index) for the proposed UKB-MDRMF are lower than published benchmarks, despite the authors' narrative of superior performance.
Cardiovascular disease (CAD). Comparator (Petrazzini B, et al.) AUC = 0.88; proposed model AUC = 0.76.
Atrial fibrillation (AF). Comparator (Mars N, et al.) C-index = 0.75; proposed model C-index = 0.69.
Prostate cancer (PC). Comparator C-index = 0.86; proposed model C-index = 0.79.
No multiple-comparison correction. Figure 2c/2f legend explicitly states two-sided Wilcoxon tests were run across 21 disease types "without" any correction, creating a high risk of false positives given the high number of tests.
Subjective feature selection. Methods (Page 10) state 542 of 7228 phenotypes were "subjectively selected," which is incompatible with the data-driven framing of the framework.
Primitive imputation. Continuous variables with missing values are imputed using the observed mean (Pattern-II, Page 12), an approach known to deflate variance and inflate significance.
Publication revision flag. Copyright line notes "© The Author(s) 2025, modified publication 2026," indicating post-publication changes.

Evidence highlights

Table 1 vs. main text (Pages 5–6). Bolded cells in Table 1 are described as the better-performing model, yet several entries favor the external comparator by margins of 0.07–0.12 in AUC/C-index.
Figure 2c, 2f legend. Explicit text: "the p values from two-sided Wilcoxon tests in each disease type, and no multiple comparison correction was applied."
Methods, Page 10. "we subjectively selected 542 phenotypes" from 7228 candidate phenotypes.
Methods, Page 12 (Pattern-II). "For continuous variables, we opt to impute missing values using the mean of the observed data."
DOI: 10.1038/s41467-025-58724-3.

Notes

Verdict is text-only; no pixel-level image inspection or code/random-seed audit was conducted.
Confidence is moderate-to-high for the Table 1 contradiction (numbers are explicit in the paper) but lower for inferring intent behind the missing statistical correction.
Recommended follow-ups: request authors' code and train/test split seeds, ask for FDR-adjusted p-values for Figure 2/3 results, file a PubPeer comment on Table 1, and contact the Nature Communications editorial office regarding statistical methodology.
AI-generated report for discussion purposes; not an institutional determination of misconduct.