Investigation report: Shao et al., 'Radiologist-like artificial intelligence for grade group prediction of radical prostatectomy' (Theranostics 2020, DOI: 10.7150/thno.48706)

Academic fraud report · Geng Detector

Summary

This report flags a Theranostics 2020 paper (DOI: 10.7150/thno.48706) by Lizhi Shao and colleagues as highly suspicious due to mathematical inconsistency in a headline result. The abstract claims the AI model 'reduced the overall downgrading rate by 6.4% (P = 0.029),' yet the Results section states that needle-biopsy (NB) downgrading was 14.7% and that PCa-GGNet achieved 6.3% downgrading in all testing samples; the difference (8.4%) or relative reduction (~57%) cannot reconcile with 6.4%. The figure 6.4% appears elsewhere in the text for a specific subgroup (GG 3), suggesting a copy-paste error in the abstract. Additional issues include a missing ethics/IRB statement for a 575-patient retrospective study, an outdated R v.3.1.0 (2014) for 2020 analyses, and a minor possible discrepancy between Figure 6A and 6E totals (278 vs. 279) that could not be verified from the text alone. Verdict: highly suspicious, primarily because of unresolvable arithmetic in a key claim.

Verdict

Highly suspicious. The central quantitative claim in the abstract is internally inconsistent with the numbers given in the Results section, indicating either a serious computational/copy error or unreliable reporting. Additional methodological and compliance concerns strengthen the suspicion.

Key findings

Arithmetic mismatch in the headline downgrading claim: abstract says 'reduced the overall downgrading rate by 6.4% (P = 0.029),' but Results report NB = 14.7% and PCa-GGNet = 6.3% (absolute reduction 8.4%, relative ~57%); neither reconciles to 6.4%.
The value 6.4% appears in the text as the downgrading rate for a specific subgroup (GG 3), strongly suggesting the abstract figure was mis-copied from that subgroup rather than independently computed.
Missing ethics/IRB approval statement and informed-consent waiver for a retrospective study of 575 human subjects (Materials & Methods / Patients section).
Statistical software reported as R v.3.1.0 (released April 2014) in a paper submitted in May 2020, an unusually outdated tool for cutting-edge deep-learning analyses.
Possible minor numerical discrepancy between Figure 6A (ring chart total = 278) and Figure 6E (confusion matrix total = 279) for the primary cohort (PC, N=279); not verifiable from the published text and flagged as unconfirmed.

Evidence highlights

Abstract statement (Theranostics 2020; 10(22): 10200-10212): 'reduced the overall downgrading rate by 6.4% (P = 0.029).'
Results section (Page 10207, Figure 4): NB overall downgrading = 14.7%; PCa-GGNet downgrading in all testing samples = 6.3%; upgrading = 12.5%.
Subgroup text where 6.4% actually appears: downgrading to GG 3 (as cited in the original Chinese report).
Materials & Methods (Page 10201): no IRB approval number or informed-consent waiver statement found in the provided text, despite inclusion of 575 patients (2010-2019).
Evaluation section (Page 10205): statistical analyses performed in 'R (v.3.1.0).'
Figure 6A vs. 6E: visual analysis of the ring chart gives 67+51+70+38+52 = 278 for the PC cohort, while the corresponding confusion matrix sums to 279; numerical text not available to confirm.
DOI: 10.7150/thno.48706

Notes

The arithmetic inconsistency is the most serious red flag: a paper's headline claim should be reproducible from its own reported numbers, and here it is not.
The ethics-statement omission is common in older Theranostics submissions but is still a compliance issue for retrospective human-subject research.
The R v.3.1.0 observation alone is not evidence of misconduct but reflects unusual software-staleness for 2020 AI/biostatistics work.
Figure 6 totals discrepancy is reported as '依据不足' (insufficient evidence) by the original reviewer; treat as unconfirmed.
All findings rely on the publicly available PDF; raw data and code were not examined. A definitive determination of misconduct requires institutional or editorial investigation.