Investigation Report: Deep Learning with Quantitative Features of Magnetic Resonance Images to Predict Biochemical Recurrence of Radical Prostatectomy: A Multi-Center Study (Cancers, 2021)

Academic fraud report · Geng Detector

Summary

Verdict: Highly suspicious. This report evaluates a 2021 multi-center deep-learning study (DOI: 10.3390/cancers13123098) that uses quantitative MRI features to predict biochemical recurrence after radical prostatectomy. Four interrelated concerns are raised. First, an internal inconsistency in cohort size: the text and Table 1 define the primary cohort (PC) as n=368, while Table 2 states n=369. Second, statistical anomalies: the CAPRA-S C-index is reported as exactly 0.654 in both independent validation cohorts (VC1 n=34; VC2 n=83), and an NCCN hazard ratio is given as 1.9022, an implausible four-decimal precision. Third, Table 2 is titled "Patient characteristics" but contains model performance metrics, indicating copy-paste errors. Fourth, decision curves in Figure 2(g-i) show implausibly jagged oscillations in the small validation sets, suggesting overfitting or uncleaned data. These findings together indicate sloppy or potentially fabricated reporting. Confidence is moderate; conclusions should be verified against raw data.

Verdict

Highly suspicious. Multiple internal inconsistencies and statistical implausibilities were identified in the paper. The cumulative pattern points to either serious methodological negligence or potential data manipulation, warranting formal investigation.

Key findings

Internal sample-size contradiction (n=368 vs n=369): The primary cohort is described as 368 patients in the text and Table 1, but Table 2 lists it as n=369.
Identical C-index across independent cohorts: VC1 (n=34) and VC2 (n=83) report the same CAPRA-S C-index of 0.654, statistically improbable given differing sample sizes and centers.
Implausible decimal precision: The NCCN hazard ratio is reported as 1.9022, retaining four decimals and ending in an unusual trailing digit.
Mismatched table title: Table 2 is captioned "Patient characteristics" but displays model performance metrics (HR, p-values, C-index).
Visually abnormal decision curves: Figure 2(g-i) DCA curves for the small validation cohorts show jagged, high-frequency oscillations suggestive of overfitting or unprocessed outliers.

Evidence highlights

Location: Section 3.1, Table 1, Table 2, Figure 2(g-i).
Numerical evidence:
PC n=368 (text, Table 1) vs PC n=369 (Table 2 header).
CAPRA-S C-index: 0.654 [0.49–0.818] in VC1 (n=34); 0.654 [0.544–0.764] in VC2 (n=83).
NCCN HR = 1.9022.
Visual evidence: Decision curves for VC1 (n=34) and VC2 (n=83) display sharp sawtooth patterns inconsistent with typical DCA plots.
DOI of the paper under review: 10.3390/cancers13123098.

Notes

The report's severity ratings (🟠/🟡) reflect the reviewer's assessment of impact on scientific validity. Identical C-index values across independent cohorts and the sample-size discrepancy are the strongest indicators; however, the possibility of transcription or copy-paste errors rather than deliberate fabrication cannot be excluded without raw data. Verification of original CAPRA-S calculation code and VC1/VC2 raw outputs is recommended. This investigation has not yet been escalated to the journal or institutional committee.