Verdict
Highly suspicious (🟠), no conclusive proof. The review identifies multiple independent unresolved concerns, but none reaches the threshold of confirmed fabrication. The core conclusion is that the authors should verify/clarify, not that misconduct is established. Note an internal discrepancy: the report narrative assigns an orange (suspicious) rating, while the automated Bayesian synthesis section labels the paper 🔴 (posterior ≥99.9%, BF = 2.01e+11); the synthesis is programmatic, assumes conditional independence of evidence, and explicitly does not constitute a finding of fraud.
Key findings
- Statistical tests excluded (Finding 1): Benford (first-digit MAD = 0.0667, χ² = 198.2, p = 0.0000) and last-digit tests (χ² = 53.8, p = 0.0000; odd/even ratio 45/121) fail their premises on engineering-spec values (55-nm CMOS, 0.98 mm², 0.07 µJ, 25.8 µW) and rounded percentages. Excluded; not evidence against the paper.
- Image forensics signals retained (Finding 2): Across 21 embedded figures, noise-variance inconsistency (CV 0.80–1.44; img-106 and img-123 flagged "highly uneven"; img-106 CV = 1.24, anomalous blocks 0/90) and PRNU weak similarity for img-105 vs. 14 other images (NCC 0.010–0.073; e.g., img-105 vs img-116: NCC = 0.010, PCE = 23.2; vs img-118: NCC = 0.026, PCE = 21.4). Candidate benign explanations (repeated axis/frame structures, die-photo illumination gradients, IEEE PDF compression) are not affirmatively supported by the paper.
- Fig. 5 waveform similarity (Finding 3, medium confidence): Easy-to-classify vs. difficult-to-classify non-seizure waveforms show highly similar noise patterns, amplitude envelopes, and high-frequency detail, suggesting possible scaling/filtering of one signal; the paper does not state the waveforms' origin.
- Fig. 18 uniform improvement (Finding 4, low confidence): All 24 patients show post-on-chip-learning accuracy no worse than before; text claims 30–40% improvement for patients 13, 14, 15, 20. Combined with claimed 100% event-based sensitivity, this raises overfitting/selective-tuning questions; the paper does not describe tuning/selection procedures.
- Editorial flaws confirmed, not misconduct (Finding 5): "Zero-Short-Retraining" vs. "zero-shot" inconsistency in Section III-C; broken figure references ("Fig.\nHAPE", "Fig.\nLCSM"); reference [22] name encoded as "R. M´'uller" (PDF extraction artifact).
- DOI: 10.1109/JSSC.2024.3446244 — IEEE JSSC, Vol. 59, No. 11, November 2024.
- img-106 noise variance: CV = 1.24, anomalous blocks 0/90 (logLR = 0.99); img-123 also flagged highly uneven.
- PRNU comparisons: img-105 vs img-116 (NCC = 0.010, PCE = 23.2, logLR = 0.82); img-105 vs img-118 (NCC = 0.026, PCE = 21.4, logLR = 0.82); batch range NCC 0.010–0.073.
- Automated synthesis: posterior ≥99.9%, 95% CI [100%, 100%], BF = 2.01e+11 — flagged as likely overestimated due to conditional-independence assumptions; prior fixed at 5%.
- 6 machine-forensics leads were excluded during review (benign explanations / failed premises / illustrative items).
- No finding meets the "smoking gun" threshold; the report explicitly frames outcomes as "requires author self-verification" rather than "fraud established."
- Recommended actions: request original high-resolution figures; ask authors about Fig. 5 waveform provenance; request per-patient accuracy data and tuning records for Fig. 18; consider a neutral PubPeer inquiry if unresponsive; do not yet escalate to journal/institution.
- Limitations: analysis based on published PDF only; no raw images or per-patient data; visual analyses carry low-to-medium confidence; AI-assisted report, subject to false positives/negatives. Final determinations require institutional investigation.