English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Integrity Review Report: A 16.4nJ/Class Patient-Independent Prototype-Based Spatio-Temporal CNN Processor with Forward-Inference-Based Adaptation for Robust and Low-Latency Seizure Detection (ISSCC 2026, DOI: 10.1109/ISSCC49663.2026.11409275)

Academic fraud report · Geng Detector

Summary

This AI-assisted integrity review of the ISSCC 2026 paper (DOI: 10.1109/ISSCC49663.2026.11409275) concludes a verdict of "questionable" (yellow flag), driven by a single substantive concern. Visual analysis found that the per-patient post-adaptation bars in Figure 36.3.5(c) align almost perfectly at approximately 96–97%, with minimal inter-patient variation, despite the paper reporting explicit inter-patient domain drift in subjects 4, 8, 9, 14, and 15 on the 24-subject CHB-MIT dataset. Such uniformity is unusual in genuine cross-patient EEG experiments, though a benign explanation (performance saturation across patients after prototype-based adaptation) is possible but unsupported in the text. The report's machine-forensics leads were assessed as inapplicable: the Benford test (MAD=0.0852, p=0.0000) was applied to controlled engineering parameters (40nm, 0.58V, 16.4nJ, 2.736kB SRAM), for which Benford premises fail; PRNU high-similarity flags (multiple NCC=1.000) concern vector-derived architecture diagrams and layouts exported from the same pipeline, a benign explanation supported by the figures themselves. Citations [1]–[11] carry verifiable DOIs, and internal numeric consistency checks (16.4 nJ/class, 23.08 µW, 0.225 mm², 0.58 V, 2 MHz) all matched the source. Confidence is limited: the key finding requires per-patient raw classification data to resolve. Recommended action is to request per-patient results from the authors; no publication to PubPeer or editorial report at this stage.

Verdict

🟡 Questionable (存疑) — This is a typical ISSCC circuit paper (chip design + measured results), not a biomedical clinical study. The only substantive concern is the anomalous regularity of post-adaptation per-patient results in Figure 36.3.5(c). All machine-forensics leads were evaluated as based on inapplicable premises for this paper type and were cleared. No misconduct finding is asserted; the open question can only be resolved with raw per-patient data.

Key findings

  • Finding 1 (substantive, 🟠): In Figure 36.3.5(c) ("Balanced Accuracy Improvements by Proposed Forward-Inference-Driven Adaptation"), the "w. adaptation" per-patient bars align almost perfectly at ~96–97% with minimal individual variation and no outliers; the reported average is 96.6%. Given the paper explicitly notes initial domain drift in subjects 4, 8, 9, 14, 15, such uniformity across 24 patients is rare in genuine cross-patient EEG experiments. Possible benign explanation — post-adaptation performance saturation — is not supported anywhere in the text and remains speculative.
  • Finding 2 (record only, ⚪): References [1]–[11] all carry verifiable DOIs; the citation timeline (2021–2025 supporting a 2026 publication) is reasonable. Comparison table (Fig. 36.3.6) "This Work" row matches the Specification table (Fig. 36.3.7) exactly: 16.4 nJ/class (0.016* µJ/class), 23.08 µW, 0.225 mm², 0.58 V, 2 MHz, 2.736 kB SRAM.
  • Finding 3 (record only, ⚪): The claimed 24 subjects are internally consistent with CHB-MIT conventions (chb01–chb24), and 6 cited data points were programmatically verified in the source text. Per-patient data authenticity cannot be independently verified from the text (links to Finding 1).
  • Finding 4 (disposition record, ⚪): Benford test (MAD=0.0852, p=0.0000) was applied to controlled engineering parameters (40nm, 0.58V, 16.4nJ, 0.0000247*) — premises not satisfied, cleared. PRNU high-similarity flags (multiple NCC=1.000) target vector-derived architecture diagrams/schematics/layout (Fig. 36.3.1–36.3.4, 36.3.7) exported from the same EDA/drawing pipeline — benign, cleared. Noise-variance flag (I1) concerns scan/compression artifacts — cleared. Note: the forensic parameter "PCE=128.0" could not be located in the source text (forensic-computation metric; NCC=1.000 is supported in the forensics block).
  • Evidence highlights

  • Reported metrics verified in source: 96.6% ("Avg. balanced accuracy w. adapt"), subject-drift list (4, 8, 9, 14, 15), "24 subjects" on CHB-MIT.
  • Spec consistency: 16.4 nJ/class ↔ 0.016* µJ/class in comparison table; 23.08 µW; 0.225 mm²; 0.58 V; 2 MHz; 2.736 kB SRAM — no internal contradictions.
  • Benford: MAD=0.0852, p=0.0000 — premise inapplicable to engineering parameters.
  • PRNU: multiple NCC=1.000 — attributable to same-pipeline vector exports.
  • Notes

  • Recommended action: contact authors for raw per-patient per-classification results (priority: Finding 1). PubPeer posting and editorial escalation are not recommended at this evidence strength.
  • Final misconduct determination requires a professional institutional investigation; this AI-assisted report is for academic discussion only, with known possibilities of false positives and false negatives.
DOI: 10.1109/ISSCC49663.2026.11409275

Tags

#academic-integrity#questionable-verdict#data-uniformity#figure-analysis#forensics-cleared#benford-test#prnu#chip-design

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/report/geng_geng_6af9bb707d8d12.16892477