Integrity review of 'Cascaded Improved Neural Network for the Reconstruction, Classification, and Unmixing of the Raman Spectra of Mixed Microplastics' (Anal. Chem., DOI: 10.1021/acs.analchem.5c04049)

Academic fraud report · Geng Detector

Summary

This automated review assesses the above paper, published in Analytical Chemistry, and rates it as highly suspicious. Three main concerns are identified. First, Table 1 reports that the ECA-ResUNet model has 13.83M parameters, exceeding SE-ResUNet (13.38M) and CBAM-ResUNet (13.82M), which contradicts the authors' own statement and the established design principle that ECA uses 1D convolutions instead of fully connected layers to reduce parameters. This suggests a coding error or fabricated comparison figures. Second, Table 4 shows that Dataset 3 (150 mW, 1000 ms) and Dataset 8 (50 mW, 3000 ms), both delivering 150 mJ total energy, produce classification accuracies of 96.78% and 91.17% respectively, a >5% gap that lacks physical explanation and is inconsistent with the optical reciprocity law. Third, in Table 6, the validation and independent test MSEs for Component 1 are identical (9.20e-7) and R2 for Component 2 is also identical (0.99955), a coincidence suggesting potential data duplication. The findings require independent verification.

Verdict

Highly suspicious (🟠). The paper presents multiple internal inconsistencies between claimed methodology, reported numerical results, and known physical principles. A manual investigation is recommended.

Key findings

Attention module parameter count anomaly (Table 1, p. 8015): The ECA-ResUNet is reported to have 13.83M parameters, which is higher than SE-ResUNet (13.38M) and CBAM-ResUNet (13.82M). This contradicts the authors' own description and the established ECA design (1D convolutions replacing FC layers to reduce parameters).
Apparent violation of optical reciprocity law (Table 4, p. 8017): Dataset 3 (150 mW × 1000 ms = 150 mJ) yields 96.78% accuracy, while Dataset 8 (50 mW × 3000 ms = 150 mJ) yields only 91.17% accuracy, a >5 percentage point gap for identical delivered energy, without a physical justification.
Suspicious validation/test set metric duplication (Table 6, p. 8019): For Component 1, both validation and independent test MSE equal 9.20×10⁻⁷ exactly. For Component 2, both R² values equal 0.99955 exactly, suggesting possible data duplication or no real test-set split.

Evidence highlights

Table 1 parameter counts: ECA-ResUNet 13.83M > SE-ResUNet 13.38M; CBAM-ResUNet 13.82M, despite the paper explicitly stating that 1D convolutions were used 'in lieu of a fully connected layer' to reduce parameters.
Table 4 energy calculation: 150 mW × 1000 ms = 150 mJ vs. 50 mW × 3000 ms = 150 mJ, with resulting accuracies 96.78% vs. 91.17%.
Table 6 duplicated values: Component 1 val MSE = 9.20×10⁻⁷, test MSE = 9.20×10⁻⁷; Component 2 val R² = 0.99955, test R² = 0.99955.
DOI: 10.1021/acs.analchem.5c04049

Notes

All quoted numbers are taken directly from the report and the paper; DOI 10.1021/acs.analchem.5c04049 is preserved.
The reciprocity-law argument presumes the acquisitions were below detector saturation; the authors should clarify whether saturation, dark-current accumulation, or sample degradation could explain the gap, but no such explanation is given in the manuscript.
The parameter anomaly could stem from a bug in the open-source code rather than deliberate fabrication; checking the repository is the most direct way to disambiguate.
Identical validation/test metrics to multiple decimal places are improbable in real deep-learning experiments but could theoretically arise from rounding at very small variance; the authors should provide the underlying data.
This review is AI-assisted; final judgment of academic misconduct requires a formal institutional investigation.