{"id":"cf6f6603-9b56-4067-9d3d-5d2448815690","arxiv_id":"2507.14557","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A derivative-multiplication preprocessing step modestly improves PCR glucose predictions from FTIR blood spectra, but the ARA method's reported accuracy depends on a fitted error term and no data are shared.","lead":"This paper tried a new way to clean up infrared measurements of blood sugar: it multiplies the measured light absorption by mathematical derivatives of the water and glucose reference signals. Glucose predictions got a little better in a 200-sample test, but the study does not share data and one of its models relies on a fitted error term.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on the unproven assertion that multiplying by the first derivative of water absorbance removes water interference; the operation only nulls one peak and reweights the rest, so the reported improvements may be artifacts.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern that I do: the derivative-multiplication step is asserted to remove water interference without a physical derivation. My analysis confirms that the transform is a pointwise reweighting, not a subtraction, and that it multiplies the glucose signal by the same derivative, so the claimed 'removal' is only a zero at one peak maximum. This is the central claim of the paper; if it fails, the improved RMSE values and the visual deductions are not evidence of a valid method. The ARA result is additionally compromised by tuning R on the same data, and the validation split does not control for subject leakage, but those are secondary to the unsupported preprocessing premise. I credit the paper for explicitly acknowledging that the original ARA was inaccurate and for giving enough equations to see the structure of DMS, but the central assumption remains unvalidated. The proposed synthetic test would settle the matter cleanly: if derivative preprocessing genuinely removes water, it should outperform NET on simulated mixtures with known glucose concentrations. Until such a test is performed, rejection is appropriate, and the verdict should remain unchanged.","tokens_in":9269,"tokens_out":2784,"duration_ms":36188,"concrete_test":"Construct a synthetic validation set: generate 200 absorbance spectra as A_i(ν̃) = c_g,i ε_g(ν̃) + c_w ε_w(ν̃) + noise, with known c_g,i spanning the reported range and c_w varying per sample. Apply DMS, DMR, DMSR, and NET, then run PCR with an identical 90/10 split (reporting predictions on held-out samples). If the derivative methods do not yield lower RMSE than NET, or if they introduce a bias/nonlinearity in recovered c_g that NET does not, the water-removal premise fails. Additionally, compute the residual water contribution ε_w(ν̃)·dA_water/dν̃ across the full range; if it is not near zero over a broad region, the claim of 'effective removal' is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DMS/DMR/DMSR preprocessing 'effectively removes' water interference and improves PCR-based glucose prediction. The justification given in the Results is that the derivative vanishes at a peak maximum, so multiplying by dA_water/dν̃ suppresses water peaks. This is physically incomplete. Modeling the measured absorbance as A = c_g ε_g + c_w ε_w + other, the DMS transform is A_DMS = A · w(ν̃), where w = dA_water/dν̃ normalized. The water term becomes c_w ε_w(ν̃) · dA_water/dν̃, which is zero only where the derivative is zero or ε_w is zero. At most wavenumbers, water absorption is nonzero and its derivative is nonzero, so water is not removed; it is scaled by a data-dependent weight. Simultaneously, the glucose term c_g ε_g(ν̃) is multiplied by the same weight, so the linear Lambert–Beer relation between glucose concentration and the processed absorbance is destroyed: the processed spectrum is bilinear in c_g and the derivative. No derivation, simulation, or independent validation shows that glucose information is preserved or that the water contribution is actually canceled. The visible disappearance of the 5200 cm⁻¹ peak is a trivial consequence of the derivative being zero at that local maximum, not evidence of water removal elsewhere. The quantitative improvements (NET 13.14 vs. DMS 12.50 mg/dL RMSE) are small, from a single 90/10 split with repeated samples from the same subjects, and no significance testing is performed. Thus the central scientific claim—that derivative multiplication is an effective water-removal preprocessing step—is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports FTIR transmittance measurements of blood samples from twenty subjects over five days (200 samples, 4000–8000 cm^-1) and compares principal component regression (PCR) glucose predictions based on net absorbances (obtained by subtracting a scaled water spectrum) with three preprocessing variants: multiplication by the normalized first derivative of the water absorbance (DMS), multiplication by (1 minus the first derivative of the glucose absorbance) (DMR), and their combination (DMSR). The authors report modest RMSE improvements over net absorbances (e.g., 13.14 to 12.50 mg/dL for DMS in range I), fewer principal components for the derivative-based methods, and an additional linear-regression approach called Absorptivity Regression Analysis (ARA) that yields an RMSE of 22.42 mg/dL after setting an error-compensation vector R to 105. The central claim is that multiplying spectra by the first derivative of water absorbance 'effectively removes' water interference and makes glucose peaks more distinguishable.","tokens_in":9631,"tokens_out":6529,"duration_ms":75362,"significance":"If the central claim were established, the proposed preprocessing would be a simple and inexpensive addition to NIR-based glucose chemometrics. The paper's strengths include an explicit acknowledgment of the scaling-factor ambiguity in water subtraction, a comparison across three wavenumber windows, and the use of Clarke error grids. However, the core mechanism is not demonstrated mathematically, the reported gains are small and unquantified, and the ARA result is obtained by fitting an error term on the same data used for evaluation. No code, data, or subject-independent validation is provided, so the claims cannot be independently checked or reproduced from the manuscript. The paper's central assertion therefore remains unsupported, and the evidence presented is not sufficient to justify its conclusions.","major_comments":[{"comment":"The claim that multiplying by dA_water/dν̃ 'effectively removes' water interference is not supported. For a Beer-Lambert model A = c_w ε_w + c_g ε_g + ..., the DMS transform multiplies the entire spectrum by w(ν̃) = (dA_water/dν̃)/||dA_water/dν̃||; the water term becomes c_w ε_w(ν̃) w(ν̃), which vanishes only at isolated wavenumbers, while the glucose term c_g ε_g(ν̃) is rescaled by the same factor. This is a data-dependent reweighting, not a subtraction or an orthogonalization, and it destroys the linear relation between A and c_g. The disappearance of the 5200 cm^-1 peak in Figure 3 is a trivial consequence of the derivative being zero at that local maximum and does not demonstrate removal of water at other wavenumbers. A derivation, a simulated two-component experiment, or subject-independent validation is required before this operation can be said to preserve glucose information or cancel water absorption.","section":"§Results and discussion (DMS/DMR definitions)"},{"comment":"The ARA result is not an independent prediction in this manuscript. Equation (1) defines δ_i arbitrarily as ν̃_i/η_i, and the authors state that the original ARA 'did not provide accurate estimation' on this dataset. The error-compensation vector R is introduced and then set to 105 in the Results section specifically to minimize error on the same measured data (Figure 8). The resulting RMSE of 22.42 mg/dL is therefore a fitted quantity, and no held-out or cross-validated assessment of ARA is reported. The conclusion that ARA provides 'acceptable accuracy' is not established by the analysis as presented.","section":"§Mathematical method (ARA, Eq. (1))"},{"comment":"The PCR comparison rests on a single 90/10 partition of 200 samples drawn from only twenty subjects with repeated daily measurements. Because samples from the same subject appear in both training and test sets, the model can memorize subject-specific baselines, and the reported RMSE differences (Table 1: 13.14 vs 12.50, 12.70, and 12.77 in range I) are within plausible sampling noise. No significance tests, confidence intervals, or subject-independent cross-validation are provided, and the 'fewer PCs' claim (e.g., 3 PCs for DMR versus 12 for NET) is also based on one split. The authors should report repeated nested cross-validation that separates subjects between training and test sets and include uncertainty intervals on the metrics.","section":"§Results and discussion (PCR validation, Table 1)"},{"comment":"The statement that DMSR provides 'the least prediction error among the other techniques' is contradicted by Table 1. In range I the lowest RMSE is DMS (12.50), in range II it is DMR (12.77), and in range III it is again DMS (12.80); DMSR is not the minimum in any range. Additionally, DMR achieves comparable or lower RMSE with only 3 principal components, versus 8 for DMSR. The ranking of the methods and the conclusion that DMSR is the most accurate should be revised to match the reported metrics, and any claim that one method is best needs statistical support rather than inspection of a single split.","section":"§Results and discussion and Table 1"}],"minor_comments":[{"comment":"The manuscript states 'III: 4000-5000 cm^-1' in the text, while Table 1 defines range III as '4000-5500 cm^-1'; these definitions should be reconciled, and Figure 11 should be checked against the intended range.","section":"§Results and discussion (wavenumber ranges)"},{"comment":"There are typographical inconsistencies in the method names: the text reads 'for the net, DMS, DMS, and DMSR cases' (DMR is omitted) and later refers to 'DMRS' instead of DMSR; these should be corrected.","section":"§Results and discussion (Figure 9 paragraph)"},{"comment":"The data availability statements are contradictory: the paper says 'This article has no additional data' while also claiming 'The data that support the findings of this study are available within the article,' yet no spectra or code are included or deposited. This prevents reproducibility of the reported RMSE values.","section":"§Data accessibility"},{"comment":"The claim that the standardized DMS spectra are 'less affected' by standardization than the net spectra is qualitative and not quantified; the authors should provide a numerical measure or remove this assertion.","section":"§Results and discussion (Figure 7)"},{"comment":"Reference 22 is cited as 'Appl. Spectrosc. 2024. 0(0)' without a volume, article number, or DOI; the entry is incomplete and should be updated.","section":"References"}],"recommendation":"reject","confidential_remarks":"I concur with the skeptical assessment: the central water-removal claim is not mathematically justified, and the ARA section openly admits that the original method failed before tuning an error term on the evaluation data. The reported PCR improvements are small and lack subject-independent validation. The paper would require a substantial reworking of the method justification, the validation protocol, and the data availability before it could be considered for publication in a serious journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine attempt at a practical problem with a genuinely new preprocessing recipe (DMS/DMR/DMSR). The authors also deserve credit for admitting that their earlier ARA method failed on this dataset and needed an error-compensation term to work at all. That is more honest than most papers.\n\nWhat is actually new: the derivative-multiplication weighting with L2 normalization is not in the cited literature. It is a minor variation on derivative spectroscopy, but it is concrete, simple, and cheap to test on other data. The reported stability gain -- the standard deviation of CV scores drops from 28-104 in the NET case to 2-4 in the derivative cases -- is striking and is the single most interesting empirical result in the paper.\n\nWhere it falls apart: the central claim that multiplying by dA_water/d-nu removes water interference is not justified. The derivative vanishes at a peak maximum, so the 5200 cm-1 peak disappears, but at most other wavenumbers the water absorption is multiplied by a data-dependent weight, not removed. The glucose term gets multiplied by the same weight, so the Lambert-Beer linearity between concentration and processed absorbance is destroyed. The paper calls it effective removal, but it is reweighting, not cancellation, and no simulation or independent validation shows that glucose information is preserved.\n\nThe quantitative support is also thin. RMSE differences are 0.5-0.9 mg/dL from a single 90/10 split with repeated samples from the same subjects. No significance testing is done. The ARA result of 22.4 mg/dL is obtained by tuning R=10^5 to minimize error on the same data, so it is a fitted value rather than a prediction. No data or code are provided, so none of this can be independently checked.\n\nWho this is for: researchers working on NIR glucose calibration who might want to test DMSR on their own data. As a paper, it needs major revision before acceptance: a derivation or simulation for the water-removal mechanism, a subject-independent split or repeated cross-validation, significance tests, and data release. I would send it to peer review rather than desk reject, because the idea is cheap to test and the stability result might be real. But I would expect the referee to demand substantial changes, and I would not accept it in its current form.","headline":"The DMS/DMR/DMSR preprocessing trick is new and the authors are honest about ARA's failure, but the central water-removal claim is asserted rather than shown, and the reported gains are too weak and possibly artifact-driven to accept as-is.","tokens_in":704,"tokens_out":924,"would_cite":false,"duration_ms":36331,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Multiplying blood spectra by the first derivative of water's absorbance suppresses water interference and improves FTIR glucose prediction in this paper's experiments.","keywords":["blood glucose","near-infrared spectroscopy","FTIR spectroscopy","principal component regression","water interference suppression","derivative multiplication","molar absorptivity","Lambert-Beer law"],"falsifier":"Measure glucose-in-water mixtures with known concentrations, apply DMS, DMR, and DMSR exactly as described, then run PCR with the same 90/10 training/test split; if the RMSE advantage over net absorbances disappears, or if the predictions track the derivative weight rather than the true concentration, the claimed water removal is not physical.","tokens_in":9018,"feed_emoji":"🧪","tokens_out":7308,"duration_ms":79195,"temperature":0.7,"pith_summary":"Near-infrared blood spectra are dominated by a strong water absorption band that obscures the glucose signal. This paper proposes multiplying each measured absorbance curve by the first derivative of the pure-water absorbance (DMS), by one minus the first derivative of the glucose absorbance (DMR), or by both (DMSR) before building a principal component regression (PCR) model for glucose. The authors report that the water peak near 5200 $cm^{-1}$ disappears, a glucose-related peak near 5100 $cm^{-1}$ becomes visible, and PCR errors drop slightly while the number of principal components falls from 12 to as few as 3. A separate Lambert-Beer regression using glucose molar absorptivity yields an RMSE of about 22.4 mg/dL. If correct, the method improves FTIR-based glucose estimation without any hardware change.","feed_headline":"Multiplying spectra by water's derivative sharpens glucose predictions","feed_subtitle":"FTIR blood-glucose models get lower RMSE and fewer principal components from a no-hardware preprocessing step.","key_machinery":"The central objects are three preprocessing transforms built from reference derivative spectra. DMS multiplies every sample absorbance by the L2-normalized first derivative of the water absorbance with respect to wavenumber, dA_water/d-nu-tilde; DMR multiplies by the L2-normalized quantity (1 - dA_glucose/d-nu-tilde); DMSR applies both factors. Because the first derivative of a spectrum is zero at a local maximum, the water band at 5200 $cm^{-1}$ is suppressed, and the glucose band near 5100 $cm^{-1}$ is retained or emphasized. These weighted spectra replace net absorbances as inputs to principal component regression, and a separate linear calibration, absorptivity regression analysis (ARA), uses glucose molar absorptivity through the Lambert-Beer relation.","core_discovery":"The central claim is that water interference in blood-glucose NIR analysis can be suppressed by pointwise multiplication of the blood absorbance by the first derivative of the water absorbance spectrum, because a derivative vanishes at a peak maximum and therefore zeros out the water band. Adding a second multiplication by one minus the first derivative of the glucose absorbance reinforces the glucose signature. Compared with the standard net-absorbance approach, which subtracts a scaled water spectrum and depends on an unknown sample-dependent scaling factor gamma, the derivative-weighted spectra are claimed to preserve glucose information and to make principal component regression more stable. In experiments on 200 blood samples from twenty subjects, the derivative-based PCR models achieve slightly lower RMSE values (12.50 to 13.00 mg/dL versus about 13.1 to 13.6 mg/dL for net absorbances) and much smaller cross-validation-score standard deviations, while requiring fewer principal components.","pith_inferences":["Because the derivative weight is fixed by pure water and pure glucose spectra, the same preprocessing could be tested on other instruments or on matrices with a dominant known interferent, provided the weight spectra are measured on the same setup.","If the improvement proves robust, the method could be combined with partial least squares or machine-learning regressors, which the paper does not test.","The mechanism is not fully settled by the experiments: the multiplication may be acting as a band-limiting weight rather than physically removing water absorption, so a synthetic-mixture study with known glucose concentrations would separate those explanations."],"forward_implications":["If the derivative-weighting claim holds, the same FTIR spectra yield glucose predictions with lower RMSE and MAE than the net-absorbance baseline in all tested wavenumber windows.","The derivative-based models become less sensitive to wavenumber range selection, since their cross-validation-score standard deviations are far smaller than those of the net-absorbance models.","DMR needs only three principal components to reach an RMSE of 12.7 mg/dL in the full 4000-8000 cm^-1 range, which would simplify the calibration model.","The ARA Lambert-Beer regression gives an RMSE of about 22.4 mg/dL with most predictions falling in Clarke error grid region A, indicating clinically plausible accuracy for a simple linear model."],"supporting_citations":[{"why":"The authors' prior work that defines the absorptivity regression analysis (ARA) method and identifies the 4000-8000 cm^-1 glucose absorption band; the current ARA implementation extends it.","marker":"22"},{"why":"The water subtraction approach (subtracting a scaled water absorbance) that the paper compares against and whose sample-dependent scaling limitation motivates DMS, DMR, and DMSR.","marker":"28"},{"why":"A published noninvasive NIR glucose assay whose reported errors are used as a reference for calling the ARA RMSE of 22.42 mg/dL acceptable.","marker":"29"},{"why":"An earlier water-interference removal method for in-vivo NIR glucose spectra, cited as context for why water suppression remains hard.","marker":"26"},{"why":"A study of scattering and absorption effects in whole-blood NIR that supports using NIR for blood glucose and motivates handling water absorption.","marker":"9"},{"why":"A comparative PLSR/PCR study used to justify the wavenumber-band investigation and the PCR approach.","marker":"14"}],"fun_headline_variants":["Derivative of water spectrum clears NIR blood glucose signal","Water-derivative preprocessing improves glucose NIR models","Derivative-weighted spectra lower RMSE for glucose","Fewer PCs and smaller errors via derivative weighting","NIR glucose analysis gains from water derivative"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that multiplying each blood absorbance by the first derivative of the water absorbance removes water's contribution at every wavenumber—asserted in the Results from the derivative vanishing at the water peak, but not derived or independently validated—so if that premise fails, the reported improvement is a reweighting artifact.","fun_headline_variants_meta":{"raw":{"variants":["Derivative of water spectrum clears NIR blood glucose signal","Water-derivative preprocessing improves glucose NIR models","Derivative-weighted spectra lower RMSE for glucose","Fewer PCs and smaller errors via derivative weighting","NIR glucose analysis gains from water derivative"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000198,"raw_usage":{"total_tokens":1322,"prompt_tokens":853,"completion_tokens":469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":397}},"tokens_in":469,"tokens_out":469,"duration_ms":5781,"temperature":1.0,"reasoning_tokens":397,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:53:35.111520+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure glucose-in-water mixtures with known concentrations, apply DMS, DMR, and DMSR exactly as described, then run PCR with the same 90/10 training/test split; if the RMSE advantage over net absorbances disappears, or if the predictions track the derivative weight rather than the true concentration, the claimed water removal is not physical.","supporting_citations":[{"cited_title":"Principal Component Analysis and Near-Infrared Spectroscopy as Noninvasive Blood Glucose Assay Methods","cited_arxiv_id":null,"evidence_quote":"The authors' prior work that defines the absorptivity regression analysis (ARA) method and identifies the 4000-8000 cm^-1 glucose absorption band; the current ARA implementation extends it."},{"cited_title":"Quantitative Infrared Spectroscopic Analysis","cited_arxiv_id":null,"evidence_quote":"The water subtraction approach (subtracting a scaled water absorbance) that the paper compares against and whose sample-dependent scaling limitation motivates DMS, DMR, and DMSR."},{"cited_title":"Noninvasive Blood Glucose Assay Using a Newly Developed Near-Infrared System","cited_arxiv_id":null,"evidence_quote":"A published noninvasive NIR glucose assay whose reported errors are used as a reference for calling the ARA RMSE of 22.42 mg/dL acceptable."},{"cited_title":"Removal of Interference Signals Due to Water from In Vivo Near -Infrared (NIR) Spectra of Blood Gluco se by Region Orthogonal Signal Correction (ROSC)","cited_arxiv_id":null,"evidence_quote":"An earlier water-interference removal method for in-vivo NIR glucose spectra, cited as context for why water suppression remains hard."},{"cited_title":"Scattering and absorption effects in the determination of glucose in whole blood by near-infrared spectroscopy","cited_arxiv_id":null,"evidence_quote":"A study of scattering and absorption effects in whole-blood NIR that supports using NIR for blood glucose and motivates handling water absorption."},{"cited_title":"Investigation of goodness of model data fit using PLSR and PCR regression models to determine informative wavelength band in NIR region for non -invasive blood glucose prediction","cited_arxiv_id":null,"evidence_quote":"A comparative PLSR/PCR study used to justify the wavenumber-band investigation and the PCR approach."}],"review_version":1}