{"id":"583ca930-eee8-44e5-97cc-7e980455d81e","arxiv_id":"2412.07821","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"Threshold-based and adaptive derivative preprocessing of MIR spectra improves Ridge and SVR glucose estimates in a 46-sample study, but the improvement is not independently validated.","lead":"This paper proposes two preprocessing tricks, TBD and ADPD, that combine raw infrared absorbance data with its derivative to help machine learning models estimate blood glucose from MIR spectra. The authors report improved R2 and lower errors, but the dataset is small and no code or data is released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Selection leakage in LOOCV undermines the claimed TBD/ADPD gains; a nested CV check is required to establish unbiased generalization.","rationale":"The reader's weakest assumption is precisely the issue: hyperparameters are chosen using the same LOOCV procedure that produces the reported performance metrics. The manuscript offers no nested CV or held-out set, and the selection ranges listed in Table 1 are broad enough to overfit 46 samples. Because the central claim rests entirely on these LOOCV numbers, the evidence is currently insufficient to establish that TBD and ADPD significantly outperform the baselines. A nested LOOCV experiment is the minimal, decisive computational check. Other issues—small sample size, missing TBD equation, no released code or data, and error-grid comparisons based on one or two zone shifts—reinforce the same conclusion but do not change it. Since the reader already reached CONDITIONAL based on this weakness, no verdict adjustment is needed.","tokens_in":8074,"tokens_out":2351,"duration_ms":21905,"concrete_test":"Run (or require the authors to run) a nested LOOCV: for each of the 46 outer folds, perform an inner LOOCV on the 45 training samples to select tau, alpha, Ridge alpha, SVR C/kernel/epsilon, and PCA n_components; then evaluate the selected configuration only on the outer held-out sample. Compare the resulting MSE/MAE/R2 for Base, Derivative, TBD, and ADPD against Table 2. If the TBD/ADPD advantage shrinks or reverses, the reported improvements are selection artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is that the reported LOOCV metrics are not unbiased estimates of generalization, because model selection shares the same 46-sample LOOCV loop used for evaluation. Section 3 states 'Using LLOC, I identified the hyperparameters for each model,' and the text describes tuning PCA components from 1 to 20 'within the hyperparameter tuning process.' Thus the TBD threshold tau, the ADPD alpha, the Ridge alpha, the SVR C/kernel/epsilon, and the PCA component count are all selected using the same folds whose held-out predictions then generate Table 2. With n=46 and no independent test set, this creates a direct selection-leakage path: configurations that fit sample-specific idiosyncrasies will look better in LOOCV than they would on new spectra. The claimed 24-36% R2 improvements could therefore be wholly or partly artifacts of this overlap. The manuscript contains no nested cross-validation, no significance tests, and no external validation, and the exact TBD transformation equation is missing from the text, so the result cannot currently be independently reproduced or checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two preprocessing methods, Threshold-Based Derivative (TBD) and Adaptive Derivative Peak Detection (ADPD), that combine mid-infrared absorbance spectra with their first derivatives to improve blood glucose concentration estimation. The methods are evaluated with Ridge Regression and Support Vector Regression under leave-one-out cross-validation (LOOCV) on 46 FTIR spectra from five healthy subjects. The authors report that TBD and ADPD outperform raw absorbance and simple derivative baselines, with lower MSE/MAE, higher R2, and better Clarke and Parkes Error Grid zones, and they argue these gains are clinically meaningful.","tokens_in":8263,"tokens_out":3358,"duration_ms":29691,"significance":"If validated, the two preprocessing ideas are simple and potentially useful additions to the chemometric toolbox for infrared-based glucose sensing. The study includes a straightforward comparison of four feature representations and two regression models, and the use of Clarke and Parkes Error Grids is appropriate for clinical relevance. However, the current evidence is not conclusive: the study uses 46 samples with no independent test set, the hyperparameters are selected with the same LOOCV folds used for evaluation, and the central TBD equation is missing. The reported gains may therefore be artifacts of selection leakage. The paper does not supply code, data, or machine-checked proofs, so reproducibility currently rests entirely on the textual description, which is incomplete.","major_comments":[{"comment":"The reported LOOCV metrics are not unbiased estimates of generalization because model selection and evaluation share the same folds. Section 3 states \"Using LLOC, I identified the hyperparameters for each model,\" and the text says the PCA component count was adjusted from 1 to 20 \"within the hyperparameter tuning process.\" Thus the TBD threshold τ, ADPD α, PCA dimension, Ridge α, and SVR C/kernel/ε are all chosen using the same 46 samples whose held-out predictions generate Table 2. This selection leakage can inflate R2 by fitting sample-specific idiosyncrasies. A nested cross-validation or a fixed, independent test set is required to support the claimed 24–36% improvements.","section":"Section 3, Tables 1-2"},{"comment":"The TBD transformation equation is missing. After the sentence \"Mathematically, this can be represented as follows:\" the equation is absent, leaving only the variable definitions. Without the explicit formula for the threshold-based combination of A and dA/dν, the method cannot be reproduced or independently verified. Please provide the full equation, including the exact condition and output value for both branches.","section":"Section 2.4"},{"comment":"The sample-exclusion rule is under-specified. The text says four samples were eliminated \"since they did not fit the requirements for inclusion,\" but no criteria are given. With only 50 samples originally and 46 used in all analyses, the exclusion rule can materially affect the conclusions. The authors should state the inclusion/exclusion criteria and, ideally, show that the main results are robust to the exclusion decision.","section":"Section 2.1"},{"comment":"The clinical-accuracy claim is overstated relative to the evidence. The differences in Table 3 are at most one point in any error-grid zone (and some cells show the baseline method performing equally or better), yet the text repeatedly says TBD and ADPD \"significantly outperform\" conventional methods. No significance test or confidence interval is provided for these counts. The authors should either tone down the language or supply a statistical analysis (e.g., a bootstrap or McNemar-style test on zone membership).","section":"Section 3, Table 3"}],"minor_comments":[{"comment":"The text says the models were evaluated using \"Root Mean Square Error (MSE)\" and then lists MSE and MAE. The acronym RMSЕ is not defined, and it is unclear whether Table 2 reports MSE or RMSE, since the equations and captions are inconsistent.","section":"Section 2.6"},{"comment":"There are several typos and inconsistencies: \"LLOC\" should be \"LOOCV\", \"EMSE\" should be \"MSE\", \"Absolye\" should be \"Absolute\", and \"Tabel\" should be \"Table\". These should be corrected throughout.","section":"Section 3"},{"comment":"The ADPD formula y = x − αz ∙ x is ambiguous because of operator precedence. Please write it as y = x − α · (z · x) or y = x · (1 − αz), with a clear statement of which form is intended.","section":"Section 2.5"},{"comment":"The phrase \"these Ridge Regression values were between 36% and 24%\" is unclear. It would be better to report the actual R2 values and the absolute or relative improvements with a clear notation, e.g., \"increasing R2 from 0.44 to 0.60 (a 36% relative increase)\".","section":"Abstract"},{"comment":"The statement \"This ensures that no information from the valid data leaked into the training process\" appears to contain a typo: \"valid\" should be \"validation\" or \"test\" data.","section":"Section 2.6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as an early draft. The missing TBD equation alone would require a revision, and the selection-leakage issue is fundamental to the empirical claims. With a nested CV or a properly split test set, and a complete methods section, the paper could be salvageable. I would not recommend acceptance in the current form; the authors should be asked to provide code or a detailed algorithmic description for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this paper proposes two derivative-based preprocessing rules, TBD and ADPD, for blood glucose estimation from MIR spectra, and reports that they lift R2 from about 0.44–0.47 to 0.6 on a 46-sample dataset. The idea is simple but not silly, and the gains are exactly the kind that would matter if real. My concern, which matches your stress-test, is that the numbers are likely inflated by selection leakage.\n\nWhat's genuinely new: the specific combination of a thresholded derivative (TBD) and an adaptive derivative weighting (ADPD) is not something I've seen in the glucose-MIR literature. The authors also do several things right: they compare against raw absorbance and plain derivative baselines, they report both regression models, they provide violin plots and error grids, and they explicitly concede the dataset is small. The preprocessing pipeline is described clearly enough to reimplement, aside from the missing TBD equation—the mathematical expression is simply absent from Section 2.4.\n\nThe soft spot is load-bearing. Section 3 states that hyperparameters for every model and preprocessing method were identified using LOOCV, and the PCA component count was tuned from 1 to 20 'within the hyperparameter tuning process.' That means the same folds that produce the Table 2 metrics were used to pick τ, α, ridge penalty, SVR C/kernel/ε, and the number of components. With 46 points, no nested CV, and no held-out set, the R2 gains from TBD and ADPD are fitted values, not unbiased predictions. Missing code and data mean we can't check whether the improvement survives an honest protocol. The clinical claims also rest on one-point differences in the error grids, and the exclusion of four samples is vaguely justified.\n\nNone of this proves the preprocessing ideas are bad; it means the paper as written doesn't demonstrate they're good. The fix is straightforward: nested CV or a proper train/test split, release code and data, report confidence intervals or significance tests, and validate on an external cohort. If the TBD gain holds under those conditions, it's a useful calibration trick. If not, it joins a long list of preprocessing variants that look good on the training set.\n\nWho should read it: anyone working on spectroscopic calibration who wants a concrete example of how easy it is to fool LOOCV when tuning and evaluation share the same loop. It's not a strong empirical contribution as it stands, but the method deserves a serious referee if resubmitted with rigorous validation.\n\nMy recommendation: I would engage with it enough to run the test on my own data, but I wouldn't cite the performance numbers until the leakage is resolved.","headline":"A plausible derivative-based preprocessing idea whose reported R2 gains are undermined by selection leakage in the shared LOOCV loop; worth testing but needs a rigorous validation redesign.","tokens_in":8816,"tokens_out":2731,"would_cite":false,"duration_ms":26099,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Selectively fusing FTIR absorbance with its first derivative raises the coefficient of determination of blood-glucose regression from 0.44 to 0.596 in ridge regression and from 0.47 to 0.60 in SVR.","keywords":["blood glucose estimation","mid-infrared spectroscopy","Fourier transform infrared spectroscopy","derivative spectroscopy","threshold-based feature selection","support vector regression","ridge regression","principal component analysis"],"falsifier":"Re-run the whole procedure so that, for each held-out sample, the threshold, $\\alpha$, PCA component count, and model hyperparameters are chosen using only the other samples, and the held-out sample is never used for tuning; if the $R^2$ gains over the base and derivative methods shrink to near zero, the reported improvements are artifacts of tuning on the scored samples.","tokens_in":7850,"feed_emoji":"🩸","tokens_out":9375,"duration_ms":78550,"temperature":0.7,"pith_summary":"Blood glucose is hard to read from mid-infrared spectra because glucose absorbs weakly and the informative peaks sit in a noisy, sloping baseline. This paper proposes two preprocessing methods—Threshold-Based Derivative (TBD) and Adaptive Derivative Peak Detection (ADPD)—that merge the absorbance spectrum with its first derivative, keeping the original peak information while adding derivative contrast only where it helps. On 46 FTIR samples from five healthy subjects, the authors report that TBD raises the coefficient of determination $R^2$ from 0.44 to 0.596 for ridge regression and from 0.47 to 0.60 for support vector regression, with lower mean squared and mean absolute errors, and that ADPD produces smaller improvements. They also report that every prediction falls inside clinically acceptable zones of the Clarke and Parkes error grids. If those numbers survive an honest out-of-sample test, TBD is a cheap preprocessing step toward making MIR spectroscopy practical for non-invasive glucose monitoring.","feed_headline":"Derivative peaks lift blood glucose prediction by 36%","feed_subtitle":"Mixing FTIR absorbance with its first derivative improves glucose estimates in Ridge and SVR models.","key_machinery":"TBD is a per-wavenumber switch: for each point $i$, if $|100 \\cdot A_i'| < \\tau$ the output keeps the absorbance $A_i$, and otherwise it uses $100 \\cdot A_i'$, with threshold $\\tau$ tuned in $[0.02, 0.3]$. ADPD forms $y = x - \\alpha\\, z \\cdot x$, where $z$ is the first derivative of $x$ with respect to wavenumber and $\\alpha$ is tuned in $[0,70]$, so the raw signal is cut back most where the derivative is large. Both operations run after rubber-band baseline correction, min-max scaling, and Savitzky-Golay smoothing, and the resulting spectra are reduced with principal component analysis before entering ridge or SVR regression.","core_discovery":"On its own terms, the paper's central claim is that derivative preprocessing works best when it is selective: instead of replacing the absorbance spectrum with its derivative everywhere, the methods keep the original absorbance where the scaled derivative is small and substitute the derivative where it is large (TBD), or downweight the absorbance by an adaptive multiple of the derivative (ADPD). In the reported leave-one-out cross-validation, TBD is the strongest variant: for ridge regression the MSE falls from 75.9 to 55.2 and $R^2$ rises from 0.44 to 0.596, and for SVR the MSE falls from 72.24 to 54.53 and $R^2$ rises from 0.47 to 0.60, corresponding to relative $R^2$ gains of roughly 36% and 27%. The authors interpret this as the methods balancing the molecular-composition information carried by absorption peaks against the resolving power of the first derivative.","pith_inferences":["If the gains hold under nested cross-validation, the same threshold-switch idea could be tried on near-infrared spectra or on urine and tissue FTIR, where glucose peaks are similarly weak.","The TBD switch is a hard binary feature selector; comparing it with a soft-threshold version would show whether the improvement comes from the binary switch or from the derivative scaling itself.","Because all settings were tuned on the same 46 samples that were then scored, the ranking of TBD over ADPD could shrink, widen, or reverse on a larger cohort with separate training, validation, and test partitions."],"forward_implications":["TBD on SVR yields $R^2 = 0.60$ with MSE 54.53 and MAE 5.63, the best reported numbers in the paper; TBD on ridge yields $R^2 = 0.596$ with MSE 55.2.","ADPD also outperforms both the raw absorbance and the pure derivative baselines for both models, but by a smaller margin than TBD.","In the Clarke error grid, TBD leaves both models with zero points outside zone A; in the Parkes type-1 grid it also reaches zero, while in the Parkes type-2 grid it ties the baseline.","Because the two methods act as a preprocessing stage before PCA and the regression models, the same features can be paired with other estimators without changing the claimed benefit."],"supporting_citations":[{"why":"Supplies the idea that derivative spectral treatment sharpens overlapping spectral features.","marker":"[13]"},{"why":"Shows first and second derivatives reduce error in FTIR multivariate calibration models.","marker":"[14]"},{"why":"Establishes that FTIR peaks carry molecular-composition information, motivating peak preservation.","marker":"[15]"},{"why":"Supports that spectral peaks are crucial for glucose estimation, motivating the peak-retaining fusion.","marker":"[16]"},{"why":"Provides the Savitzky-Golay filter used to smooth spectra before derivative computation.","marker":"[11]"},{"why":"Provides a comparison set of machine-learning models and PCA-based feature selection for glucose prediction.","marker":"[6]"},{"why":"Supplies the Clarke error grid used to judge clinical accuracy of predictions.","marker":"[27]"},{"why":"Supplies the Parkes error grid used to judge clinical accuracy of predictions.","marker":"[28]"}],"fun_headline_variants":["Derivative math boosts glucose prediction accuracy by 36%","Selective derivative peaks lift glucose estimates by 36%","Ridge regression plus derivative sharpens glucose readings","Derivative-based MIR spectroscopy improves glucose R2 by 36%","Hybrid absorbance-derivative model elevates glucose prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the reported leave-one-out scores honestly measure how well the method will predict on new people, even though the same 46 blood samples were used to choose the method settings and then to produce those scores.","fun_headline_variants_meta":{"raw":{"variants":["Derivative math boosts glucose prediction accuracy by 36%","Selective derivative peaks lift glucose estimates by 36%","Ridge regression plus derivative sharpens glucose readings","Derivative-based MIR spectroscopy improves glucose R2 by 36%","Hybrid absorbance-derivative model elevates glucose prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000551,"raw_usage":{"total_tokens":2612,"prompt_tokens":912,"completion_tokens":1700,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":1619}},"tokens_in":528,"tokens_out":1700,"duration_ms":11746,"temperature":1.0,"reasoning_tokens":1619,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:38:38.019150+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the whole procedure so that, for each held-out sample, the threshold, $\\alpha$, PCA component count, and model hyperparameters are chosen using only the other samples, and the held-out sample is never used for tuning; if the $R^2$ gains over the base and derivative methods shrink to near zero, the reported improvements are artifacts of tuning on the scored samples.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the idea that derivative spectral treatment sharpens overlapping spectral features."},{"cited_title":"Detection of orange essential oil, isopropyl myristate, and benzyl alcohol in lemon essential oil by ftir spectroscopy combined with chemometrics","cited_arxiv_id":null,"evidence_quote":"Shows first and second derivatives reduce error in FTIR multivariate calibration models."},{"cited_title":"Garcia Martinez and Ihtesham Ur Rehman","cited_arxiv_id":null,"evidence_quote":"Establishes that FTIR peaks carry molecular-composition information, motivating peak preservation."},{"cited_title":"A single wavelen gth mid -infrared photoacoustic spectroscopy for noninvasive glucose detection using machine learning","cited_arxiv_id":null,"evidence_quote":"Supports that spectral peaks are crucial for glucose estimation, motivating the peak-retaining fusion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Savitzky-Golay filter used to smooth spectra before derivative computation."},{"cited_title":"Accurate prediction of glucose concentration and identi fication of major contributing features from hardly distinguishable near-infrared spectroscopy","cited_arxiv_id":null,"evidence_quote":"Provides a comparison set of machine-learning models and PCA-based feature selection for glucose prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Clarke error grid used to judge clinical accuracy of predictions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Parkes error grid used to judge clinical accuracy of predictions."}],"review_version":1}