{"id":"b72e95b9-6ce6-4685-b592-879b2e10a81e","arxiv_id":"2504.15324","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A graph-based spectral filtering pipeline using PageRank centrality and forward feature selection raises exosome classification accuracy to 0.76 for spectra and 0.857 for surface-averaged data under group 10-fold cross-validation.","lead":"This paper describes a new way to clean up Raman spectroscopy data before using machine learning to classify exosomes, which are tiny particles in blood and other fluids. The method uses a graph-based ranking technique to remove noisy measurements and select the most informative spectral features, improving classification accuracy for samples from normal, hyperglycemic, and hypoglycemic cells.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be inflated because tuning decisions are not demonstrably nested inside the cross-validation; the paper's own conclusion flags nested cross-validation as future work.","rationale":"The paper is read in good faith: the group k-fold split is the right general idea, the graph construction is described, and the authors acknowledge in the conclusion that nested cross-validation is needed. The concern is not that the method is impossible or that the authors are hiding results, but that the central comparison could be an artifact of tuning on evaluation data. The reader's verdict of CONDITIONAL matches this: the method is plausible, but the evidence is not yet sufficient. I agree with the reader's weakest_assumption; the systematically tested cutoffs in Section 3.3 and the feature selection in Section 3.4 are the exact places where leakage can enter. The proposed concrete test would settle the concern by forcing all tuning inside the training folds and reporting variability across folds. If the accuracy advantage persists under nested CV, the core claim is supported. If it disappears, the paper needs to be reframed as a preprocessing proposal with weaker empirical support. No ad hominem is intended; the issue is protocol transparency and experimental design. Verdict remains UNCHANGED because the reader already assigned CONDITIONAL, which is the appropriate disposition.","tokens_in":6992,"tokens_out":4177,"duration_ms":38644,"concrete_test":"Implement a nested group cross-validation protocol. For each outer group-10-fold split, use inner group CV on the training groups only to choose the IQR proportion cutoff, PageRank centrality threshold, Gaussian kernel scale, and number of forward-selected wavenumbers; then train the final model on the full training groups with those chosen values and evaluate on the outer test groups. Report the selected parameter values for every fold and the mean plus/minus standard deviation of accuracy over repeated outer splits. If the resulting OSC+PRF+DR accuracy falls below 0.70 for spectra or is within roughly 0.03 of OSC, the claimed advantage of the graph-based filter is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the OSC+PRF+DR pipeline improves accuracy because graph-based PageRank filtering preserves biomarker signal. The evidence in Tables 1 and 2 depends on an evaluation protocol that is never fully specified. Section 3.3 says the IQR proportion cutoff was 'systematically tested' and a PageRank centrality cutoff was applied; Section 3.4 says a forward sequential feature selection chose the 50 most influential wavenumbers. The manuscript does not state that these cutoffs, the Gaussian kernel width, or the feature count were selected inside each training fold of the group 10-fold cross-validation. If they were selected using all groups, information from held-out groups entered the model construction, and the reported improvement over OSC would be optimistically biased, since OSC does not undergo comparable tuning. The conclusion explicitly lists nested cross-validation as future work, which is an in-text admission that the current evaluation is not nested. With only 63 SERS surfaces in the group splits, the accuracy differences (0.760 vs 0.635 for spectra, 0.857 vs 0.683 for surfaces) are also reported without standard deviations or repeated CV, so it is unclear whether they exceed sampling noise. The load-bearing condition, that all tuning decisions are made inside training folds, is not stated and the paper's language suggests the opposite.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-stage Raman spectral processing pipeline, OSC+PRF+DR, for classifying exosomes from normal, hyperglycemic, and hypoglycemic cells. OSC is a standard cleaning sequence (despiking, baseline correction, smoothing, scaling); PRF adds graph-based PageRank filtering of spectra per SERS surface, after constructing Gaussian-kernel similarity graphs in Neo4j; DR applies forward sequential feature selection to retain 50 wavenumbers. The authors evaluate with group 10-fold cross-validation on 3,045 spectra from 63 SERS surfaces, using SVM, Random Forest, and Extra Trees. They report accuracies of 0.760 for spectra and 0.857 for surfaces with OSC+PRF+DR, compared with 0.635 and 0.683 for the OSC baseline, and claim that graph-based filtering plus dimensionality reduction significantly improves classification.","tokens_in":7259,"tokens_out":2955,"duration_ms":27320,"significance":"If the reported gains are real, the combination of graph-centrality-based outlier removal and feature selection would be a useful addition to Raman spectral preprocessing, and the use of group cross-validation is methodologically commendable. The paper also introduces a publicly relevant biomedical task and presents a clearly structured comparison of three processing pipelines. The main weakness is that the evaluation protocol does not demonstrate that all tuning decisions are nested inside the cross-validation folds; the paper's own conclusion lists nested cross-validation as future work. Because the reported accuracy improvements are the paper's central claim, this is a load-bearing issue. The absence of any variance estimates or repeated cross-validation further weakens the claim that the differences are significant. The core idea is plausible, but the current evidence does not yet establish superiority over the OSC baseline.","major_comments":[{"comment":"The tuning decisions—the IQR proportion cutoff, the PageRank centrality threshold, the Gaussian kernel width, and the number of selected wavenumbers—are not stated to be chosen inside each training fold of the group 10-fold cross-validation. The text in Section 3.3 says the IQR proportion cutoff was 'systematically tested' and a PageRank cutoff was applied, and Section 3.4 says features were selected by forward sequential selection 'iteratively selecting the wave numbers that contributed most to model performance.' The conclusion explicitly lists nested cross-validation as future work, which is an in-text admission that the current evaluation is not nested. If any of these choices used held-out groups, the reported accuracies are optimistically biased, and the comparison against OSC (which receives no comparable tuning) is unfair. Please either demonstrate that every tuning step was performed inside the training folds, or re-run the evaluation with fully nested cross-validation and report the resulting accuracies.","section":"Section 3.3, 3.4, 5"},{"comment":"The reported accuracies are point estimates with no standard deviations, confidence intervals, or repeated cross-validation runs. With only 63 unique SERS surfaces split into 10 groups, the observed differences (0.760 vs 0.635 for spectra; 0.857 vs 0.683 for surfaces) may be within sampling noise. The conclusion describes OSC+PRF+DR as 'significantly better' and 'significantly improved,' but no statistical test or variance estimate supports this wording. Please report per-fold results, repeated group-10-fold cross-validation with mean and standard deviation, and, ideally, a paired significance test across folds (or another appropriate procedure).","section":"Tables 1 and 2"},{"comment":"The PageRank filtering procedure is underspecified. It is unclear whether the IQR proportion cutoff and the PageRank centrality cutoff are global constants applied identically to every surface, or are adaptively chosen per surface. The Gaussian kernel width used to build the similarity subgraphs is never reported, although it determines which spectra are considered close enough to share an edge. The choice of the cutoff values affects how many spectra survive filtering and, consequently, all downstream accuracy numbers. Please give the exact parameter values used in the final experiments, the range tested, and the criterion used to pick the final values, and explain whether these choices were made per surface or globally.","section":"Section 3.3"},{"comment":"The forward sequential feature selection is not described precisely enough to assess whether it leaks information. The text says wave numbers are selected 'iteratively' based on 'model performance,' but it does not state what performance measure is used, what inner validation procedure (if any) is employed, or how the 50-feature limit interacts with the group cross-validation. If the feature selection is run on the entire dataset before the outer cross-validation loop, the chosen wavenumbers carry information from the held-out groups, biasing the reported accuracy. Feature selection must be embedded inside each training fold (or in a nested inner loop), and the selected wavenumbers should be reported for at least one fold to assess stability.","section":"Section 3.4"}],"minor_comments":[{"comment":"The caption contains a typo: 'Extra Tress model' should read 'Extra Trees model.'","section":"Figure 5 caption"},{"comment":"The caption says 'after OSC proccesing'; 'proccesing' should be 'processing.'","section":"Figure 2 caption"},{"comment":"The sentence 'To minimize high-frequency noise, the Savitzky-Golay filter, a robust digital signal processing technique widely used in spectral analysis to preserve signal characteristics while reducing noise [9].' is grammatically incomplete; a main verb is missing.","section":"Section 3.2"},{"comment":"The 50 selected wavenumbers are never listed or characterized, making it difficult to judge biological plausibility or reproducibility. Please provide the selected wavenumber indices or a representative list, and discuss whether the selection is stable across folds.","section":"Section 3.4 and Tables 1-2"},{"comment":"The 'Best Model' row of Tables 1 and 2 reports only the best-performing classifier, but the manuscript does not give the hyperparameters used for SVM, Random Forest, or Extra Trees. Without this information, the comparison across pipelines cannot be fully reproduced.","section":"Section 4.2"},{"comment":"The benefit of using Neo4j as a graph database, as opposed to constructing the same graphs in memory, is not explained. If the graph database is incidental to the method, please clarify this; otherwise, describe what graph features or computational advantages it provides.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's own conclusion states that nested cross-validation is future work, which is effectively an admission that the current evaluation does not keep tuning decisions inside the folds. Requiring a fully nested evaluation and variance estimates would address the central weakness. The idea is publishable in principle if the revised evaluation supports the claimed gains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nPunchline: this is a genuinely novel preprocessing combination for Raman-based exosome classification, but the headline accuracy numbers are not well-supported as reported. The gain over OSC could easily be tuning bias rather than a real effect.\n\nWhat the paper actually does: after standard spectral cleaning (OSC), it builds a graph where spectral peaks are nodes, with edges connecting peaks from the same spectrum or located close in wavenumber. PageRank centrality selects representative spectra, then a forward feature selector reduces the set to 50 wavenumbers. That specific sequence—PageRank-based spectral filtering plus sequential feature selection—is not in the prior Raman/exosome literature, as far as I know. Using group 10-fold cross-validation with 63 SERS surfaces is a solid instinct and addresses the most common leakage problem in spectral classification.\n\nThe soft spots are real and non-trivial. The evaluation protocol is underspecified. Section 3.3 says the IQR proportion cutoff was \"systematically tested\" and a PageRank centrality threshold was applied; Section 3.4 says forward feature selection chose the 50 most influential wavenumbers. Nothing in the paper states these are selected inside each training fold of the group 10-fold CV. If they were selected on the full dataset, information from held-out groups leaks into model construction, and the reported 0.760/0.857 accuracies over OSC's 0.635/0.683 are optimistically biased. The authors' conclusion lists nested cross-validation as future work, which is effectively an admission that the current evaluation is not nested. Also, with only 63 surfaces, the accuracy differences have no error bars or repeated CV, so we don't know if they exceed sampling noise.\n\nMinor issues: the limit of 50 features is arbitrary, and the smoothing/normalization choices appear to be evaluated on the same data. No code or data is provided, which makes reproduction harder.\n\nNet read: the method is worth a serious referee, but the manuscript needs major revision. The authors should specify exactly which parameters are tuned inside the CV loop, run nested CV or an inner validation loop, and report variability across folds. If the gain survives that, the contribution is a modest but useful preprocessing technique for liquid biopsy Raman analysis.\n\nFor your own work: not a must-cite, but worth watching. I would send it for peer review.","headline":"Plausible new Raman preprocessing combination, but the reported accuracy gains may be inflated because key tuning decisions are not demonstrably nested inside the cross-validation.","tokens_in":7753,"tokens_out":2474,"would_cite":false,"duration_ms":22966,"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":"A graph-based PageRank filter plus forward feature selection lifts Raman exosome classification accuracy to 0.760 on spectra and 0.857 on surfaces under group cross-validation.","keywords":["Exosome classification","Raman spectroscopy","Surface-enhanced Raman spectroscopy","PageRank filter","Graph database","Spectral preprocessing","Forward feature selection","Group cross-validation"],"falsifier":"Repeat the full evaluation with nested group 10-fold cross-validation, re-selecting the interquartile-range cutoff, the PageRank threshold, the Gaussian kernel width, and the 50 wave numbers inside each training fold and applying them only to the held-out group. If held-out accuracy drops to the OSC-only baseline (0.635 spectra, 0.683 surfaces), the reported gains are an artifact of tuning on the evaluation groups.","tokens_in":6835,"feed_emoji":"🔬","tokens_out":10614,"duration_ms":89817,"temperature":0.7,"pith_summary":"The paper claims that standard Raman preprocessing leaves noisy, unrepresentative spectra in exosome datasets, and that this noise can be removed with a graph-based filter: each spectrum is placed in a graph built from Raman peaks, scored by PageRank centrality, and low-scoring spectra are discarded. Adding a forward sequential feature selection that keeps the 50 most informative wave numbers then gives a compact feature set that classifies exosomes from normal, hyperglycemic, and hypoglycemic cells at 0.760 accuracy on Raman spectra and 0.857 accuracy on Raman surfaces under group 10-fold cross-validation. The practical interest is that exosomes circulate in blood, saliva, and urine, so a reliable Raman-plus-graph pipeline would support liquid-biopsy diagnostics without invasive tissue sampling. For the reported numbers to be unbiased, every tuning choice in the pipeline must be made inside each training fold.","feed_headline":"Graph filter plus feature selection hits 0.857 on exosome Raman","feed_subtitle":"A PageRank-based cleaning step followed by 50 selected wave numbers separates normal, hyperglycemic, and hypoglycemic exosomes.","key_machinery":"The load-bearing mechanism is the PageRank Filter (PRF) over a peak-based graph. Each Raman peak is a node; edges connect peaks that come from the same spectrum or that lie close in wave number, so the graph encodes co-occurrence and spectral proximity. PageRank centrality—a score that ranks nodes by their connectedness in the graph—is computed for each spectrum, and spectra below a centrality cutoff are dropped as noise. The dimensionality-reduction stage then runs a forward sequential search over wave numbers, keeping the 50 that most improve model performance. Extra Trees is the classifier that best exploits this reduced representation in the paper's experiments.","core_discovery":"The authors' central discovery is that a three-stage pipeline—standard cleaning (cosmic-spike removal, baseline correction, smoothing, scaling), a PageRank Filter that removes spectra with low graph centrality, and an optimal dimensionality-reduction step that forward-selects 50 wave numbers—classifies exosome samples substantially better than standard cleaning alone. On the raw Raman spectra, the pipeline reaches 0.760 accuracy with an Extra Trees model versus 0.635 for cleaning alone; on the SERS surfaces, it reaches 0.857 versus 0.683. The authors attribute the gain to the graph filter suppressing unrepresentative spectra while the feature selection keeps biomarker-relevant peaks, and they report that surface-level Raman data are consistently easier to classify than single-spectrum data.","pith_inferences":["The decisive test the paper leaves undone is nested group cross-validation: if the interquartile-range cutoff, PageRank threshold, kernel width, and 50 selected wave numbers were tuned on held-out groups, reported accuracies would be optimistic; a nested re-run would settle whether the gains persist.","The graph encodes more information than the filter uses: node embeddings or centrality features could be fed directly to the classifier, so the paper's pipeline likely under-exploits the graph structure it builds.","The 50 selected wave numbers are a testable biomarker claim: checking them against known exosome Raman bands (amide I/III, CH2, phosphate backbone) would show whether the pipeline preserves molecular signal or merely separates surface preparation batches.","The gap between surface-level and spectrum-level accuracy suggests the SERS surface itself carries predictive information; distinguishing a true exosome signal from a batch effect would require surfaces prepared from independent biological replicates."],"forward_implications":["The complete OSC+PRF+DR pipeline improves accuracy from 0.635 to 0.760 on spectra and from 0.683 to 0.857 on surfaces under group 10-fold cross-validation.","The PageRank filter alone accounts for part of the gain (0.696 on spectra, 0.825 on surfaces), so graph-based outlier removal helps even before feature selection is added.","Surface-level Raman data classify more cleanly than single spectra under every preprocessing variant, including perfect recall for the hyperglycemic class and perfect precision for the normal class under the full pipeline.","Because group 10-fold cross-validation never splits spectra from the same surface across training and test sets, the reported accuracies are estimates for classifying previously unseen experimental surfaces rather than new spectra from known surfaces."],"supporting_citations":[{"why":"Supplies PageRank centrality, the scoring rule used to rank and filter spectra in the PageRank Filter.","marker":"[3]"},{"why":"Defines Raman spectroscopy and the peak-to-composition mapping that the graph edges are built from.","marker":"[7]"},{"why":"Supplies the asymmetric least-squares baseline correction that is part of the OSC cleaning chain.","marker":"[8]"},{"why":"Supplies the smoothing filter used in the OSC cleaning chain.","marker":"[9]"},{"why":"Supplies the group-k-fold protocol that the paper's headline accuracies rely on to prevent surface-level data leakage.","marker":"[12]"},{"why":"Documents the spectral acquisition and calibration protocol that produced the 3,045 spectra used in the evaluation.","marker":"[20]"},{"why":"Motivates the dimensionality-reduction stage by linking high-dimensional Raman feature spaces to degraded classifier performance.","marker":"[23]"},{"why":"Supplies the cosmic-spike removal algorithm used at the start of the OSC cleaning chain.","marker":"[29]"}],"fun_headline_variants":["PageRank-filtered Raman peaks classify exosomes at 0.857","Graph centrality cleans Raman spectra to 0.857 exosome accuracy","Neo4j + PageRank sharpen exosome Raman classification","Graph filter lifts exosome Raman accuracy to 0.857"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracies assume that every tuning decision—the interquartile-range outlier cutoff, the PageRank centrality threshold, the Gaussian kernel width, and the 50 selected wave numbers—was made inside each training fold of the group 10-fold cross-validation, so that no held-out information influenced the final model.","fun_headline_variants_meta":{"raw":{"variants":["PageRank-filtered Raman peaks classify exosomes at 0.857","Graph centrality cleans Raman spectra to 0.857 exosome accuracy","Neo4j + PageRank sharpen exosome Raman classification","Graph filter lifts exosome Raman accuracy to 0.857"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000478,"raw_usage":{"total_tokens":2347,"prompt_tokens":904,"completion_tokens":1443,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":1369}},"tokens_in":520,"tokens_out":1443,"duration_ms":9762,"temperature":1.0,"reasoning_tokens":1369,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:35:38.461340+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the full evaluation with nested group 10-fold cross-validation, re-selecting the interquartile-range cutoff, the PageRank threshold, the Gaussian kernel width, and the 50 wave numbers inside each training fold and applying them only to the held-out group. If held-out accuracy drops to the OSC-only baseline (0.635 spectra, 0.683 surfaces), the reported gains are an artifact of tuning on the evaluation groups.","supporting_citations":[{"cited_title":"Computer Networks 56(18), 3825–3833 (2012)","cited_arxiv_id":null,"evidence_quote":"Supplies PageRank centrality, the scoring rule used to rank and filter spectra in the PageRank Filter."},{"cited_title":"Vibrational Spectroscopy57(2), 163–176 (2011)","cited_arxiv_id":null,"evidence_quote":"Defines Raman spectroscopy and the peak-to-composition mapping that the graph edges are built from."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the asymmetric least-squares baseline correction that is part of the OSC cleaning chain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the smoothing filter used in the OSC cleaning chain."},{"cited_title":"Analytical Methods9(30) (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies the group-k-fold protocol that the paper's headline accuracies rely on to prevent surface-level data leakage."},{"cited_title":"Master’s thesis, Dublin City University, Ireland (2022)","cited_arxiv_id":null,"evidence_quote":"Documents the spectral acquisition and calibration protocol that produced the 3,045 spectra used in the evaluation."},{"cited_title":"Journal of Raman Spectroscopy 45(10), 930–940 (2014)","cited_arxiv_id":null,"evidence_quote":"Motivates the dimensionality-reduction stage by linking high-dimensional Raman feature spaces to degraded classifier performance."},{"cited_title":"Chemometrics and Intelligent Laboratory Systems179, 82–84 (2018)","cited_arxiv_id":null,"evidence_quote":"Supplies the cosmic-spike removal algorithm used at the start of the OSC cleaning chain."}],"review_version":1}