{"id":"0fcb21f3-2009-4168-89e7-053d7cffc7e5","arxiv_id":"1908.06137","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Near-infrared spectra from hand skin, classified by a support vector machine, point to 1270 nm as the most informative wavelength for systemic sclerosis, yet the result rests on unverified synthetic data and weak statistics.","lead":"Researchers tested whether near-infrared light scans of the hand plus machine learning can identify systemic sclerosis, a rare autoimmune disease. The work is worth a look because it promises a cheap, non-invasive route to earlier diagnosis, but the supporting evidence is fragile.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unexplained 'false samples' used to balance groups are the load-bearing risk: reported accuracy and the 1270 nm band may be artifacts of synthetic spectra rather than real SSc signals.","rationale":"I agree with the reader's weakest assumption: the synthetic 'false samples' are the load-bearing point. They enter before feature selection, so every reported accuracy and every selected wavelength is conditional on their validity. The paper gives no description of how these samples were made, which is a missing-support flag under the reviewing rules. The statistical reporting is also internally problematic: a U-test result 'greater than 0.99' is used to support a claim that one region is more accurate than another, which is the reverse of a significance claim. That adds to the correctness risk but is secondary to the synthetic-sample issue. The paper does include ethics approval and a plausible measurement setup, and the cited singlet-oxygen literature is relevant to the proposed mechanism; those are credits. However, without code, data, or a real held-out set of patients, the central diagnostic claim is unsupported. The reader's rejection stands, so I leave the verdict unchanged.","tokens_in":6403,"tokens_out":4544,"duration_ms":50618,"concrete_test":"Obtain the synthetic-sample generation routine used in Section 2.1, or the dataset, then re-run the exact RFECV+SVC pipeline on only real spectra: 17 SSc patients versus 17 randomly drawn control spectra, repeated over many draws, with no fabricated samples. If the median test accuracy no longer separates the groups or the top selected wavelength shifts from 1270 nm, the reported result is an artifact of the balancing procedure. Report the distribution of accuracy and selected wavelengths across all draws.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the classifier separates real SSc spectra from real control spectra. Section 2.1 states: 'false samples were created, so the number of 25 samples was established for each of the two groups.' With 25 controls and 17 SSc patients, this implies that roughly 8 fabricated spectra were added to the SSc group, and no procedure for creating them is described. Every downstream result—RFECV-selected wavelengths, the 1270 nm band, and the reported accuracy distribution—is computed on this mixed dataset. If the synthetic spectra were generated by resampling with additive noise, by interpolation, or by any transformation of real data, the classifier can learn the difference between synthetic and real spectra rather than between SSc and controls. A 20% held-out split that includes synthetic samples does not measure real diagnostic performance, because held-out synthetic spectra can be recognized by their synthetic structure. The 1270 nm feature is therefore not established as a biological marker; it may simply be the most discriminative artifact. The claim would hold only if the synthetic spectra are statistically exchangeable with real SSc spectra and the model generalizes to unseen real patients; neither condition is shown. The absence of code and data makes this impossible to audit from the manuscript alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports a pilot study using near-infrared spectroscopy (NIRS) of three hand regions to distinguish patients with systemic sclerosis (SSc) from healthy controls. Spectra were preprocessed by baseline removal and normalization; to balance group sizes, the authors created 'false samples' so that both groups had 25 spectra. A support vector classifier with recursive feature elimination and cross-validation (RFECV) was run 100 times with a 20% test split, and the most frequently selected wavelength bands were recorded. The paper claims that the proximal interphalangeal joints region yields the best accuracy and that the 1270 nm band, attributed to singlet oxygen luminescence, is the most diagnostically important.","tokens_in":6651,"tokens_out":3416,"duration_ms":35074,"significance":"If the central claim were established, a low-cost, non-invasive optical diagnostic for systemic sclerosis would be a valuable clinical tool, and the specific wavelength marker at 1270 nm would provide a testable biological hypothesis. The authors use standard open-source machine-learning software, report a reproducible procedure repeated 100 times, and identify a concrete wavelength band, which are commendable features. However, the significance of the result depends entirely on the validity of the synthetic data used to balance the groups and on the statistical interpretation of the accuracy comparisons; as presented, the evidence does not support the diagnostic claim.","major_comments":[{"comment":"The creation of 'false samples' is unexplained and load-bearing. The text states only that 'false samples were created, so the number of 25 samples was established for each of the two groups.' With 25 controls and 17 SSc patients, this implies that roughly eight fabricated spectra were added to the SSc group, but the generation procedure (e.g., resampling, interpolation, or noise injection) is not described. Every downstream result—the RFECV-selected wavelengths, the 1270 nm band, and the reported accuracy distribution—is computed on this mixed dataset. If the synthetic spectra are not statistically exchangeable with real SSc spectra, the classifier may learn to distinguish synthetic from real spectra rather than SSc from controls, and a 20% held-out split that includes synthetic samples does not measure real diagnostic performance. The authors must either describe the synthetic-data generation method and justify exchangeability, or validate the classifier on a real patient cohort.","section":"Section 2.1, 'Spectra acquiring and preprocessing'"},{"comment":"The statistical comparison between regions is contradictory and does not support the claim that the proximal interphalangeal joints region is best. The text states: 'U-tests were performed... all ρ values resulted were greater than 0.99.' If these are p-values, values greater than 0.99 indicate no statistically significant difference; if these are some other statistic, the test description is incomplete. Either way, the conclusion that one region is superior is not backed by the reported numbers. In addition, accuracy is reported as histograms without confidence intervals or standard errors, and the 100 random 80/20 splits on a dataset of at most 50 samples are not independent replications, so the displayed distributions do not provide a reliable estimate of generalization performance.","section":"Section 3, 'Results'"},{"comment":"The identification of the 1270 nm band as the 'most important wavelength' is derived from the same cohort used to evaluate the classifier. The frequency of wavelength selection is obtained by running RFECV on random training subsets of the same 50 samples, and the accuracy is measured on the corresponding test subsets of the same cohort. This makes the wavelength importance a descriptive statistic of this dataset rather than an independent predictor. The singlet oxygen explanation in Section 4 is a post hoc attribution and does not establish a biological link. Supporting the claim would require an external validation cohort, or at least a nested cross-validation in which feature selection is performed without seeing the test data, and ideally an independent experiment designed to probe the 1270 nm feature.","section":"Sections 2.2 and 3, feature selection and evaluation"}],"minor_comments":[{"comment":"The phrase 'hand's volunteers' should be 'volunteers' hands' or 'the hands of volunteers'; there are also minor grammatical errors elsewhere (e.g., 'The results suggests', 'better accuracy's scores').","section":"Abstract and Introduction"},{"comment":"The preprocessing description is vague: 'the baseline of the raw data had to be removed' by subtracting the minimum value, but the exact order of operations (subtraction, division by source spectrum, filtering) is not fully specified, making the results difficult to reproduce.","section":"Section 2.1"},{"comment":"The grid search over SVC parameters is mentioned but not described: it is not reported which parameter ranges were searched, how performance was evaluated during the search, or whether the 'penalty parameter of 103 order of magnitude' means C = 10^3.","section":"Section 2.2"},{"comment":"Figures 2-4 show histograms of selected wavelengths, but the x-axis is partially omitted 'for better visualization', and the y-axis is not labeled; the number of runs represented by each histogram is stated only in the captions. Table 1 would benefit from a definition of how the 'means' were computed from the counts.","section":"Section 3 and Figures"},{"comment":"The paper does not state where the data or analysis code can be obtained; given the small and synthetic-augmented dataset, public availability would be essential for auditing the results.","section":"General"}],"recommendation":"reject","confidential_remarks":"The unexplained synthetic-data step is a fundamental flaw that cannot be remedied by textual revision: it affects all reported results. Even if the synthetic data were described, the study would still lack external validation and the statistical comparisons are misreported. This is better suited as a methods paper describing a protocol, pending proper validation on a real cohort."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the clinical motivation is real: NIRS is cheap and non-invasive, SSc diagnosis is often delayed, and the specific 1270 nm band is a new claim. Second, the analysis as written cannot support that claim. The unexplained synthetic samples are load-bearing, and the statistics are internally contradictory.\n\nWhat is good: the authors collected actual spectra from 42 subjects (25 controls, 17 SSc) at three hand regions, with ethics approval, a specified commercial spectrometer, and a standard preprocessing chain. The RFECV/SVC pipeline is standard and they report 100 runs. I believe the 1270 nm finding is new, and they are appropriately cautious about the 1360/1690 nm bands.\n\nThe soft spots are not minor. Section 2.1 says 'false samples were created, so the number of 25 samples was established for each of the two groups.' That is roughly eight fabricated SSc spectra, with no description of how they were made. If they were generated by resampling or adding noise to real spectra, the classifier can learn to separate synthetic from real rather than SSc from control, and the 20% held-out split cannot catch it because held-out synthetic spectra have the same artifact. Every reported accuracy and every RFECV wavelength count is computed on that mixed dataset. The 1270 nm band may be the most discriminative artifact, not a biological marker.\n\nThe statistics have a second, independent problem. The paper reports U-tests with all p-values greater than 0.99 and then concludes the Proximal Interphalangeal Joints region is best. That is the wrong direction: high p-values do not support 'greater,' they show no evidence for it. There are also no confidence intervals or effect sizes for the accuracies, and the feature importance is derived from the same small cohort used for evaluation, so it is a description of this dataset rather than an independent prediction. The singlet oxygen explanation is post hoc.\n\nWho is this for? A reader wanting a cautionary example of synthetic data in clinical machine learning, or a starting point for a properly designed pilot. I would not cite it as evidence.\n\nFor peer review: I would send it out rather than desk reject, because the underlying question matters and the data are real. But the referee should require disclosure of the synthetic-sample procedure, code and data, and a re-analysis on real spectra only. Without that, the diagnostic claim should not stand.","headline":"The 1270 nm singlet-oxygen claim is not supported by the analysis as written; the synthetic-sample step and the contradictory statistics make the central result unverifiable.","tokens_in":7153,"tokens_out":3929,"would_cite":false,"duration_ms":40477,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that near-infrared spectra of the hand, classified by a linear support-vector machine with recursive feature elimination, distinguish systemic sclerosis patients from controls, with the 1270 nm singlet-oxygen band as the…","keywords":["systemic sclerosis","scleroderma","near-infrared spectroscopy","support vector classification","recursive feature elimination","singlet oxygen","1270 nm band","non-invasive diagnosis"],"falsifier":"The decisive check is to rerun the same SVC+RFECV pipeline on real spectra only, using leave-one-out classification of the 17 SSc patients against the 25 controls, and see whether accuracy holds and 1270 nm remains the most-selected wavelength; if either fails, the reported diagnostic signal is an artifact of the synthetic samples.","tokens_in":6252,"feed_emoji":"🔬","tokens_out":7242,"duration_ms":70065,"temperature":0.7,"pith_summary":"This paper tries to show that near-infrared spectroscopy (NIRS) of hand skin, plus a machine-learning classifier, can distinguish people with systemic sclerosis (SSc), a chronic autoimmune disease that hardens skin and internal organs, from healthy controls. The authors report that the single most diagnostic wavelength is 1270 nm, which they connect to the luminescence of singlet oxygen, a reactive oxygen species implicated in the skin damage of SSc. They also report that the proximal interphalangeal joints are the best of the three hand regions tested. The evidence is a small pilot: 25 control spectra against 25 patient spectra, of which 8 patient spectra were synthetic balancing samples, so the claim is conditional on those fabricated samples. If correct, the approach would be a low-cost, non-invasive diagnostic aid for a disease usually recognized late.","feed_headline":"1270 nm light separates scleroderma patients from controls","feed_subtitle":"Near-infrared spectra plus a support-vector classifier read hand skin non-invasively, pointing to singlet oxygen.","key_machinery":"The central machinery is the pairing of near-infrared absorbance spectroscopy with Recursive Feature Elimination with Cross-Validation (RFECV) driving a linear Support Vector Classifier (SVC). RFECV repeatedly trains the classifier, removes the least informative wavelength, and uses cross-validation to decide how many wavelengths to keep, so that the output is a short list of diagnostic wavelengths rather than a full spectrum. Repeating the selection over 100 random train/test splits turns that list into a histogram, and the most frequent band, 1270 nm, is then interpreted biochemically as singlet-oxygen luminescence. This gives the argument a concrete optical marker instead of a purely statistical black box.","core_discovery":"In its own terms, the paper finds that systemic sclerosis leaves a measurable optical signature in the near-infrared absorbance of hand skin. A linear support-vector classifier, guided by recursive feature elimination, separates SSc patients from controls across 100 training/test runs, and the wavelength selected most often is 1270 nm, attributed to singlet-oxygen luminescence, consistent with the oxidative-stress component of SSc pathogenesis. The accuracy distribution is best for spectra taken at the proximal interphalangeal joints, and statistical tests rank that region above the metacarpal and proximal phalanges sites. The authors therefore propose NIRS plus machine learning as a path toward faster, non-invasive SSc diagnosis.","pith_inferences":["A direct test the paper leaves implicit: retraining on the 17 real SSc spectra with leave-one-out validation would reveal whether the reported accuracy survives without the synthetic balancing samples; this is the single most informative follow-up.","If the 1270 nm association is real, the same pipeline could be tested on patients with Raynaud's phenomenon alone to see whether the optical signal precedes clinical skin fibrosis, an early-diagnosis claim the paper does not make.","The authors interpret 1270 nm as singlet-oxygen luminescence but do not establish a dose-response link; correlating the 1270 nm intensity with a clinical fibrosis score, such as the modified Rodnan skin score, would test that biochemical reading.","More broadly, the method implies a cheap optical oxidative-stress readout for autoimmune disease, which could extend to other fibrotic conditions, but that extension is speculative until the SSc result is replicated independently."],"forward_implications":["A standard, inexpensive near-infrared spectrometer paired with SVC could serve as a screening aid in rheumatology clinics, shortening the time from symptom onset to diagnosis.","The proximal interphalangeal joints should be the standard measurement region in any follow-up protocol, since they produced the most accurate classifications.","The 1270 nm band gives a concrete optical target for future work: instrument designers and clinicians can focus on a narrow spectral window rather than whole-spectrum diagnostics.","Because NIRS uses non-ionizing light and no biopsy, the approach could be repeated safely to track disease progression or response to treatment."],"supporting_citations":[{"why":"Supplies the support-vector machine classification theory on which the binary SSc/control classifier rests.","marker":"[19]"},{"why":"Provides the implementation of SVC and the cross-validated recursive feature elimination used to select wavelengths.","marker":"[20]"},{"why":"Supplies the recursive feature elimination methodology used to identify the most diagnostic wavelengths and to guard against overfitting.","marker":"[23]"},{"why":"Demonstrates direct detection of singlet-oxygen luminescence in human skin and cells, the basis for interpreting the 1270 nm band.","marker":"[25]"},{"why":"Provides the clinical overview of systemic sclerosis against which the optical diagnostic claims are positioned.","marker":"[13]"},{"why":"Links reactive oxygen species to systemic autoimmune pathology, supporting the singlet-oxygen interpretation.","marker":"[32]"},{"why":"Reviews continuous-wave near-infrared spectroscopy instrumentation and methodology, grounding the NIR measurement approach.","marker":"[17]"},{"why":"Supplies the baseline-plus-noise signal model used to preprocess the spectra before classification.","marker":"[18]"}],"fun_headline_variants":["NIR spectra plus ML spot scleroderma in hand joints","1270 nm singlet oxygen signal flags scleroderma","Hand NIR with SVM reads scleroderma non-invasively","Proximal joints' NIR best for SSc diagnosis via ML","Singlet oxygen at 1270 nm distinguishes SSc patients"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire accuracy result rests on the assumption that the eight synthetic spectra added to balance the SSc group behave exactly like real patient spectra; if they do not, the reported accuracies and the 1270 nm selection are artifacts of the balancing procedure.","fun_headline_variants_meta":{"raw":{"variants":["NIR spectra plus ML spot scleroderma in hand joints","1270 nm singlet oxygen signal flags scleroderma","Hand NIR with SVM reads scleroderma non-invasively","Proximal joints' NIR best for SSc diagnosis via ML","Singlet oxygen at 1270 nm distinguishes SSc patients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000404,"raw_usage":{"total_tokens":2049,"prompt_tokens":835,"completion_tokens":1214,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":1129}},"tokens_in":451,"tokens_out":1214,"duration_ms":7714,"temperature":1.0,"reasoning_tokens":1129,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:53:51.787490+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The decisive check is to rerun the same SVC+RFECV pipeline on real spectra only, using leave-one-out classification of the 17 SSc patients against the 25 controls, and see whether accuracy holds and 1270 nm remains the most-selected wavelength; if either fails, the reported diagnostic signal is an artifact of the synthetic samples.","supporting_citations":[{"cited_title":"A tutorial on support vector machines for pattern recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the support-vector machine classification theory on which the binary SSc/control classifier rests."},{"cited_title":"Gene selection for cancer classification using support vector machines","cited_arxiv_id":null,"evidence_quote":"Supplies the recursive feature elimination methodology used to identify the most diagnostic wavelengths and to guard against overfitting."},{"cited_title":"Direct detection of singlet oxygen generated by uva irradiation in human cells and skin.Journal of Investigative Dermatology, 127(6):1498– 1506, 2007","cited_arxiv_id":null,"evidence_quote":"Demonstrates direct detection of singlet-oxygen luminescence in human skin and cells, the basis for interpreting the 1270 nm band."},{"cited_title":"Systemic sclerosis","cited_arxiv_id":null,"evidence_quote":"Provides the clinical overview of systemic sclerosis against which the optical diagnostic claims are positioned."},{"cited_title":"Oxygen free radicals and systemic autoimmunity","cited_arxiv_id":null,"evidence_quote":"Links reactive oxygen species to systemic autoimmune pathology, supporting the singlet-oxygen interpretation."},{"cited_title":"A review on continuous wave functional near-infrared spectroscopy and imaging instrumenta- tion and methodology","cited_arxiv_id":null,"evidence_quote":"Reviews continuous-wave near-infrared spectroscopy instrumentation and methodology, grounding the NIR measurement approach."},{"cited_title":"An optimization perspec- tive on baseline removal for spectroscopy","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline-plus-noise signal model used to preprocess the spectra before classification."}],"review_version":1}