{"id":"676a5076-4cd3-46f8-a04e-6fca0482590b","arxiv_id":"2412.10567","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"Semi-supervised models at best match, but do not outperform, supervised models on CVD detection from BRFSS data, and the paper lacks proper baselines and error bars.","lead":"This paper applies five existing semi-supervised learning methods to predict cardiovascular disease from the public BRFSS survey, claiming they outperform supervised models when few labels are available. The reported results do not support the headline claim: a supervised gradient boosting model beats the best semi-supervised model on accuracy and AUC.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that semi-supervised models reduce labeling dependence is untested: Table I compares semi-supervised models using 50% labeled data against supervised models using 100% labeled data, with no supervised baseline on the same 50% labeled subset, so any observed parity cannot be attributed to…","rationale":"The reader's verdict of REJECT is well-founded. The paper's own results fail to support the headline claim, and the experimental design cannot isolate the effect of semi-supervised learning. My concern is the most load-bearing because it targets the causal inference at the heart of the contribution: the assertion that unlabeled data reduces labeling dependence. A simple control experiment would resolve it. The paper also lacks error bars, hyperparameter details, and class-imbalance handling, but these are secondary. I agree with the reader's weakest assumption. No ad hominem intended; the critique is on the evidence structure. The verdict should remain REJECT unless the missing baselines are supplied and show a real SSL advantage. Therefore verdict_should_be is UNCHANGED (no change to the reader's rejection), and agreement_with_reader is agree.","tokens_in":6008,"tokens_out":3403,"duration_ms":28563,"concrete_test":"Retrain each supervised model (at least gradient boosting and logistic regression) using only the same 50% labeled subset that was given to the SSL models, with identical preprocessing and hyperparameter choices, and evaluate on the same test set. If any supervised 50%-label model achieves F1/AUC statistically indistinguishable from or higher than Self-Training's 0.8425 AUC / 0.5175 F1, then the unlabeled data provides no demonstrated benefit and the central claim fails. Run this over multiple random 50/50 splits with confidence intervals to account for sampling variability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is that semi-supervised learning (SSL) can match or beat supervised learning while using only half the labels. This requires demonstrating that unlabeled data provides a performance boost beyond what the labeled subset alone gives. Section IV.B (Table I) does not test this: SSL models are trained on 50% labeled plus 50% unlabeled training data, while supervised models are trained on 100% labeled data. There is no supervised model trained on the same 50% labeled subset. Therefore, the comparison conflates three factors: percentage of labels, model class, and the SSL mechanism. For example, Self-Training achieves AUC 0.8425 with 50% labels; a gradient boosting model trained on 100% labels achieves 0.8439. This does not tell us whether unlabeled data helped; the 50%-label gradient boosting baseline might be equally good or better. The conclusion in Section IV.B that SSL 'can match the top supervised models' metrics with half the labeled data' is therefore unsupported. Furthermore, the abstract's 'outperform' is contradicted by Table I itself: Gradient Boosting has higher accuracy (0.8027 vs 0.8003) and AUC (0.8439 vs 0.8425) than the best SSL model, with essentially tied F1. The missing matched-label baseline is the load-bearing gap because without it there is no evidence for the reduced-labeling claim that motivates the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies five semi-supervised learning (SSL) methods (Semi-Supervised SVM, Self-Training, Pseudo-Labeling, Mean Teacher, Pi-Model) to cardiovascular disease detection using a cleaned subset of 20,544 BRFSS records. The experimental protocol trains SSL models on 50% labeled plus 50% unlabeled training data and supervises five standard classifiers (Logistic Regression, Decision Tree, Random Forest, KNN, Gradient Boosting) on the full 100% labeled training set. The central claim, stated in the abstract and repeated in Section IV.B and the conclusion, is that SSL models outperform traditional supervised learning while reducing dependence on labeled data. However, the paper's own Table I shows that the best SSL model (Self-Training) achieves accuracy 0.8003, F1 0.5175, and AUC 0.8425, whereas supervised Gradient Boosting achieves 0.8027, 0.5179, and 0.8439, respectively, and Logistic Regression achieves AUC 0.8443. The comparison is also not matched on the amount of labeled data: no supervised model is trained on the same 50% labeled subset used by the SSL models.","tokens_in":6322,"tokens_out":3021,"duration_ms":27767,"significance":"If the central claim were established, the paper would provide useful evidence that SSL can lower annotation costs for cardiovascular risk screening. The authors are to be credited for evaluating several SSL variants and reporting three standard metrics on a public dataset, and for describing the methods with equations. However, the load-bearing claim that SSL outperforms supervised learning is contradicted by the paper's own Table I, and the reduced-labeling claim is untested because the comparison lacks a supervised baseline on the same 50% labeled subset. The empirical contribution is therefore not sufficient to support the abstract's conclusions. Disagreement with the current consensus is not the issue; the internal inconsistency between the stated claims and the reported numbers is the central problem.","major_comments":[{"comment":"The abstract claims that 'semi-supervised models outperform traditional supervised learning techniques,' but Table I directly contradicts this. Self-Training, the best SSL model, has accuracy 0.8003 and AUC 0.8425, while supervised Gradient Boosting has accuracy 0.8027 and AUC 0.8439, and Logistic Regression has AUC 0.8443. The F1 scores are essentially tied (0.5175 vs. 0.5179). The paper should either remove the outperform claim or provide a statistically validated setting in which SSL is actually superior on at least one primary metric.","section":"Abstract and Section IV.B"},{"comment":"The assertion that SSL 'can match the top supervised models' metrics with half the labeled data' is not supported by the experimental design. SSL models are trained on 50% labeled plus 50% unlabeled data, whereas supervised models are trained on 100% labeled data. Without a supervised baseline trained on the same 50% labeled subset, any parity between the two families could be due to the supervised model class, the model capacity, or the reduced data size, rather than to the SSL mechanism. The authors should add matched-label supervised baselines at 10%, 20%, 30%, 40%, and 50% labeled data and report the corresponding metrics.","section":"Section IV.B"},{"comment":"Figures 2 and 3 are described as showing 'performance enhancements' in F1 and AUC as labeled data increases from 10% to 50%, but no numerical values, standard deviations, or confidence intervals are reported. The paper also does not state how many random splits or training runs were performed. Without variance estimates or significance tests, the apparent improvements and the ranking of models (e.g., Self-Training versus Semi-Supervised SVM) are not statistically grounded.","section":"Section IV.A"},{"comment":"Several implementation details required for reproducibility are missing. For Semi-Supervised SVM, the penalty parameters C and C* are not specified. For Self-Training and Pseudo-Labeling, the confidence threshold is not given. For Mean Teacher and Pi-Model, the network architecture, optimizer, learning rate, batch size, number of epochs, and the regularization coefficient λ are not provided. The paper should also describe how the 50% labeled/50% unlabeled split was constructed (e.g., random, stratified, seeded) so that the experiment can be repeated.","section":"Section III"}],"minor_comments":[{"comment":"The sentence ending 'with only 50' is incomplete; it should read 'with only 50% labeled data.'","section":"Section IV.B"},{"comment":"The text contains grammatical errors, such as 'In out dataset' and 'The response variable is Heart Disease , shows'.","section":"Section II.A"},{"comment":"The reference list contains many works that are not cited in the text or are unrelated to the topic, including several self-citations involving the same authors. Examples include references on PCM microcapsules, augmented reality, and hardware accelerators. These citations should be removed or replaced with relevant literature on semi-supervised learning and cardiovascular disease prediction.","section":"References"},{"comment":"The equation for Semi-Supervised SVM is not fully introduced; in particular, the meaning and range of the slack variables ξ_j for unlabeled data and the role of C* should be explained in the text.","section":"Section III.A"}],"recommendation":"reject","confidential_remarks":"The central claim of the paper is internally contradicted by its own Table I, and the comparison protocol does not isolate the effect of unlabeled data. These are load-bearing issues that cannot be resolved by minor editing. In addition, the reference list appears to contain a substantial number of irrelevant and self-cited entries, which may indicate citation padding and should be investigated editorially."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the abstract overstates what the data show. Table I has Gradient Boosting beating every semi-supervised model on accuracy and AUC, and the comparison that would actually support the title—semi-supervised with 50% labels versus supervised with the same 50% labels—isn't in the paper. The 'reduced labeling dependence' claim is therefore not tested.\n\nThe paper does have some virtues. It uses a well-known public dataset (BRFSS), runs five standard SSL methods, and reports accuracy, F1, and AUC for labeled proportions from 10% to 50%. The descriptions of the methods are clear enough for a conference paper, and the workflow is easy to follow. For someone wanting a quick benchmark of off-the-shelf SSL on a tabular health dataset, the figures give a rough sense of what to expect.\n\nThe soft spots are significant. The missing supervised baseline at 50% labels is the load-bearing gap: without it, the parity between Self-Training (50% labels) and Gradient Boosting (100% labels) tells you nothing about whether unlabeled data contributed. A 50%-label gradient boosting might be just as good. The abstract's 'outperform' is simply wrong relative to the paper's own results. Beyond that, there are no error bars or repeated runs, hyperparameters for the SSL methods are not reported, and the reference list contains several papers irrelevant to the topic (construction materials, AR overlays) plus a heavy block of self-citations that don't support the claims.\n\nI don't think this is a case of deliberate misdirection; more likely a rushed write-up where the conclusion ran ahead of the experiments. But as a scientific contribution, the main point fails on its own evidence. A serious editor should desk-reject this. If the authors add the matched-label baseline, tighten the abstract, and report proper uncertainty, they might have a small empirical note—but that's for a workshop-level venue, not a journal.","headline":"The paper's core claim—that semi-supervised learning reduces labeling dependence—is untested because Table I lacks a supervised baseline on the same 50% labeled subset, and the abstract's 'outperform' is contradicted by the paper's own results.","tokens_in":6844,"tokens_out":3429,"would_cite":false,"duration_ms":30577,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that semi-supervised learning, trained on just half the labeled training data, detects cardiovascular disease as accurately on BRFSS self-reported health data as fully supervised models, with Self-Training matching the…","keywords":["cardiovascular disease detection","semi-supervised learning","self-training","pseudo-labeling","BRFSS","labeled data scarcity","AUC","F1-score"],"falsifier":"Train Gradient Boosting, Logistic Regression, and Random Forest on exactly the same 50% labeled subset used by the SSL models, and compare their Accuracy, F1, and AUC on the same test set. If any of these supervised models matches or beats Self-Training's 0.8003/0.5175/0.8425 without seeing the unlabeled data, the paper's attribution of the gain to unlabeled data is undercut.","tokens_in":5811,"feed_emoji":"❤️","tokens_out":12514,"duration_ms":100511,"temperature":0.7,"pith_summary":"Cardiovascular disease detection typically depends on large, carefully labeled datasets, which are hard to produce in clinical settings. This paper tries to show that semi-supervised learning relaxes that requirement: on a cleaned 20,544-record BRFSS survey sample, five semi-supervised models trained with only 50% of the training data labeled are compared against five fully supervised models trained on all labels. The headline result is Self-Training, which reaches Accuracy 0.8003, F1 0.5175, and AUC 0.8425, within a few thousandths of a point of Gradient Boosting's 0.8027, 0.5179, and 0.8439. The abstract states the claim as 'semi-supervised models outperform traditional supervised learning techniques,' while the table shows the best semi-supervised model matching rather than beating the best supervised one. If the comparison holds, initial CVD screening could be built with far fewer expert annotations at nearly the same accuracy.","feed_headline":"Half the labels, same heart-disease detection","feed_subtitle":"On 20,544 BRFSS records, Self-Training's AUC 0.8425 nearly matches Gradient Boosting's 0.8439 with half the labels.","key_machinery":"The load-bearing mechanism is the confidence-thresholded pseudo-labeling loop of Self-Training: a logistic-regression seed classifier predicts labels for unlabeled points, keeps only predictions above a threshold, adds them to the training set, and retrains repeatedly. Pseudo-Labeling is the same idea with a fixed high threshold, while Semi-Supervised SVM adds slack terms for unlabeled points to the margin objective, and Mean Teacher and Pi-Model enforce consistency between EMA-averaged or perturbed network outputs. The comparison in Table I is what carries the paper's conclusion: it attributes Self-Training's near-parity with Gradient Boosting to the unlabeled half of a 50%-labeled training set.","core_discovery":"On the paper's own terms, the central discovery is that a confidence-based pseudo-labeling loop can substitute unlabeled survey data for missing labels: with half the training set labeled, the Self-Training model attains Accuracy 0.8003, F1 0.5175, and AUC 0.8425, versus Gradient Boosting's 0.8027, 0.5179, and 0.8439 on the full labeled set. The authors interpret the near-parity as evidence that unlabeled data carries enough signal to match the top supervised model, and they generalize this to the claim that semi-supervised learning outperforms traditional supervised techniques for cardiovascular disease detection. The other semi-supervised models—Semi-Supervised SVM, Pseudo-Labeling, Mean Teacher, and Pi-Model—also stay above 0.75 AUC at 50% labels, though they trail Self-Training.","pith_inferences":["The paper does not include a supervised baseline trained on the same 50% labeled subset; adding that matched-label baseline would separate the contribution of unlabeled data from the contribution of model class and data quantity.","The F1-scores near 0.5 show the dataset is class-imbalanced, so a natural extension is to test whether Self-Training's gain persists under class-weighting, resampling, or higher-threshold operating points.","The reported numbers come from a single 75/25 train/test split without resampling, so a repeated-split or bootstrap replication would show whether Self-Training's near-tie with Gradient Boosting is stable or within noise.","The same pseudo-labeling recipe could be applied to other chronic-disease flags in BRFSS (diabetes, stroke, kidney disease), where the survey structure and labeling costs are identical."],"forward_implications":["On the BRFSS-derived data, the best semi-supervised model needs only half the labeled training data to match the best fully supervised model's Accuracy, F1, and AUC, so the labeling burden for this screening task could be cut roughly in half.","All five semi-supervised models improve steadily as the labeled fraction rises from 10% to 50%, indicating that they make effective use of added labels as well as unlabeled data.","Self-Training is the strongest semi-supervised method in the comparison, while consistency-based methods (Mean Teacher and Pi-Model) lag, suggesting that pseudo-labeling is the more productive direction on this kind of tabular health-survey data.","The authors propose semi-supervised screening as an initial triage step in clinical environments, where fully labeled datasets are the bottleneck."],"supporting_citations":[{"why":"Supplies the BRFSS dataset and its reliability/validity background; every experiment in the paper uses this data source.","marker":"[15]"},{"why":"The paper cites it as the source for the Semi-Supervised SVM margin objective that extends SVM with unlabeled-data slack terms.","marker":"[31]"},{"why":"The paper cites it as the source for the iterative Self-Training loop that pseudo-labels high-confidence unlabeled points.","marker":"[6]"},{"why":"The paper cites it as the source for Pseudo-Labeling, the high-confidence-threshold variant of self-training.","marker":"[9]"},{"why":"The paper cites it for the Mean Teacher exponential-moving-average weight update used by one neural semi-supervised baseline.","marker":"[11]"},{"why":"The paper cites it for the Pi-Model consistency-regularization loss that penalizes disagreement between perturbed views.","marker":"[29]"},{"why":"The paper cites it for Random Forest, one of the five supervised baselines in the comparison.","marker":"[20]"},{"why":"The paper cites it for Gradient Boosting, the strongest supervised baseline that Self-Training is compared against.","marker":"[17]"},{"why":"The paper cites it for Logistic Regression, which serves both as a supervised baseline and as the seed classifier inside Self-Training.","marker":"[4]"}],"fun_headline_variants":["Semi-supervised CVD model needs half the labels","Self-training matches gradient boosting with half labels","Unlabeled data steps in for missing heart disease labels","CVD detection: half labels, on par with full supervision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central comparison assumes that an SSL model trained on 50% labeled plus 50% unlabeled data differs from a supervised model trained on 100% labeled data only in the unlabeled data, yet no supervised model is trained on the same 50% labeled subset, so the observed parity could stem from data quantity, model class, or procedure rather than from semi-supervision itself.","fun_headline_variants_meta":{"raw":{"variants":["Semi-supervised CVD model needs half the labels","Self-training matches gradient boosting with half labels","Unlabeled data steps in for missing heart disease labels","CVD detection: half labels, on par with full supervision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1179,"prompt_tokens":844,"completion_tokens":335,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":272}},"tokens_in":460,"tokens_out":335,"duration_ms":3291,"temperature":1.0,"reasoning_tokens":272,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:49:47.494515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Gradient Boosting, Logistic Regression, and Random Forest on exactly the same 50% labeled subset used by the SSL models, and compare their Accuracy, F1, and AUC on the same test set. If any of these supervised models matches or beats Self-Training's 0.8003/0.5175/0.8425 without seeing the unlabeled data, the paper's attribution of the gain to unlabeled data is undercut.","supporting_citations":[{"cited_title":"A systematic review of publications assessing reliability and validity of the behavioral risk factor surveillance system (brfss), 2004–2011","cited_arxiv_id":null,"evidence_quote":"Supplies the BRFSS dataset and its reliability/validity background; every experiment in the paper uses this data source."},{"cited_title":"Demonstration of chromocloth: Re-programmable multi-color textures through flexible and portable light source","cited_arxiv_id":null,"evidence_quote":"The paper cites it as the source for the Semi-Supervised SVM margin objective that extends SVM with unlabeled-data slack terms."},{"cited_title":"6: Simultaneous tracking, tagging and mapping for augmented reality","cited_arxiv_id":null,"evidence_quote":"The paper cites it as the source for Pseudo-Labeling, the high-confidence-threshold variant of self-training."},{"cited_title":"Predicting 30-day hospital readmission in medicare patients: Insights from an lstm deep learning model","cited_arxiv_id":null,"evidence_quote":"The paper cites it for the Mean Teacher exponential-moving-average weight update used by one neural semi-supervised baseline."},{"cited_title":"Optimizing distributed networking with big data scheduling and cloud computing","cited_arxiv_id":null,"evidence_quote":"The paper cites it for the Pi-Model consistency-regularization loss that penalizes disagreement between perturbed views."},{"cited_title":"Influencing factors of housing price in new york-analysis: Based on excel multi-regression model","cited_arxiv_id":null,"evidence_quote":"The paper cites it for Random Forest, one of the five supervised baselines in the comparison."},{"cited_title":"A mixed- heuristic quantum-inspired simplified swarm optimization algorithm for scheduling of real-time tasks in the multiprocessor system","cited_arxiv_id":null,"evidence_quote":"The paper cites it for Gradient Boosting, the strongest supervised baseline that Self-Training is compared against."},{"cited_title":"Green building material with superior thermal insulation and energy storage properties fabricated by paraffin and foam cement composite","cited_arxiv_id":null,"evidence_quote":"The paper cites it for Logistic Regression, which serves both as a supervised baseline and as the seed classifier inside Self-Training."}],"review_version":1}