{"id":"46841567-78a4-4cad-9359-59007720746e","arxiv_id":"1908.02338","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A 1D convolutional neural network trained on balanced windowed segments of fetal heart rate traces classifies abnormal birth outcomes with an AUC of 0.86, outperforming classical baselines tested by the authors.","lead":"Researchers trained a one-dimensional convolutional neural network on windowed segments of fetal heart rate traces to classify birth outcomes from cardiotocography recordings. The method balances the heavily skewed dataset by generating equal numbers of control and case windows, and the best model achieved an area under the curve of 0.86 for distinguishing abnormal from normal deliveries.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Positive class includes 24 caesareans with no confirmed pathology; if these labels are wrong, the model may learn delivery mode rather than fetal abnormality, undermining the reported AUC.","rationale":"The reader's weakest assumption correctly identifies the unvalidated caesarean labels as the most load-bearing concern. Unlike the window-independence issue, which affects the precision and interpretation of the performance metrics but not their construct, a mislabelled positive class changes what the model is actually predicting. The paper is transparent about the assumption, which is commendable, but the central claim ('early detection of abnormal birth outcomes') cannot be evaluated without knowing whether the 24 unconfirmed caesareans are truly abnormal. The proposed concrete test would settle this by training on only confirmed cases; if performance collapses, the model is likely exploiting caesarean-related signal rather than fetal pathology. Since this is fixable with a cleaner outcome definition and the method itself may be sound, the conditional verdict remains appropriate. We therefore recommend no change to the reader's verdict.","tokens_in":15681,"tokens_out":4701,"duration_ms":49479,"concrete_test":"Retrain the W=200 1DCNN with the same architecture and windowing, but define the positive class as only the 22 confirmed pathological cases (18 with pH ≤ 7.20; 4 with 7.20 < pH ≤ 7.25), and evaluate with record-level cross-validation against the 506 normal controls. Compare the resulting AUC, sensitivity, and specificity to the reported 0.86, 0.80, and 0.79; if the AUC falls below approximately 0.75, the reported result depends on the 24 unvalidated caesarean labels.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed 1DCNN distinguishes normal from abnormal birth outcomes, with AUC 0.86. However, Section 2.1 defines the positive class as 46 caesarean deliveries, of which 24 (52%) have no pathological outcome measures (no acidosis, no deterioration). The authors state they assume the caesarean decision was supported by pathological concerns and explicitly note 'there is no way to validate this'. If any of these 24 caesareans were performed for non-pathological reasons (e.g., maternal request, previous caesarean, malpresentation), the positive class is contaminated with fetuses that are actually normal. Because caesarean delivery itself is associated with different clinical management and potentially different FHR patterns, a classifier could learn to separate vaginal deliveries from caesarean deliveries rather than normal from pathological fetuses. Since these 24 subjects constitute the majority of the positive class, the reported sensitivity, specificity, and AUC may reflect delivery mode, not fetal compromise. This is a construct validity threat to the central claim, and the paper provides no evidence to rule it out.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for classifying abnormal birth outcomes from intrapartum cardiotocography (CTG) traces. The method splits each fetal heart rate (FHR) signal into non-overlapping fixed-length windows, retains all windows from the 46 abnormal deliveries, randomly subsamples an equal number of windows from the 506 normal deliveries, and trains a one-dimensional convolutional neural network (1DCNN) on the raw FHR windows. On the CTU-UHB dataset the authors report a best test performance with window size 200: sensitivity 80%, specificity 79%, and AUC 0.86, outperforming MLP, SVM, RF, FLDA, and their earlier SMOTE-based pipeline. The paper also discusses limitations, including the absence of clinical annotations in the CTG traces and the unverifiable assumption behind the positive-label definition.","tokens_in":15743,"tokens_out":5926,"duration_ms":63849,"significance":"If the reported result were valid, the windowing-plus-1DCNN pipeline would be a useful and simple contribution to imbalanced clinical time-series classification, removing handcrafted feature engineering and using only real data. Strengths include the use of a public dataset (CTU-UHB), head-to-head comparisons against several conventional classifiers under the same windowing protocol, and transparent acknowledgment of the unverifiable label assumption in Section 2.1 and of the lack of trace annotations in the Conclusion. However, the current evaluation does not establish the central claim because the test metrics are computed at the window level, the best window size is chosen after inspecting test results, and the positive-label definition may encode mode of delivery rather than fetal pathology.","major_comments":[{"comment":"All performance metrics and confidence intervals are computed over test windows, with multiple non-overlapping windows from the same CTG recording treated as independent observations. For W=200, the test set contains 620 case windows and 620 control windows drawn from only 10 case recordings and 101 control recordings, so the effective sample size for estimating sensitivity, specificity, and AUC is much smaller than 1,240. Adjacent windows from one trace are highly correlated, and the model may exploit per-recording artifacts. This inflates the reported confidence intervals and can make a model look substantially better than its recording-level performance. Please report results at the recording level (for example, majority vote over windows per recording) or use clustered cross-validation that leaves entire recordings out, and adjust confidence intervals for clustering.","section":"Section 2.2 and Tables 2, 6; Eq. (7)"},{"comment":"The window size W=200 is selected as best after inspecting test-set ROC curves and test AUC in Table 6. The same test-set selection pattern appears for the MLP, SVM, RF, and FLDA baselines. Selecting a hyperparameter such as window size on the test set and then reporting that test performance as the headline result is a form of test-set leakage; the expected best-of-five AUC is larger than the expected AUC for a fixed configuration. The reported 0.86 AUC for W=200 is therefore optimistic and is not a valid estimate of generalization performance. Please select window size and other hyperparameters using a validation split or nested cross-validation, and report the performance of the selected configuration on a truly held-out test set.","section":"Section 3.2.2 and Table 6; also Sections 3.1.2, 3.3.2, 3.3.4, 3.3.6"},{"comment":"The positive class is defined as 46 caesarean deliveries, of which 18 had pH≤7.20, 4 had pH between 7.20 and 7.25 (fetal deterioration), and 24 had no recorded pathological outcome measures. The authors state that they assume the decision to deliver by caesarean was supported by underlying pathological concerns, with the explicit caveat that there is no way to validate this. Because these 24 subjects constitute 52% of the positive class, the classifier could be learning to distinguish caesarean deliveries from vaginal deliveries rather than pathological from normal fetal states, especially since caesarean delivery is associated with different clinical management and possibly different FHR patterns. This is a construct-validity threat to the central claim of detecting abnormal birth outcomes. Please provide a sensitivity analysis restricted to the 22 cases with confirmed pathological outcomes (acidosis or deterioration), and clearly interpret results for the full 46-case label definition.","section":"Section 2.1"},{"comment":"The paper describes two different split protocols. Section 3 says the data set is split randomly into training (80%), validation (10%), and testing (10%), while Section 2.2 describes only training and test sets (405/101 controls, 36/10 cases) with no validation set. It is unclear how the validation set used for early stopping and model selection was constructed, whether it was drawn from the training records before windowing, and whether any test windows came from records used in training. In addition, because control windows are randomly subsampled to balance classes, the test set is stochastic, yet no random seeds or repeated subsampling results are reported. Please specify the exact split at the recording level, including the validation set, and report results across multiple subsampling seeds or with a fixed seed.","section":"Sections 2.2 and 3"}],"minor_comments":[{"comment":"The definitions of sensitivity and specificity appear reversed: the text says sensitivity describes true positives (normal deliveries) and specificity describes true negatives (pathological deliveries), which contradicts the usual convention and the paper's own treatment of caesarean/pathological cases as the positive class. Please correct these definitions and make the positive-class convention explicit throughout.","section":"Section 2.4"},{"comment":"The confidence-interval formula uses 'i' both as an index and as the quantile, and the phrase '1 − α/2)-quantile' has a formatting error; this should be the standard normal quantile z_{1−α/2}. Please rewrite the formula and its notation.","section":"Equation (7)"},{"comment":"The last row of Table 9 has a formatting error: the specificity confidence interval is written as '0.53,0.70\")' with a stray quotation mark and parenthesis.","section":"Table 9"},{"comment":"The text says 'Fig. 9 shows that the ROC curves for all SVM models...', but Figure 9 is the 1DCNN ROC plot; the SVM ROC curves appear in Figure 10. The cross-reference should be corrected.","section":"Section 3.3.2"},{"comment":"The text refers to 'Table 8' when presenting the SMOTE oversampling results, but the relevant table is Table 10. Please correct the cross-reference.","section":"Discussion, SMOTE comparison"},{"comment":"The 1DCNN architecture description is not fully reproducible: the kernel size is said to be half the windowing strategy with the example '150 for 300 data points', but no kernel size is given for W=200, and the number of filters, stride, pooling size, and dropout are not specified. Please provide a complete architecture table.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The label-assumption problem in Section 2.1 is the deepest issue: with 24 of 46 positives lacking any confirmed pathological outcome, the paper's framing as 'detection of abnormal birth outcomes' may not match what the classifier actually learns. The authors should consider either re-framing the task as caesarean versus vaginal delivery classification or restricting the positive class to confirmed pathological cases. The window-level evaluation and test-set hyperparameter selection are fixable with standard methodological changes, but they currently prevent the reported AUC from being taken at face value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the windowing + 1DCNN idea is neat and the paper is worth a look, but I don't trust the headline numbers as an estimate of clinical performance. The main result (AUC 0.86, W=200) is built on three assumptions that the authors mostly acknowledge but don't fix: windows from the same recording treated as independent, test-set used to select the window size, and a positive class where 24 of 46 cases have no confirmed pathology.\n\nWhat's actually new: combining non-overlapping windowing as a class-balancing strategy with a simple 1DCNN on raw FHR traces. That's a sensible way to use real data rather than synthetic oversampling, and it's a plausible contribution to the CTG analysis literature. The paper does some things well: public dataset, standard baselines (MLP, SVM, RF, FLDA), and a comparison with their own SMOTE work. They also tell you exactly what the label limitation is, which is more transparent than many papers in this space.\n\nSoft spots, in order of severity. First, the label problem. The positive class is 46 caesareans, and for 24 of them there is no recorded acidosis or deterioration; the authors assume the caesarean was for pathological reasons but admit they can't validate it. Since those 24 are more than half the positives, the model may be learning \"delivery mode\" rather than \"fetal compromise.\" That threatens the construct validity of the whole experiment. Second, the evaluation ignores the recording structure: multiple windows from one recording are treated as independent observations. The confidence intervals use window counts, not recording counts, so they're far too tight and the AUC is likely inflated by within-recording correlation. Third, the best window size is picked from test-set results. The text says validation pointed to W=200, but Table 5 actually shows W=500 with higher validation AUC, so the selection is at least inconsistent. A pre-specified window or a nested validation loop would fix this.\n\nThese are all fixable with a recording-level bootstrap or cross-validation, a pre-registered window size, and a cleaner outcome definition (e.g., only confirmed acidosis or a composite of known pathology). None of the problems makes the method itself nonsensical; the approach is reasonable and the data is public, so someone could rerun it properly.\n\nMy bottom line: this paper deserves a serious referee. It's a useful example of both a practical class-imbalance trick and a set of common evaluation traps. But I wouldn't cite the 0.86 AUC as evidence of a working clinical tool. If the authors tighten the evaluation, the result might hold in a weaker form; as it stands, it's a \"conditional\" rather than a \"validated\" claim.\n\nRecommendation: send it to review with a request for major revision, focusing on the three points above.","headline":"Reasonable windowing + 1DCNN idea on public CTG data, but the headline AUC is undermined by label contamination, non-independent windows, and test-set model selection.","tokens_in":16410,"tokens_out":2836,"would_cite":false,"duration_ms":29002,"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":"Windowing fetal heart-rate traces into non-overlapping 200-point segments lets a one-dimensional convolutional network detect abnormal birth outcomes with 80% sensitivity and 79% specificity.","keywords":["cardiotocography","fetal heart rate","one-dimensional convolutional neural network","class imbalance","time-series windowing","birth outcome prediction","deep learning","CTG classification"],"falsifier":"Retrain the 200-window 1DCNN using only the 18 deliveries with umbilical-artery pH $\\le 7.20$ as the abnormal class and an equal number of normal windows; if the window-level AUC stays near 0.86 the separation reflects real pathology, whereas a drop toward chance or a model that labels the 24 unconfirmed caesareans as normal would show that delivery-mode labels drove the reported result.","tokens_in":15348,"feed_emoji":"🩺","tokens_out":10180,"duration_ms":103721,"temperature":0.7,"pith_summary":"This paper claims that the class imbalance that plagues machine learning on cardiotocography (CTG) recordings can be removed without synthetic data by segmenting traces into non-overlapping windows, keeping every window from abnormal deliveries and randomly subsampling an equal number from normal deliveries. A one-dimensional convolutional neural network trained on these raw windows learns discriminative features automatically, removing the need for manually engineered clinical features. With 200-sample windows the test performance reaches 80% sensitivity and 79% specificity with an AUC of 0.86, and this combination outperforms an MLP, SVM, random forest, FLDA, and the authors' previous SMOTE approach on the same 552-recording dataset. A sympathetic reader would take this as evidence that windowed real data plus learned features is a workable route to automated intrapartum monitoring.","feed_headline":"200-point CTG windows let a 1D CNN flag abnormal births at AUC 0.86","feed_subtitle":"Balancing skewed birth data with real windows beats handcrafted features and synthetic oversampling.","key_machinery":"The load-bearing machinery is a one-dimensional convolutional neural network applied to segmented raw signals: a 1D convolution sweeps learnable kernels along the time axis to detect local patterns, a max-pooling layer downsamples them, and two dense layers classify the resulting representation. Its companion mechanism is the windowing scheme: non-overlapping windows of fixed length per trace, all windows from case records retained and an equal number randomly drawn from control records, yielding balanced training sets built from real data. The design choices include a kernel size half the window length, ReLU after the convolution, sigmoid output, Adam optimization, and binary cross-entropy loss, tuned on training and validation windows.","core_discovery":"The central claim is that raw fetal heart-rate traces, cut into fixed-length non-overlapping windows, contain enough temporal morphology to separate normal from abnormal birth outcomes with no handcrafted feature extraction. The training set is balanced by retaining all windows from abnormal cases and randomly drawing the same number of windows from normal controls, so the model sees only real signal segments. A 1DCNN with one convolutional layer of 20 filters, a kernel half the window length, one max-pooling layer, and two dense layers is trained directly on these windows. At a window size of 200 the authors report sensitivity 80% (95% CI 75% to 85%), specificity 79% (95% CI 73% to 84%), and AUC 0.86 (95% CI 0.81 to 0.91), the best among the five window sizes tested. The same setup outperforms an MLP on raw windows and beats SVM, random forest, and FLDA classifiers under identical conditions, as well as the authors' earlier SMOTE ensemble.","pith_inferences":["A testable extension follows from the label assumption: retraining the 200-window model with only the 18 pH-confirmed acidosis cases as abnormal would show whether the reported discrimination is fetal pathology or merely delivery mode.","The reported metrics are window-level, and each patient contributes many correlated windows, so a per-patient aggregation rule (for example, majority voting across windows) could change the apparent sensitivity and specificity and should be evaluated before clinical use.","The same recipe of real-data window balancing plus 1D convolution could transfer to other long, imbalanced physiological recordings, such as EEG seizure detection or continuous ECG monitoring, where handcrafted features and class skew are twin barriers."],"forward_implications":["Automated CTG screening could operate directly on raw monitor signals, bypassing the time-consuming and subjective step of extracting clinical features by hand.","Windowing supplies a real-data alternative to SMOTE, so balanced training sets are produced without synthetic minority samples and their associated bias.","Because the network learns features from data rather than from fixed rules, the same pipeline can be transferred to new CTG devices or new clinical questions with little re-engineering.","Window size is a genuine tuning parameter: the best separation in this study occurs at 200 samples, with both shorter and longer windows degrading performance.","On the same dataset, the windowed 1DCNN improves on the authors' previous SMOTE ensemble and on established classifiers including FLDA, random forest, SVM, and a plain MLP."],"supporting_citations":[{"why":"Supplies the open-access intrapartum cardiotocography database of 552 recordings used for training and testing.","marker":"[14]"},{"why":"Defines the SMOTE oversampling method that the windowing strategy is designed to replace.","marker":"[17]"},{"why":"The authors' earlier SMOTE-based ensemble on the same dataset, which the new windowing 1DCNN is compared against.","marker":"[18]"},{"why":"An automated fetal-heart-rate evaluation using random forest on the same database, used as a comparison point.","marker":"[12]"},{"why":"A nonlinear-feature SVM study on the same data, used as another comparison baseline.","marker":"[15]"}],"fun_headline_variants":["1D CNN on 200-point CTG windows hits AUC 0.86 for abnormal births","Raw 200-point CTG windows let a 1D CNN flag abnormal births at AUC 0.86","No handcrafted features: 1D CNN on 200-point CTG windows gets AUC 0.86","1D CNN with real 200-point CTG windows beats handcrafted features, AUC 0.86","AUC 0.86 from raw 200-point CTG windows with a single-layer 1D CNN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 24 caesarean deliveries with no recorded pathological outcome measures were performed because of underlying fetal pathology, even though the dataset curators give no reason; those 24 cases make up more than half of the abnormal class.","fun_headline_variants_meta":{"raw":{"variants":["1D CNN on 200-point CTG windows hits AUC 0.86 for abnormal births","Raw 200-point CTG windows let a 1D CNN flag abnormal births at AUC 0.86","No handcrafted features: 1D CNN on 200-point CTG windows gets AUC 0.86","1D CNN with real 200-point CTG windows beats handcrafted features, AUC 0.86","AUC 0.86 from raw 200-point CTG windows with a single-layer 1D CNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000857,"raw_usage":{"total_tokens":3784,"prompt_tokens":1073,"completion_tokens":2711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":2577}},"tokens_in":689,"tokens_out":2711,"duration_ms":19478,"temperature":1.0,"reasoning_tokens":2577,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:47:20.846968+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the 200-window 1DCNN using only the 18 deliveries with umbilical-artery pH $\\le 7.20$ as the abnormal class and an equal number of normal windows; if the window-level AUC stays near 0.86 the separation reflects real pathology, whereas a drop toward chance or a model that labels the 24 unconfirmed caesareans as normal would show that delivery-mode labels drove the reported result.","supporting_citations":[{"cited_title":"Open access intrapartum ctg database: Stepping stone towards generalization of technical findings on ctg signals,","cited_arxiv_id":null,"evidence_quote":"Supplies the open-access intrapartum cardiotocography database of 552 recordings used for training and testing."},{"cited_title":"Smote: synthetic minority over-sampling technique,","cited_arxiv_id":null,"evidence_quote":"Defines the SMOTE oversampling method that the windowing strategy is designed to replace."},{"cited_title":"Machine learning ensem- ble modelling to classify caesarean section and vaginal delivery types using cardiotocography traces,","cited_arxiv_id":null,"evidence_quote":"The authors' earlier SMOTE-based ensemble on the same dataset, which the new windowing 1DCNN is compared against."},{"cited_title":"Discriminating normal from abnormal pregnancy cases using an automated fhr evaluation method,","cited_arxiv_id":null,"evidence_quote":"An automated fetal-heart-rate evaluation using random forest on the same database, used as a comparison point."},{"cited_title":"Using nonlinear features for fetal heart rate classification,","cited_arxiv_id":null,"evidence_quote":"A nonlinear-feature SVM study on the same data, used as another comparison baseline."}],"review_version":1}