{"id":"8168e040-646c-414d-87a7-025fc464152c","arxiv_id":"2606.20172","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"ML pipeline on fetal MRI features predicts gestational age at birth (R²=0.13, MAE=2.74 weeks) and preterm status (accuracy 0.77) from 426 cases.","lead":"A machine learning pipeline using multi-modal fetal MRI data predicts gestational age at birth with R² of 0.13 and classifies term versus preterm with 0.77 accuracy. This offers a proof-of-concept approach for using advanced imaging to flag preterm birth risk in pregnancy care.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Small preterm cohort (n=93) plus R²=0.13 leaves regression stability and generalization unproven under internal 10-fold CV only","rationale":"The reader's weakest_assumption matches the load-bearing statistical limitation exactly; the abstract metrics already make the concern visible, and the acknowledged need for larger cohorts confirms it. No stronger internal inconsistency appears from the given performance numbers.","tokens_in":1815,"tokens_out":295,"duration_ms":17891,"concrete_test":"Reserve a random 30-case preterm hold-out set before any CV or feature selection; retrain the full pipeline on the remaining 396 cases and report R²/MAE on the untouched hold-out. If R² falls below 0.05 or MAE exceeds 3.5 weeks, the current CV numbers do not support stable generalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central regression claim rests on predicting continuous GA at birth from 426 total cases (333 term + 93 preterm). With bespoke imputation + feature selection inside each fold, the effective sample for the minority class is small; the reported R²=0.13 shows the selected features (cervical length, placental T2* stats) capture little variance. Stratified 10-fold CV on this single cohort does not test whether the fitted mapping holds under site/protocol shift or finer preterm stratification, which the authors themselves flag as future work.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript describes development of a bespoke ML pipeline (imputation, feature selection, regression) to predict continuous gestational age at birth from multi-modal fetal MRI in 426 pregnancies (333 term + 93 preterm). Predictions are then thresholded into term/preterm classes. Using stratified 10-fold CV the pipeline reports R²=0.13, MAE=2.74 weeks, accuracy=0.77, sensitivity=0.59, specificity=0.82; cervical length and placental T2* statistics dominate the selected features. An ablation study is included and code is released.","tokens_in":1943,"tokens_out":531,"duration_ms":24868,"significance":"If the modest performance holds under external validation, the work supplies a first proof-of-concept for regression-based GA prediction (rather than binary classification) in the preterm setting using MRI. The ablation study and public code release are positive for reproducibility. The low R² nevertheless constrains immediate clinical utility and underscores the need for larger, multi-site cohorts.","major_comments":[{"comment":"Abstract and Results: the headline R²=0.13 indicates that the selected features explain only 13 % of GA variance; this low value, obtained on only 93 preterm cases, is load-bearing for the claim that the pipeline constitutes a viable prediction method rather than a marginal signal.","section":"Abstract / Results"},{"comment":"Methods (cross-validation description) and Discussion: stratified 10-fold CV is performed entirely within a single cohort; no external-site or protocol-shift test is reported, leaving open whether the mapping from cervical length / T2* statistics to GA generalizes beyond the present imaging protocol and population.","section":"Methods / Discussion"},{"comment":"Results (performance reporting): neither per-fold standard deviations nor confidence intervals accompany the aggregate R², MAE, or classification metrics, so stability of the regression on the minority (preterm) class cannot be assessed from the given numbers.","section":"Results"}],"minor_comments":[{"comment":"The exact decision threshold used to convert the continuous GA prediction into the binary term/preterm label is not stated; adding this detail would improve reproducibility.","section":"Methods"}],"recommendation":"major_revision","confidential_remarks":"The modest R² and single-cohort design make the contribution borderline for a high-impact venue; the editor may wish to weigh whether the proof-of-concept framing and code release are sufficient to offset the limited explanatory power."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the thoughtful review and constructive feedback. We address each major comment below, proposing revisions where feasible while acknowledging inherent limitations of the single-cohort design.","responses":[{"response":"We agree that an R² of 0.13 reflects modest explanatory power and that the preterm subsample (n=93) constrains stronger claims. The manuscript already frames the work as a proof-of-concept for regression-based GA prediction (rather than binary classification) in this setting, supported by the ablation study and feature interpretability. We will revise the abstract and discussion to more explicitly qualify the performance as preliminary and to avoid implying immediate clinical viability, while retaining the emphasis on the novel regression framing and public code release.","revision_made":"yes","referee_comment":"[Abstract / Results] Abstract and Results: the headline R²=0.13 indicates that the selected features explain only 13 % of GA variance; this low value, obtained on only 93 preterm cases, is load-bearing for the claim that the pipeline constitutes a viable prediction method rather than a marginal signal."},{"response":"We concur that internal stratified 10-fold CV alone cannot establish generalizability across sites or protocols. The current study is limited to a single-center cohort acquired under one imaging protocol; external validation would require additional multi-site data that are not available. We will expand the discussion to explicitly state this limitation and to underscore the need for future multi-center cohorts, but cannot perform external testing within the scope of the present manuscript.","revision_made":"partial","referee_comment":"[Methods / Discussion] Methods (cross-validation description) and Discussion: stratified 10-fold CV is performed entirely within a single cohort; no external-site or protocol-shift test is reported, leaving open whether the mapping from cervical length / T2* statistics to GA generalizes beyond the present imaging protocol and population."},{"response":"We thank the referee for this observation. We will update the results section and tables to report per-fold standard deviations and bootstrap confidence intervals for all metrics (R², MAE, accuracy, sensitivity, specificity), allowing readers to evaluate stability particularly on the preterm subset.","revision_made":"yes","referee_comment":"[Results] Results (performance reporting): neither per-fold standard deviations nor confidence intervals accompany the aggregate R², MAE, or classification metrics, so stability of the regression on the minority (preterm) class cannot be assessed from the given numbers."}],"tokens_in":1493,"tokens_out":569,"duration_ms":11705,"standing_objections":["Absence of external-site or protocol-shift validation, which cannot be addressed without new multi-center data collection beyond the current single-cohort study."]},"desk_editor":{"model":"grok-4.3","letter":"The core result here is a supervised pipeline that regresses gestational age at birth from multi-modal fetal MRI (morphological plus functional) on 426 pregnancies, 93 of them preterm. It reports R²=0.13, MAE=2.74 weeks, and after binning reaches 0.77 accuracy, 0.59 sensitivity, 0.82 specificity on stratified 10-fold CV. Cervical length and placental T2* statistics dominate the selected features. An ablation is included and the code is on GitHub.\n\nWhat is new is the explicit regression framing instead of the usual binary preterm classification; the abstract states this had not been tried before with this data type. The work also handles imputation and feature selection in a way that keeps them inside the CV loop, which is better than many similar papers.\n\nThe numbers themselves are the main limitation. R²=0.13 shows the chosen features capture little of the outcome variance, and the preterm arm is small enough that stability under any shift in site or protocol remains untested. All evaluation stays within one cohort; the authors correctly flag larger data and finer stratification as next steps.\n\nThis is for perinatal imaging and obstetrics ML groups who want to see an initial continuous-prediction attempt and can reuse the released code. It is not yet strong enough for clinical claims.\n\nSend it to peer review. The methods are transparent, the ablation and code release give referees something concrete to check, and the low performance is reported plainly rather than hidden.","headline":"Proof-of-concept regression for continuous GA at birth from fetal MRI hits R²=0.13 on n=93 preterm cases under internal CV only, with code released.","tokens_in":2437,"tokens_out":388,"would_cite":false,"duration_ms":15027,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A machine learning pipeline using multi-modal fetal MRI predicts gestational age at birth to within 2.74 weeks on average.","keywords":["preterm birth","gestational age prediction","fetal MRI","machine learning","placental T2*","cervical length","multi-modal imaging","regression model"],"falsifier":"Applying the same pipeline to an independent set of pregnancies scanned under different protocols and obtaining a mean absolute error above 4 weeks or classification accuracy below 65 percent.","tokens_in":2724,"feed_emoji":"🤰","tokens_out":660,"duration_ms":38446,"temperature":0.7,"pith_summary":"The paper develops and tests a pipeline of machine learning steps for imputing missing values, selecting features, and performing regression to forecast the gestational age at which birth will occur. It draws on comprehensive MRI scans from 333 term pregnancies and 93 preterm ones to move beyond simple yes-no classification of preterm risk. Cervical length and statistics from placental T2* measurements stand out as the strongest predictors in the model. Accurate timing information matters for scheduling care that protects newborns in high-risk pregnancies. The authors frame the work as an initial proof of concept that continuous prediction from such scans is feasible.","feed_headline":"MRI pipeline predicts birth timing to 2.74 weeks error","feed_subtitle":"Cervical length and placental T2* features yield 77 percent accuracy distinguishing preterm from term cases in 426 pregnancies.","key_machinery":"A custom pipeline of machine learning methods for data imputation, feature selection, and regression applied to multi-modal fetal MRI data.","core_discovery":"The authors present a pipeline that imputes missing data, selects features, and applies regression to multi-modal fetal MRI to predict gestational age at birth. On 426 cases evaluated with stratified 10-fold cross-validation, the model reaches an R2 score of 0.13 and a mean absolute error of 2.74 weeks for the continuous GA prediction, while achieving 0.77 accuracy, 0.59 sensitivity, and 0.82 specificity for term versus preterm classification. Cervical length and placental T2* statistics are the predominant selected features. This supplies the first regression treatment of preterm birth from multi-modal MRI rather than classification alone.","pith_inferences":["If performance holds on new data, the predictions could guide decisions on when to administer antenatal steroids or plan delivery.","Adding routine clinical variables to the MRI features might raise the modest R2 value.","External validation across scanner vendors would reveal whether the selected features remain stable."],"forward_implications":["Cervical length and placental T2* values carry usable information about birth timing.","Regression on continuous gestational age extends beyond binary preterm classification.","Larger cohorts will support finer stratification inside the preterm group.","Fast motion-robust multi-modal MRI makes repeated fetal imaging practical for this purpose."],"fun_headline_variants":["Fetal MRI pipeline predicts birth to 2.74 weeks error","Cervical length drives multi-modal MRI birth prediction","Fetal MRI regression forecasts GA at birth at 0.77 accuracy","Placental T2* features aid preterm birth timing model"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The 93 preterm cases supply enough signal to train a regression model that generalizes beyond this cohort and imaging protocol.","fun_headline_variants_meta":{"raw":{"variants":["Fetal MRI pipeline predicts birth to 2.74 weeks error","Cervical length drives multi-modal MRI birth prediction","Fetal MRI regression forecasts GA at birth at 0.77 accuracy","Placental T2* features aid preterm birth timing model"]},"model":"grok-4.3","cost_usd":0.004555,"raw_usage":{"total_tokens":2324,"prompt_tokens":789,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":45549500,"prompt_tokens_details":{"text_tokens":789,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1467,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":789,"tokens_out":68,"duration_ms":13411,"temperature":1.0,"reasoning_tokens":1467,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T18:08:51.418253+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Applying the same pipeline to an independent set of pregnancies scanned under different protocols and obtaining a mean absolute error above 4 weeks or classification accuracy below 65 percent.","supporting_citations":[],"review_version":1}