{"id":"0d28c056-f665-46de-96ea-ff7b0bc36567","arxiv_id":"2501.13587","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CPC with full fine-tuning yields a small but significant AUROC gain over target-only training at 5% target data in one of two tasks, and temporal patterns transfer better than point-of-care decisions.","lead":"Contrastive pre-training on source-hospital data, followed by full fine-tuning, slightly improves extubation-risk prediction on a small target dataset compared to training on that dataset alone. The paper also reports that temporal progression patterns transfer across institutions more easily than point-of-care decisions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The low-data transfer gain may stem from TTE-guided negative sampling, which injects extubation-timing information into pre-training; no random-negative or non-outcome ablation is reported.","rationale":"The reader's weakest-assumption analysis correctly identifies the missing ablation as the load-bearing issue. The paper itself acknowledges the modification is novel ('we incorporate this domain knowledge through a simple modification'), but provides no comparison to standard CPC, so the mechanism is untested. This is not a claim of misconduct; it is an unverified design choice in the central comparison. The clinical setup and evaluation across data regimes are credible, and the direct comparison to Source+FTF at 5% (0.736 vs 0.708) suggests that the effect is not merely from using source data. However, because TTE is the exact continuous variable underlying Task 2 and a plausible correlate of Task 1, the pre-training objective can shortcut to the target concept. The proposed ablation settles whether the reported transfer benefit is due to contrastive learning generally, to TTE-guided hard negatives, or to auxiliary-label leakage. Until that ablation is run, the paper's central claim is conditional, matching the reader's verdict. No change to the verdict is needed.","tokens_in":8015,"tokens_out":7999,"duration_ms":78922,"concrete_test":"Retrain CPC-FTF under three negative-sampling conditions: (a) TTE-guided as in the paper, (b) uniform random negatives, and (c) negatives guided by non-outcome similarity (age, PIM-3, diagnosis, vasoactive support). Keep all other hyperparameters and the 65/15/20 split fixed, use the same 5 seeds, and report Task 1 and Task 2 AUROC and AUPRC at 5% target data with paired significance tests. If condition (b) or (c) reproduces the 0.736 vs 0.716 advantage over Target-Only, the result does not depend on TTE guidance; if only (a) does, the paper must re-scope its claim to TTE-guided pretraining.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CPC pre-training plus full fine-tuning improves target-institution performance in low-data regimes, with the headline numbers being Task 1 AUROC 0.736 vs 0.716 at 5% target data. The pre-training loss in Eq. (4) uses Eq. (3) with m set to time-to-extubation (TTE) in Section 2.4, so negative samples are drawn preferentially from episodes with similar remaining ventilation time. Task 2's label (extubation within the next 12 hours) is a thresholded function of exactly that remaining time, and Task 1's extubation-failure label is correlated with weaning timing. The pretext task can therefore encode the downstream concept before fine-tuning, meaning the observed transfer benefit may reflect auxiliary-label leakage rather than the quality of contrastive representations. The manuscript does not ablate Eq. (3) against standard random-negative CPC, nor against a similarity measure built from non-outcome covariates, so the active ingredient in the reported benefit is unidentified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a cross-institutional transfer learning framework for pediatric ventilation management, using contrastive predictive coding (CPC) pretrained on a general PICU and then fine-tuned on a smaller cardiac-focused PICU. Three learning strategies are compared (target-only, full fine-tuning FTF, and decoder-only fine-tuning FTD) at 100%, 30%, and 5% of the target training data, for two tasks: point-of-care extubation failure prediction (Task 1) and prospective 12-hour extubation-window identification (Task 2). The main empirical claim is that CPC with full fine-tuning (CPC-FTF) improves low-data transfer performance, with a reported Task 1 AUROC gain of 0.736 versus 0.716 at 5% target data, while the secondary claim is that temporal progression patterns transfer more readily than point-of-care decisions. The pretraining loss uses a time-to-extubation (TTE) similarity measure for guided negative sampling.","tokens_in":8216,"tokens_out":2072,"duration_ms":21685,"significance":"If the claims hold, the paper would provide practical evidence on how self-supervised representations can support model deployment across pediatric intensive care units with different patient populations, which is a clinically important and understudied problem. The use of two real PICUs, multiple data-regime comparisons, several metrics, and multiple random seeds are strengths. However, the headline result rests on a small reported AUROC difference and on a pretraining sampling scheme that is directly tied to the downstream extubation outcome; without an ablation isolating that design choice, the core mechanism behind the claimed transfer benefit remains unidentified. The secondary asymmetry claim also lacks a formal comparison across tasks.","major_comments":[{"comment":"The TTE-guided negative sampling is a load-bearing design choice because TTE is not an outcome-free covariate: Task 2 labels extubation within the next 12 hours, a direct function of remaining ventilation time, and Task 1 predicts extubation failure, which is correlated with weaning timing. Thus the pretext task can encode the downstream target concept before fine-tuning, so the reported transfer gain may reflect auxiliary-label leakage rather than generally transferable representations. The paper does not ablate Eq. (3) against standard random-negative CPC, nor against a similarity measure built from non-outcome covariates (e.g., demographics, PIM score, diagnosis), so the active ingredient in the transfer benefit is unquantified. This needs to be addressed for the central claim to stand.","section":"Section 2.4, Eqs. (3)-(4)"},{"comment":"The headline Task 1 result of CPC-FTF (AUROC 0.736 ± 0.059) versus target-only (0.716 ± 0.047) has overlapping standard deviations, and the manuscript reports only a p-value threshold (p < 0.05) without describing the statistical test, whether it is paired across the five seeds, or how multiple comparisons over models, tasks, and data fractions are handled. With five seeds, a formal paired test and effect-size reporting are necessary to support the claim of significant improvement.","section":"Section 3, Table 2, 5% data row"},{"comment":"The claim that temporal progression patterns (Task 2) transfer more readily than point-of-care decisions (Task 1) is based on descriptive comparisons of baseline performance and transfer gaps. No statistical test is provided for the difference in transfer benefit between tasks, and the asymmetry could partly reflect the different base rates, label definitions, or the leakage mechanism in the TTE-guided sampling. A formal interaction test or an ablation separating outcome-related leakage from representation quality is needed before this secondary claim can be accepted.","section":"Section 3, 'Task-specific analysis'"},{"comment":"Restricting the analysis to the first extubation attempt per admission is a defensible independence choice, but it may also change the case mix and the difficulty of Task 2 by removing subsequent weaning windows; the manuscript does not discuss how this selection affects the transfer conclusions or whether results are robust to including later episodes with appropriate clustering.","section":"Section 2.2.1, inclusion criteria"}],"minor_comments":[{"comment":"The sentence reporting the direct transfer drop (AUROC from 0.785 to 0.709) cites p < 0.01 but does not specify the test or whether it compares Source-Only against Target-Only at 100% data; please clarify.","section":"Section 3, first paragraph"},{"comment":"The notation m(x_{\\le t}, x'_{\\le t}) is introduced but neither its exact functional form nor how it is computed from TTE is defined; the operationalization of the sampling probability in Eq. (3) should be stated explicitly.","section":"Section 2.3.2"},{"comment":"It is unclear how the temperature parameter beta in Eq. (3) was selected and whether it was tuned jointly with the other contrastive hyperparameters; please report the chosen value and its sensitivity.","section":"Section 2.4 and Appendix Table 3"},{"comment":"The hyperparameter table lists ranges but not the final selected configurations for the reported models; providing the actual settings would improve reproducibility.","section":"Appendix Table 3"},{"comment":"The phrase 'significantly outperforms' in the discussion of the 5% Task 1 result is stronger than the evidence warrants given the overlapping standard deviations and the absent test description; please temper the language accordingly.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The TTE-guided sampling concern is central and should be resolved with a random-negative ablation and, ideally, a non-outcome similarity control. If the authors show that the benefit persists without outcome-derived sampling, the paper could be a solid contribution; without that, the headline result is consistent with target leakage. The statistical reporting also needs strengthening before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's central claim—contrastive pre-training with CPC plus full fine-tuning transfers knowledge from a general PICU to a cardiac-focused PICU, with the biggest gain at 5% target data—is plausible but not established. The pre-training negative sampler (Eq. 3) uses time-to-extubation as the similarity function, and TTE is the same signal that defines both downstream tasks (extubation within 12 hours, and extubation failure within 48 hours). So the 'unsupervised' pre-training is doing weakly supervised learning on the target concept. The headline gain at 5% data is 0.736 vs 0.716 AUROC, which overlaps at one standard deviation, and the paper gives no statistical test details. Without a random-negative CPC ablation, you can't attribute the effect to representation quality.\n\nWhat is genuinely new: the application to pediatric ventilation across two real PICUs, with a systematic comparison of full vs decoder-only fine-tuning across data regimes. The asymmetry finding—temporal window identification transfers better than point-of-care extubation risk—is an interesting empirical pattern, and the clinical setup is sensible. I also credit the decision to analyze only first extubation attempts and to keep censored episodes in pre-training.\n\nThe soft spots are in proportion to the central claim. The leakage issue is load-bearing, not minor. The paper highlights TTE as a 'clinical relevance' measure, but doesn't acknowledge that it encodes outcome information. The result is a small effect under a questionable pretext task. Also, no code, no data, no external validation; the framework is essentially the McDermott et al. 2021 benchmark applied to a new setting, so novelty is incremental.\n\nWho should read this: anyone working on cross-institutional EHR transfer or on outcome leakage in self-supervised learning. It's a good case study for a reading group discussion of why negative sampling choices matter. A serious referee could fix the study with one random-negative baseline and proper statistical reporting. I'd send it to review rather than desk reject—the question is worth answering and the data is real—but I wouldn't cite it for the transfer claim until the ablation exists.","headline":"The transfer gain is probably inflated by outcome leakage in the negative sampling; the paper deserves a referee to test that, but it is not citable as-is.","tokens_in":8736,"tokens_out":3182,"would_cite":false,"duration_ms":27020,"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":"Contrastive pre-training on a general pediatric ICU transfers to a cardiac-focused unit, and full fine-tuning makes the difference at low data.","keywords":["contrastive predictive coding","cross-institutional transfer","pediatric ventilation","clinical time series","fine-tuning","extubation prediction","ICU knowledge transfer","self-supervised learning"],"falsifier":"Run the same transfer protocol with guided negatives replaced by random negatives; if the AUROC advantage of CPC-FTF over target-only training at 5% data disappears or reverses, the claimed transfer benefit is an artifact of the guided sampling. A second check is to replace time-to-extubation by a clinically irrelevant similarity such as gender and see whether the gain persists.","tokens_in":7836,"feed_emoji":"🩺","tokens_out":3682,"duration_ms":31024,"temperature":0.7,"pith_summary":"The paper claims that contrastive predictive coding (CPC) pre-trained on a large general pediatric ICU can transfer useful knowledge to a smaller cardiac-focused ICU, provided the full model is fine-tuned on target data. This is demonstrated in pediatric ventilation management, where direct transfer of a source-trained model loses substantial AUROC on both prediction tasks. The strongest evidence is a low-data gain: with 5% of target data, CPC plus full fine-tuning reaches Task 1 AUROC 0.736 versus 0.716 for target-only training, and matches full-data target-only performance in Task 2. The paper also finds that continuous temporal progression tasks transfer more readily than point-of-care extubation-risk decisions, suggesting that cross-institutional deployment could begin with monitoring tasks rather than high-stakes decisions.","feed_headline":"Contrastive pre-training lifts ICU transfer at 5% data","feed_subtitle":"With full fine-tuning, a source-ICU model reaches AUROC 0.736 vs 0.716 for training from scratch on a cardiac unit.","key_machinery":"The central object is guided negative sampling inside the InfoNCE loss of Contrastive Predictive Coding. Standard CPC draws negative samples uniformly, while this framework biases negative pairs by clinical relevance $m(\\cdot,\\cdot)$, instantiated as time-to-extubation, so pre-training learns to separate clinically similar but physiologically distinct patients. The encoder $f_\\theta$ is an MLP feature extractor plus a GRU autoregressive model, and the downstream adaptation is either full fine-tuning (FTF) or decoder-only fine-tuning (FTD) of this encoder with a task-specific linear head.","core_discovery":"The paper claims that self-supervised contrastive pre-training on source-institution EHR time series, followed by fine-tuning all parameters, reduces the cross-institutional performance gap that direct transfer suffers. In the five-percent target-data regime, the pretrained and fully fine-tuned model outperforms training from scratch for point-of-care extubation risk (AUROC 0.736 vs 0.716, p<0.05), and in the window-identification task it reaches parity with target-only training at full data. Conversely, decoder-only fine-tuning fails in the data-poor regime (Task 1 AUROC drops to 0.510), showing that adapting the feature extractor matters. The paper interprets this pattern as evidence that general physiological representations transfer across institutions, but institution-specific decision thresholds do not.","pith_inferences":["Because the guided negative sampling uses time-to-extubation as a clinical similarity proxy, the reported 5%-data gain may partly reflect leakage of the task concept into pre-training; an ablation against standard CPC with random negatives is the natural test.","The observed asymmetry between temporal progression and point-of-care decisions likely generalizes to other readiness prediction settings: trend and window prediction should be evaluated before high-stakes point decisions.","Similarity-guided negative sampling could be ported to other clinical domains that have a natural clinical-stage proxy, such as time to sepsis onset or ICU length-of-stay.","If representation quality scales with source diversity, pretraining on a larger multi-unit source cohort would further improve few-shot target performance."],"forward_implications":["Smaller specialized units can train competitive models with very little local data by fine-tuning a source-ICU pretrained encoder.","Decoder-only transfer with fixed representations is a poor default; feature extractors need to be updated for the target institution.","Continuous monitoring-type tasks are the safer first deployment for cross-institutional clinical transfer.","Direct source-to-target model deployment is not a viable shortcut for institutions with different patient populations.","The transfer benefit is concentrated in low-data settings; at 30% and 100% target data, pretraining does not hurt but does not help much."],"supporting_citations":[{"why":"Supplies the CPC framework and InfoNCE loss that the pre-training phase is built on.","marker":"[Oord et al., 2018]"},{"why":"Provides the transfer-learning evaluation setup (Target-Only, FTF, FTD) and the baseline architecture.","marker":"[McDermott et al., 2021]"},{"why":"Provides the hourly aggregation and imputation preprocessing pipeline used to build the ventilation time series.","marker":"[Wang et al., 2020]"},{"why":"Defines extubation failure as reintubation within 48 hours, the outcome for Task 1.","marker":"[Thille et al., 2011]"},{"why":"Motivates contrastive learning for medical time series and links it to clinical reasoning.","marker":"[Liu et al., 2023]"},{"why":"Supports the applicability of CPC to EHR time-series data.","marker":"[Bouchattaoui et al., 2024]"}],"fun_headline_variants":["Contrastive pretraining boosts cross-hospital ICU transfer at 5% data","With full fine-tuning, contrastive pretraining beats scratch on 5% ICU data","General ICU physiology transfers; decision thresholds don't","At 5% target data, contrastive pretraining closes ICU gap","Pretrained ICU models need only 5% fine-tuning data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on time-to-extubation being a valid measure of clinical similarity for choosing contrastive negatives, so that the pre-training objective encodes the same concept the downstream tasks test.","fun_headline_variants_meta":{"raw":{"variants":["Contrastive pretraining boosts cross-hospital ICU transfer at 5% data","With full fine-tuning, contrastive pretraining beats scratch on 5% ICU data","General ICU physiology transfers; decision thresholds don't","At 5% target data, contrastive pretraining closes ICU gap","Pretrained ICU models need only 5% fine-tuning data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000546,"raw_usage":{"total_tokens":2575,"prompt_tokens":875,"completion_tokens":1700,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":1607}},"tokens_in":491,"tokens_out":1700,"duration_ms":12232,"temperature":1.0,"reasoning_tokens":1607,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:47:08.438018+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same transfer protocol with guided negatives replaced by random negatives; if the AUROC advantage of CPC-FTF over target-only training at 5% data disappears or reverses, the claimed transfer benefit is an artifact of the guided sampling. A second check is to replace time-to-extubation by a clinically irrelevant similarity such as gender and see whether the gain persists.","supporting_citations":[{"cited_title":"Mimic-extract: A data extraction, preprocessing, and representation pipeline for mimic-iii","cited_arxiv_id":null,"evidence_quote":"Provides the hourly aggregation and imputation preprocessing pipeline used to build the ventilation time series."},{"cited_title":"Out- comes of extubation failure in medical intensive care unit patients","cited_arxiv_id":null,"evidence_quote":"Defines extubation failure as reintubation within 48 hours, the outcome for Task 1."},{"cited_title":"Causal Contrastive Learning for Counterfactual Regression Over Time","cited_arxiv_id":"2406.00535","evidence_quote":"Supports the applicability of CPC to EHR time-series data."}],"review_version":1}