{"id":"91aca987-01a2-42fd-b4cb-6f770fdab510","arxiv_id":"2606.20823","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"A YOLO keypoint model trained on 37k+ public images plus 1k neonatal frames achieves SOTA NME and low failure rates for 68-point neonatal landmark detection in clinical conditions.","lead":"The paper trains a YOLO-based model on mixed public and neonatal face images to locate 68 facial landmarks on newborns in hospital settings. This could support automated systems that assess pain from facial expressions without physical contact.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Standardization of 37k heterogeneous adult landmark sets to 68-point markup may embed annotation inconsistencies that explain the reported pre-fine-tune generalization rather than true domain transfer.","rationale":"The reader's weakest_assumption directly identifies the data-standardization step as load-bearing; the full-text description does not add verifiable detail that would remove the risk, so the concern remains the single most material uncertainty for the generalization claim.","tokens_in":1927,"tokens_out":363,"duration_ms":15241,"concrete_test":"Release the exact per-dataset landmark remapping rules (or code) used for the 11 public sets; then recompute NME/DFR on the clinical test set after (a) replacing the standardized adult labels with a small manually re-annotated subset (~200 images) and (b) ablating the adult data entirely. If DFR rises above 15% in either case, the generalization result is sensitive to the standardization step.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline claim of strong generalization (DFR=5.3% on unseen clinical neonatal frames before fine-tuning) rests on the mixed training distribution having a small domain gap. The 37,459 public images come from 11 datasets whose original landmark definitions differ; the paper states they were 'standardised to 68-point markup' but supplies no explicit mapping, inter-annotator agreement, or verification that the resulting labels remain geometrically consistent with neonatal anatomy. With only 1,123 native neonatal frames, any systematic offset introduced by the adult-to-68 conversion would be absorbed into the learned regressor and could produce the observed low DFR without the model actually solving the neonatal-specific challenges (occlusions, motion blur, head pose).","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces NeoLoc-68, claimed as the first end-to-end 68-point facial landmark detector for neonatal faces in clinical environments. It trains a YOLO-based keypoint regressor on 37,459 images from 11 public (mostly adult) datasets standardized to 68-point markup plus 1,123 manually annotated neonatal frames, initialized from pretrained neonatal face detector weights. Reported results include SOTA on public datasets (NME=5.37, FR=12.5%, AUC0.08=38.00%) and strong generalization on a held-out clinical neonatal test set (DFR=5.3% before fine-tuning; after fine-tuning NME=6.36, FR=22.30%, DFR=1.77%).","tokens_in":2106,"tokens_out":552,"duration_ms":30555,"significance":"If the generalization results hold after addressing verification gaps, the work would be significant for enabling non-contact neonatal pain assessment via facial expression analysis under real clinical conditions (occlusions, motion blur, varied poses). The mixed-dataset training strategy and multi-metric evaluation (including DFR and AUC) are positive aspects; however, the absence of protocol details limits immediate utility for downstream tasks.","major_comments":[{"comment":"Data section: The standardization of 37,459 images from 11 heterogeneous public datasets to a common 68-point markup is described at a high level but supplies no explicit landmark mapping, inter-annotator agreement, or geometric consistency checks against neonatal anatomy. This is load-bearing for the central generalization claim (DFR=5.3% on unseen clinical neonatal frames before fine-tuning), as systematic offsets from adult-to-68 conversion could be absorbed into the regressor without true domain transfer.","section":"Data section"},{"comment":"Experiments section: The manuscript reports numeric performance metrics, SOTA claims, and baseline comparisons but provides no experimental protocol, train-test split information, baseline implementation details, or error analysis. This prevents verification of the reported results (e.g., pre-fine-tune DFR=5.3% and post-fine-tune metrics) and undermines the soundness of the generalization statements.","section":"Experiments section"}],"minor_comments":[{"comment":"Abstract: The parenthetical '(totalling over 76,000 landmarks)' appears to count only the 1,123 neonatal frames (≈76k landmarks) while omitting the contribution from the 37,459 public images; this is a minor arithmetic/presentation inconsistency.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive comments, which highlight important areas for improving the clarity and verifiability of our work. We address each major comment point-by-point below and commit to revisions that will strengthen the manuscript without altering its core claims.","responses":[{"response":"We acknowledge that the Data section currently provides only a high-level description of the standardization process. To directly address this concern and support the generalization results, the revised manuscript will include an expanded Data section with explicit per-dataset landmark mappings to the 68-point format, available inter-annotator agreement statistics for the manually annotated neonatal frames, and a description of the geometric consistency checks applied during standardization to align with neonatal anatomy. These additions will clarify that the reported pre-fine-tuning DFR reflects genuine domain transfer rather than absorbed offsets.","revision_made":"yes","referee_comment":"[Data section] The standardization of 37,459 images from 11 heterogeneous public datasets to a common 68-point markup is described at a high level but supplies no explicit landmark mapping, inter-annotator agreement, or geometric consistency checks against neonatal anatomy. This is load-bearing for the central generalization claim (DFR=5.3% on unseen clinical neonatal frames before fine-tuning), as systematic offsets from adult-to-68 conversion could be absorbed into the regressor without true domain transfer."},{"response":"We agree that the absence of detailed experimental protocol information limits independent verification. In the revised manuscript, we will add a new Experimental Protocol subsection that specifies the train-test splits used for the public datasets and the held-out neonatal test set, full implementation details and hyperparameters for all baselines, the exact training procedure including initialization from the neonatal face detector, and an error analysis with qualitative examples of failure cases. This will enable full reproduction and verification of all reported metrics, including the pre- and post-fine-tuning DFR, NME, FR, and AUC values.","revision_made":"yes","referee_comment":"[Experiments section] The manuscript reports numeric performance metrics, SOTA claims, and baseline comparisons but provides no experimental protocol, train-test split information, baseline implementation details, or error analysis. This prevents verification of the reported results (e.g., pre-fine-tune DFR=5.3% and post-fine-tune metrics) and undermines the soundness of the generalization statements."}],"tokens_in":1618,"tokens_out":507,"duration_ms":23004,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"This paper delivers the first reported end-to-end 68-point facial landmark detector aimed at neonatal clinical images. It adapts a YOLO keypoint model, starts from neonatal face detector weights, and trains on 37,459 standardized adult images plus 1,123 neonatal frames. The headline result is a 5.3% detection failure rate on held-out clinical neonatal test frames before any fine-tuning, with further gains afterward.\n\nThe work fills a clear practical gap. Adult landmark models struggle with medical equipment occlusions, motion blur, and infant head poses, and the authors correctly flag this for downstream neonatal pain scoring. Combining multiple public datasets with native neonatal annotations and reporting competitive numbers on both public and clinical sets is a reasonable step.\n\nThe soft spot sits in the data preparation. Standardizing 37k images from 11 adult datasets to a single 68-point scheme is described but not broken down—no mapping rules, no inter-annotator checks, no confirmation that the converted labels stay geometrically consistent with neonatal proportions. With adult data outnumbering neonatal frames by more than 30 to 1, any systematic offset in those converted labels could be absorbed into the regressor and produce the observed low failure rate without the model actually solving neonatal-specific problems. The abstract also omits train-test splits, baseline implementation details, and error analysis, so the SOTA claims on public data and the generalization claim cannot be checked from what is shown.\n\nThis is for researchers building automated neonatal monitoring or pain assessment pipelines. Readers who need a starting 68-point model for infant faces in real wards will find the numbers and the mixed-training approach useful.\n\nIt deserves peer review. The targeted clinical need and the new performance numbers on an under-served domain are enough to justify referee time, provided the data standardization and experimental protocol receive close scrutiny.","headline":"First claimed end-to-end 68-point neonatal landmark model via mixed adult-neonatal data shows reported pre-fine-tune generalization, but adult label standardization lacks verification details.","tokens_in":2586,"tokens_out":452,"would_cite":false,"duration_ms":16083,"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":"Mixing standardized adult face images with neonatal frames trains the first end-to-end model to locate 68 landmarks on babies in clinical settings.","keywords":["facial landmark detection","neonatal faces","clinical environments","68-point landmarks","YOLO keypoint model","pain assessment","generalization","end-to-end detection"],"falsifier":"A fresh clinical neonatal test set annotated by separate experts on which the model shows detection failure rates substantially above the reported levels would show the generalization claim does not hold.","tokens_in":2854,"feed_emoji":"👶","tokens_out":686,"duration_ms":23493,"temperature":0.7,"pith_summary":"The paper develops a detector that places 68 specific points on newborn faces even when the infants are in hospital environments surrounded by equipment and showing sudden movements. It standardizes over 37,000 images from public adult datasets to a common landmark scheme and combines them with over 1,000 manually labeled neonatal frames to train a regression model. The system starts from weights of a neonatal face detector and uses a keypoint approach to output the points directly. This step matters because accurate landmarks form the basis for automated systems that read facial expressions to judge pain without touching the baby. If the approach holds, it provides a foundation for contact-free monitoring tools that clinicians could use in neonatal care.","feed_headline":"Model locates 68 points on neonatal faces in clinics","feed_subtitle":"Standardized adult images mixed with neonatal frames let it generalize to hospital conditions with low failure rates.","key_machinery":"The YOLO-based keypoint model adapted to regress 68 facial landmarks, initialized with weights from a pretrained neonatal face detector.","core_discovery":"The paper claims that standardizing 37,459 single-face images from 11 public datasets to a 68-point markup and mixing them with 1,123 annotated neonatal frames produces a training set that lets an adapted YOLO-based keypoint model regress 68 landmarks on neonatal faces in clinical conditions, delivering the lowest detection failure rate among tested baselines on a clinical test set before fine-tuning and further gains after fine-tuning, establishing the first end-to-end 68-point neonatal model.","pith_inferences":["Similar mixing of public adult data with small specialized sets could reduce the need for large new annotations in other medical imaging domains with limited samples.","The landmark outputs could support video-based tracking of expression changes over time rather than single frames.","Deployment in neonatal units might allow continuous non-contact observation of distress signals during routine care."],"forward_implications":["The model reaches state-of-the-art numbers on public face datasets using the standardized 68-point markup.","It generalizes to clinical neonatal images with the lowest failure rate among baselines even before any neonatal fine-tuning.","Fine-tuning on additional neonatal frames further lowers error rates and failure rates.","The outputs can feed directly into downstream neonatal health monitoring and facial expression analysis tasks."],"fun_headline_variants":["YOLO regresses 68 neonatal facial landmarks in clinics","Mixed datasets enable 68-point neonatal face localization","End-to-end 68-point neonatal landmark model in hospitals","68 neonatal landmarks localized with low clinical failure rate"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That standardizing adult images to a shared 68-point scheme and mixing them with limited neonatal frames creates a training distribution whose gap to real clinical neonatal images remains small enough for the observed generalization to hold.","fun_headline_variants_meta":{"raw":{"variants":["YOLO regresses 68 neonatal facial landmarks in clinics","Mixed datasets enable 68-point neonatal face localization","End-to-end 68-point neonatal landmark model in hospitals","68 neonatal landmarks localized with low clinical failure rate"]},"model":"grok-4.3","cost_usd":0.005,"raw_usage":{"total_tokens":2520,"prompt_tokens":825,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":49999500,"prompt_tokens_details":{"text_tokens":825,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1634,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":825,"tokens_out":61,"duration_ms":13322,"temperature":1.0,"reasoning_tokens":1634,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T18:11:20.107289+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A fresh clinical neonatal test set annotated by separate experts on which the model shows detection failure rates substantially above the reported levels would show the generalization claim does not hold.","supporting_citations":[],"review_version":1}