{"id":"2ab5a5be-684e-471c-bf67-4e6f4e6810f8","arxiv_id":"2606.20449","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"YOLOv11m model fine-tuned on neonatal infant videos reaches AP50 of 0.96 for face detection, outperforming three general face detectors.","lead":"This paper develops a YOLOv11m-based model for detecting infant faces in neonatal clinical environments, first trained on general face datasets then fine-tuned on 228 videos from 113 infants. It could support non-contact video monitoring of infant pain, breathing, and distress in hospitals.","discovery_kind":"unclear","skeptic_critique":{"model":"grok-4.3","headline":"Neonatal dataset (113 infants, 114 sessions) may lack diversity and proper held-out evaluation, risking overstated domain-adaptation gains to AP50 0.96","rationale":"Reader's weakest assumption directly identifies the same load-bearing point; full-text access does not remove the small-N and split-verification risk for the adaptation claim.","tokens_in":1741,"tokens_out":323,"duration_ms":15525,"concrete_test":"In the methods or supplementary material, locate the exact train/validation/test split protocol for the 228 videos; if it is a per-frame or per-video random split without infant- or session-level blocking, re-run the fine-tuning with leave-one-infant-out cross-validation and report the mean AP50 on the held-out infants. A drop below 0.90 would indicate the original 0.96 is not generalizable.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline result is the jump from AP50 0.87 (public-data model) to 0.96 after fine-tuning on the 228-video neonatal set. This requires (a) the 113-infant corpus to be representative of the target clinical distribution and (b) the reported metric to be computed on data unseen during fine-tuning. With only one session per infant on average and no public neonatal benchmark, both conditions rest on unverified assumptions about diversity, session independence, and train/test partitioning. If the split leaks infant identity or clinical condition, the 0.09 gain is not evidence of robust adaptation.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes a YOLOv11m-based one-stage detector for infant faces in neonatal clinical settings. It first trains on the union of VGGFace2, CelebA, FDDB and WIDER FACE, reporting AP50 = 0.87 that exceeds three unnamed general-purpose face detectors; it then fine-tunes on a private neonatal corpus of 228 videos from 114 sessions of 113 infants and reports an improved AP50 of 0.96. The work notes the absence of public neonatal benchmarks and calls for their creation under appropriate privacy safeguards.","tokens_in":1881,"tokens_out":492,"duration_ms":16400,"significance":"A reliably validated domain-adapted detector would directly support downstream non-contact neonatal monitoring tasks (pain scoring, cardiorespiratory extraction, apnoea alerts). The reported 0.09 AP50 gain is potentially impactful, but its significance cannot be assessed without evidence that the fine-tuning evaluation used infant-disjoint, session-disjoint held-out data.","major_comments":[{"comment":"Abstract: the headline claim that fine-tuning raises AP50 from 0.87 to 0.96 is load-bearing for the central contribution, yet the abstract (and, by the provided text, the manuscript) supplies no information on train/test partitioning of the 228-video neonatal set, whether the 114 sessions were split by infant identity, or any cross-validation protocol. Without these details the 0.09 gain cannot be distinguished from overfitting or identity leakage.","section":"Abstract"},{"comment":"Abstract: the statement that the pre-fine-tuning model 'surpass[es] the performance of three state-of-the-art general face detectors' is unsupported by any named baselines, per-detector AP50 scores, or evaluation protocol on the public datasets, preventing verification of the claimed superiority.","section":"Abstract"}],"minor_comments":[{"comment":"Abstract: the phrase '114 recording sessions of 113 independent infants' is internally inconsistent (one extra session) and should be clarified.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The neonatal corpus is not released; while privacy constraints are legitimate, the lack of even a de-identified public benchmark or detailed split statistics makes independent reproduction impossible and weakens the paper's reproducibility claims."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on our manuscript. The comments highlight the need for greater transparency in our evaluation protocol and baseline comparisons, which we will address through revisions.","responses":[{"response":"We agree that explicit details on the train/test partitioning are required to substantiate the reported performance gain and rule out identity leakage or overfitting. The neonatal corpus comprises 228 videos from 114 sessions of 113 independent infants. In the revision we will add a dedicated subsection describing the partitioning protocol, confirming that the splits are infant-disjoint (and session-disjoint where relevant) and specifying the cross-validation procedure employed.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the headline claim that fine-tuning raises AP50 from 0.87 to 0.96 is load-bearing for the central contribution, yet the abstract (and, by the provided text, the manuscript) supplies no information on train/test partitioning of the 228-video neonatal set, whether the 114 sessions were split by infant identity, or any cross-validation protocol. Without these details the 0.09 gain cannot be distinguished from overfitting or identity leakage."},{"response":"We acknowledge that the manuscript does not name the three general face detectors, report their individual AP50 scores, or detail the evaluation protocol on the combined public datasets. In the revision we will explicitly identify the detectors, provide their per-detector AP50 values on the union of VGGFace2, CelebA, FDDB and WIDER FACE, and describe the evaluation protocol used.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the statement that the pre-fine-tuning model 'surpass[es] the performance of three state-of-the-art general face detectors' is unsupported by any named baselines, per-detector AP50 scores, or evaluation protocol on the public datasets, preventing verification of the claimed superiority."}],"tokens_in":1426,"tokens_out":419,"duration_ms":27029,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core of this paper is a practical fine-tune of YOLOv11m on neonatal clinical video. They first train on a mix of public face datasets to reach AP50 0.87, beating a few general detectors, then adapt on 228 videos from 114 sessions of 113 infants and report 0.96. That is the headline result.\n\nWhat works is the problem framing. Neonatal face detection really does face extra issues from equipment, lighting shifts, and partial occlusions that standard models miss. Starting with a one-stage detector and doing domain adaptation is a reasonable choice for real-time clinical use. The closing call for more public neonatal datasets with proper ethics is also fair.\n\nThe weak part is verification. The abstract gives no information on how the neonatal data was split, whether infants or sessions were held out, or any error breakdown. With roughly two videos per infant, it is easy for identity or condition leakage to produce an inflated number. Without those controls the 0.09 gain cannot be read as robust adaptation. The dataset itself is not released, so the claim stays hard to check.\n\nThis is for people building camera-based neonatal monitoring tools or doing narrow-domain face detection. A reader who needs an example of applying current detectors to a medical niche can extract the setup, but the work does not move the broader literature much.\n\nSend it to review if the full manuscript supplies the missing protocol and split details; the motivation is real and the base approach is sound even if the numbers need scrutiny.","headline":"Straightforward YOLOv11 fine-tuning on a 113-infant neonatal set claims a solid AP50 lift to 0.96, but missing split and leakage details make the gain hard to trust.","tokens_in":2344,"tokens_out":397,"would_cite":false,"duration_ms":19371,"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 YOLOv11m model fine-tuned on neonatal videos detects infant faces in clinical settings at 0.96 AP50.","keywords":["infant face detection","neonatal clinical environments","YOLOv11","domain adaptation","face localization","non-contact monitoring","neonatal video analysis"],"falsifier":"Running the same fine-tuned model on a fresh collection of neonatal videos recorded under different hospitals, lighting, or equipment and finding that AP50 falls well below 0.96 would show the adaptation does not generalize.","tokens_in":2656,"feed_emoji":"👶","tokens_out":641,"duration_ms":20976,"temperature":0.7,"pith_summary":"The paper seeks to build a face detector that works reliably on newborns inside hospitals, where standard models lose accuracy because of changing light, background clutter, and medical equipment blocking the view. It starts with a one-stage detector trained on large public face datasets, then adapts the model using 228 videos from 113 infants recorded in real neonatal sessions. If the adaptation succeeds, the detector becomes accurate enough to support camera-based checks for pain, breathing pauses, and heart-rate signals without attaching sensors to the baby. Before adaptation the model already beats three general-purpose face detectors at 0.87 AP50; after adaptation the score rises to 0.96.","feed_headline":"Neonatal face detector reaches 0.96 AP50 after fine-tuning","feed_subtitle":"One-stage model trained on public faces then adapted to 113 infants overcomes clinical clutter and equipment obstructions.","key_machinery":"The one-stage YOLOv11m-based detector with clinical-domain adaptation on neonatal video data.","core_discovery":"A one-stage YOLOv11m model trained on combined public face datasets and then fine-tuned on a neonatal dataset of 228 videos from 114 sessions of 113 infants reaches an AP50 of 0.96 for localizing infant faces, outperforming three state-of-the-art general face detectors evaluated on the same clinical data.","pith_inferences":["Similar domain-adaptation steps could be tested on older children or other pediatric wards where face detection is also obstructed.","Public release of additional neonatal face datasets, with privacy protections, would allow direct comparison of future detectors.","The performance jump after adaptation suggests that small targeted clinical datasets can close large gaps left by adult-trained models."],"forward_implications":["Reliable face localization becomes available as the starting point for camera-based pain scoring and distress analysis in neonates.","Non-contact cardiorespiratory signal extraction and breathing-cessation alerts can operate without physical sensors attached to the infant.","The same adapted model can be evaluated against other general face detectors to quantify the gain from neonatal-specific fine-tuning."],"fun_headline_variants":["YOLOv11m reaches 0.96 AP50 for infant faces after fine-tuning","Infant detector hits 0.96 AP50 on neonatal clinical dataset","Domain-adapted YOLOv11m achieves 0.96 AP50 in NICU","Neonatal face detection at 0.96 AP50 with YOLOv11m adaptation","One-stage model scores 0.96 AP50 on 113 infant faces"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The collection of 228 videos from 113 infants is varied enough to represent real clinical conditions and to support domain adaptation without overfitting or hidden bias.","fun_headline_variants_meta":{"raw":{"variants":["YOLOv11m reaches 0.96 AP50 for infant faces after fine-tuning","Infant detector hits 0.96 AP50 on neonatal clinical dataset","Domain-adapted YOLOv11m achieves 0.96 AP50 in NICU","Neonatal face detection at 0.96 AP50 with YOLOv11m adaptation","One-stage model scores 0.96 AP50 on 113 infant faces"]},"model":"grok-4.3","cost_usd":0.0043,"raw_usage":{"total_tokens":2172,"prompt_tokens":689,"num_sources_used":0,"completion_tokens":108,"cost_in_usd_ticks":42999500,"prompt_tokens_details":{"text_tokens":689,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1375,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":689,"tokens_out":108,"duration_ms":10076,"temperature":1.0,"reasoning_tokens":1375,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T18:07:08.835071+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Running the same fine-tuned model on a fresh collection of neonatal videos recorded under different hospitals, lighting, or equipment and finding that AP50 falls well below 0.96 would show the adaptation does not generalize.","supporting_citations":[],"review_version":1}