{"id":"421cb4a1-1988-4869-9c58-2684dce2d71a","arxiv_id":"2502.04365","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A thermal-video AI model detects time of birth from 3-second clips with 91.4% precision, 97.4% recall, and finds the birth in 24 of 25 test videos with median absolute error around 1.5 seconds.","lead":"This paper reports a machine learning system that watches thermal video of childbirth and automatically marks the moment of birth, so staff do not have to record it by hand. On a 25-birth test set the system found the birth in 24 cases, with typical error around one to two seconds versus human review.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth ToB labels are derived from the same thermal videos (Section 2), so the reported 1.5 s median error may measure model-human agreement on one modality rather than accuracy against an independent birth-time definition; this premise propagates to every Exp. 2 statistic.","rationale":"The paper's goal is automated ToB documentation with second-level precision. The strongest support is Table 2: with FIR and gamma=0.9, median absolute error is 1.5 s (the abstract says 1 s), found in 96% of videos. For this to substantiate the central claim, manual ToB annotations must be accurate to sub-second or at most one second. The paper explicitly says annotations were made by inspecting thermal videos because Liveborn app entries were delayed. This creates circularity: the model is trained and evaluated against human interpretation of the same sensor modality. A model can be highly consistent with a human annotator while being wrong against an external event definition, especially if both rely on the same ambiguous heat cues. The clip-level precision/recall figures depend on labels assigned from the same annotations, but they are not directly about ToB timing; the timing claim is most affected. The test set is small (25 manually selected videos), and threshold gamma=0.9 is selected from FPR analysis on this set, so results are optimistic. The concrete test of inter-annotator agreement directly measures label reliability; without it, the second-level error claim is not established. This does not undermine the contribution as a proof-of-concept, but it keeps the verdict conditional.","tokens_in":7555,"tokens_out":3336,"duration_ms":34366,"concrete_test":"Re-annotate the 25 test videos with two independent clinicians viewing the thermal videos, and also compare against any contemporaneous delivery-room log or RGB recording if available. Compute the inter-annotator median absolute difference; if it is around 1.5 s or larger, the model's reported median error is indistinguishable from label noise, so the second-level accuracy claim is not supported. If annotators agree to, say, under 0.5 s and an independent source confirms the annotations, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is second-level ToB accuracy. Section 2 states that manual registration in the Liveborn app was sometimes delayed, so manual annotation of the ToB with second precision was done by carefully inspecting the thermal videos. Thus the ground truth used in Eq. 6 is a human reading of the very same thermal feed the model sees. The clinical definition of ToB (head, torso, and nates fully outside the perineum) could in principle be scored from images, but if the human annotator is uncertain or biased by the same heat-signature cues the model uses, the 1.5 s median absolute error in Table 2 is an agreement measure, not a measure against an independent definition of birth time. The 25-video test set is also manually selected, and threshold gamma = 0.9 appears to be chosen after examining FPR on this set, which can inflate apparent performance. The label premise enters in Section 2 and propagates to clip classification labels and to every error statistic in Exp. 2.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spatiotemporal deep-learning system to detect the Time of Birth (ToB) from thermal videos in a privacy-preserving manner. Videos are normalized with a GMM-based method, fed as sliding-window clips to a 3D CNN (I3D, X3D, MoViNet) trained for binary ToB/no-birth classification, and the predicted scores are post-processed with an FIR filter and thresholded to estimate ToB. On a test set of 25 manually selected thermal birth videos, the MoViNet-A2 backbone achieves 91.4% precision and 97.4% recall at clip level, and the inference pipeline yields a median absolute error of 1.5 s (Q2=1.5 s) with ToB found in 96% of videos, while the abstract and conclusion claim a 1-second median absolute deviation. The authors compare against their earlier image-based method.","tokens_in":7795,"tokens_out":6356,"duration_ms":54222,"significance":"If the reported accuracy holds, this would be a valuable step toward automated, second-level ToB documentation during newborn resuscitation, with the added privacy advantage of thermal imaging. The paper includes a reasonable comparison of multiple backbones, a causal inference design with FIR filtering, and a clear statement of the clinical motivation. However, the strength of the quantitative claims is limited by the fact that the threshold and model are selected using the same test set, and the ground-truth labels are derived from the same thermal modality; the headline numbers should therefore be interpreted as model-human agreement under favorable selection, not as a validated clinical measurement. The work is nonetheless a useful case study and the pipeline is well described.","major_comments":[{"comment":"The abstract and conclusion state that the system estimates ToB with an 'absolute median deviation of 1 second' from manual annotations, but Table 2 reports Q2 = 1.5 seconds for the final video-based method with FIR filter, and the conclusion repeats the 1-second claim. This internal inconsistency directly misstates the central quantitative result and must be corrected, with the abstract and conclusion updated to match the actual Q2 value or with an explanation of why 1 second is reported.","section":"Abstract, Section 5 (Table 2), Section 6"},{"comment":"The confidence threshold gamma = 0.9 is chosen after inspecting the FPR curve (Figure 4a) computed on the same 25-video test set, and the same set is then used to report the ToB errors in Table 2 and Figure 4b. This constitutes selection on the test set and makes the reported Q2 and B.F. values optimistic. A separate validation set (or nested cross-validation) should be used for threshold selection, and if that is not possible, the paper should explicitly state the bias and provide a sensitivity analysis over thresholds on a held-out set.","section":"Section 4.2, Section 5 (Figure 4a, Table 2)"},{"comment":"The manual ToB annotations used as ground truth in Eq. (6) are obtained by carefully inspecting the same thermal videos (Section 2). Consequently, the reported median absolute error of 1.5 s (Table 2) measures agreement between the model and a human reading of the same thermal feed, not accuracy against an independent clinical definition of birth time (e.g., the head, torso, and nates fully outside the perineum). This premise propagates to every error statistic in Exp. 2 and to the clip-level labels. Please discuss this limitation explicitly, describe the annotation procedure in detail, and, if feasible, validate the annotations against an independent source (e.g., obstetric records, RGB video, or multiple annotators).","section":"Section 2, Section 4.2, Eq. (6)"},{"comment":"The model selection in Exp. 1 is performed on the test set: Table 1 compares backbones on the same 25-video test set, and the best-performing model (MoViNet-A2) is then used for the final evaluation. This double use of the test set for model selection and performance reporting compromises the statistical validity of the reported precision, recall, and MCC values. With 321 videos, a dedicated validation split (or repeated cross-validation) should be feasible; at minimum, the paper should acknowledge this limitation and provide estimates of the selection bias.","section":"Section 4, Table 1, Section 5"}],"minor_comments":[{"comment":"The sentence 'Weight decay of 0.97 is applied every 1k steps' appears to mix learning-rate scheduling with weight decay; please clarify whether this is a learning-rate decay factor or an L2 regularization weight decay schedule.","section":"Section 4.1.2"},{"comment":"The classification threshold used to compute precision/recall in Exp. 1 is not specified; please state whether a default 0.5 threshold or a tuned threshold was used.","section":"Section 4.1.2"},{"comment":"Table 2 reports quartiles and B.F. for the video-based method, but it is not explicitly stated that Q1/Q2/Q3/mean are computed over the 24 videos where ToB was found (excluding the missing video); please add a note to that effect.","section":"Section 5, Table 2"},{"comment":"The sentence 'It is important to note that the dataset used in this work is larger than the one used in the image-based method' is potentially confusing because Table 2 is said to use the same 25 test videos; clarify whether the image-based model was retrained on the enlarged dataset or evaluated as originally trained.","section":"Section 5"},{"comment":"The test set is described as 'manually selected,' and it is not stated whether the selection criteria (other than excluding twins and preserving position distribution) could introduce bias; please describe the selection process and its potential impact.","section":"Section 2"},{"comment":"The GMM normalization uses 'empirical constraints' for the temperature range of interest; please provide the specific constraint values so that the method is reproducible.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a compact case study with a small test set; the technical pipeline is sound but the evaluation protocol needs strengthening. The ground-truth issue and threshold selection are the main concerns; they are addressable with additional experiments or careful caveats, so I recommend major revision rather than rejection. The abstract/conclusion discrepancy should be fixed before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Genuinely new application: first spatiotemporal video-based ToB detector, and the improvement over the authors' own image-based method is credible. But the abstract's 1-second median claim doesn't match Table 2 (Q2=1.5 for the final method), and the evaluation protocol has two problems that inflate the accuracy numbers.\n\nThe good first. The clinical motivation is real—manually logged ToB is minute-precision and that matters for resuscitation research. Using thermal video for privacy is a sensible design choice, and they have a substantial dataset (321 births) with ethical approvals and a comparison of three backbones. GMM normalization and the FIR filter are straightforward but effective. The video-based method beating the image-based baseline on the same test set (Q2 1.5 s vs 3.2 s, mean 2.1 s vs 87.8 s) is a meaningful step forward, not just a toy result.\n\nNow the soft spots, in proportion. First, the internal inconsistency: the abstract says 1 second median absolute deviation; Table 2 reports 1.5 s for the method described in the text. That must be corrected. Second, the confidence threshold gamma=0.9 is selected after inspecting the FPR curve on the same 25-video test set (Fig 4a, Section 5). That is test-set tuning; the reported error statistics are partly fitted to the test set. A held-out threshold-selection protocol (or a proper validation split) is needed. Third, the ground-truth ToB labels come from manual annotation of the same thermal videos, as Section 2 states. So the error in Eq. 6 measures agreement between the model and a human reading the same thermal feed, not accuracy against an independent birth-time definition. The clinical definition of ToB is visually defined, so thermal video might be a reasonable ground truth, but it is not independent, and any bias in the human annotations (e.g., delayed reaction, miscounting) propagates directly into every number in Table 2. The manually selected test set (25 videos, keeping the maternal-position distribution) is a lesser concern but still worth noting.\n\nNone of these are load-bearing flaws in the sense that the qualitative conclusion—video-based thermal analysis can detect ToB with second-level precision—is likely wrong. But the numerical claims as stated are not fully supported. The fix is straightforward: correct the abstract, add a proper validation split for gamma, and either obtain an independent ground-truth annotation (e.g., from RGB video or a second annotator) or at least report inter-annotator agreement and discuss the limitation.\n\nWho's it for: researchers in clinical video analytics, neonatal care, privacy-preserving sensing. It deserves a serious referee—it's a real applied contribution with a clear use case and the issues are fixable. I'd accept it with major revision, not desk-reject.","headline":"A genuinely new privacy-preserving clinical video application, but the headline accuracy doesn't match its own table and the evaluation protocol needs tightening.","tokens_in":8314,"tokens_out":2946,"would_cite":false,"duration_ms":29036,"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":"Thermal video plus a 3D CNN estimates time of birth to within about 1.5 seconds.","keywords":["time of birth detection","thermal video analysis","privacy-preserving healthcare","newborn resuscitation","spatiotemporal video classification","MoViNet","Gaussian mixture model normalization","video-based clinical documentation"],"falsifier":"Have two or more clinicians independently re-annotate the 25 test videos (and ideally a fresh set of births) without seeing the model output, and measure inter-annotator agreement. If the spread between human annotators is comparable to or larger than the reported 1.5-second median error, the system's accuracy against a true time of birth is not established; if the model matches human consensus better than individual humans match each other, the claim is strengthened.","tokens_in":7377,"feed_emoji":"👶","tokens_out":6674,"duration_ms":59782,"temperature":0.7,"pith_summary":"This paper sets out to show that the time of birth can be detected automatically and with second-level precision from thermal video, without capturing identifiable images of mothers or healthcare staff. The proposed system classifies short sliding-window clips of ceiling-mounted thermal footage as birth or no-birth, smooths the per-second scores, and reads out the first moment the smoothed score crosses a confidence threshold. On a test set of 25 deliveries it detects the birth in 96% of videos, with a median absolute error of 1.5 seconds against manual annotation (the abstract states 1 second). This matters because time of birth is currently logged by hand, usually to the nearest minute, and the first 60 seconds after birth are the window in which resuscitation decisions are made.","feed_headline":"Thermal-video AI finds birth time in 96% of cases","feed_subtitle":"A 3D CNN plus smoothing filter estimates the moment of birth to about 1.5 seconds, without recording identifiable images.","key_machinery":"The load-bearing object is the sliding-window classifier over GMM-normalized thermal video. Each second, the model sees the previous 25 frames and outputs a birth probability; the sequence of probabilities is smoothed by a 3-tap rectangular FIR filter, a simple moving average over the last three prediction scores, and the estimated time of birth is the first timestamp at which the filtered score reaches a fixed threshold of 0.9. The GMM normalization fits three Gaussian components to each video's temperature distribution (background, clothing and bedding, and skin), rescaling intensities around the skin-temperature peak so that predictions depend on relative heat rather than absolute camera readings.","core_discovery":"The paper's central claim is that spatiotemporal information from thermal video is sufficient to pinpoint the moment of birth, and that a lightweight 3D CNN (MoViNet-A2) with a 3-sample smoothing filter and a threshold of 0.9 achieves clip-level precision of 91.4% and recall of 97.4%, estimating ToB in 24 of 25 test videos with a median absolute deviation of 1.5 seconds (abstract: 1 second). The method outperforms the authors' earlier image-based approach, whose median error was 3.2 seconds and mean error 87.8 seconds under the same threshold, by exploiting temporal birth dynamics instead of single frames. The system also demonstrates that relative temperature normalization via a Gaussian mixture model can make the network robust to camera autocalibration and room-temperature variation while preserving privacy.","pith_inferences":["If thermal-video timestamps are ever used in clinical documentation, the ground-truth question shifts: the current evaluation compares the model with a human reading of the same thermal feed, so the reported accuracy is agreement between two readers of one modality, not validation against an independent definition of birth time.","The non-causal whole-video GMM normalization means the reported numbers are for retrospective analysis; a causal, streaming version of the normalization would be needed before the system can issue real-time alerts during a delivery.","Because the test set is small (25 videos), excludes twins, and includes a known failure in the hands-and-knees position, the second-level precision claim should be re-measured on a larger, more diverse cohort before deployment decisions.","The same relative-temperature reasoning could transfer to other privacy-sensitive settings where timing matters, for example documenting when an intervention starts or when a procedure ends in emergency care."],"forward_implications":["Time-of-birth documentation can move from manual, minute-precision logging to automatic, second-precision timestamps generated from ceiling cameras.","Resuscitation debriefs and research timelines can be built from objective video-derived timestamps, including the interval from birth to first stimulation or ventilation.","The same privacy-preserving thermal pipeline may be reused for other clinical events that have a clear thermal or movement signature, such as cord clamping or first breath.","The threshold and filter settings give a tunable trade-off: raising the threshold suppresses false detections but increases the chance of missing a birth in positions where the newborn is hard to see.","A fallback image-based detector can cover the limited-visibility cases where the video-based model fails, as the paper explicitly proposes."],"supporting_citations":[{"why":"Provides the image-based ToB detector and GMM normalization baseline that the video system must beat.","marker":"[10]"},{"why":"Describes the consent-driven semiautomatic data collection that produced the 321 birth videos.","marker":"[11]"},{"why":"Supplies the Liveborn Observation App used to log manual ToB and save the 30-minute thermal videos.","marker":"[13]"},{"why":"Reports earlier thermal-video detection of time of birth and cord clamping that motivates the clinical application.","marker":"[14]"},{"why":"Documents thermal-sensor distortions such as autocalibration that justify the GMM normalization.","marker":"[15]"},{"why":"Supplies the I3D backbone compared in Table 1.","marker":"[16]"},{"why":"Supplies the X3D backbone compared in Table 1.","marker":"[17]"},{"why":"Supplies the MoViNet family from which the best-performing MoViNet-A2 is taken.","marker":"[18]"}],"fun_headline_variants":["Thermal-video AI times birth to within a second in 96% of cases","Privacy-safe AI uses thermal video to pinpoint the birth second","Thermal AI detects birth moment with 97% recall, no visible images","Heat-sensing AI records birth time to the second, preserving privacy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the manual ToB annotations, produced by carefully watching the same thermal videos, are accurate enough to serve as ground truth; if those human timestamps are biased or noisy, the reported precision measures agreement between the model and one human reading of the same camera feed rather than true birth time.","fun_headline_variants_meta":{"raw":{"variants":["Thermal-video AI times birth to within a second in 96% of cases","Privacy-safe AI uses thermal video to pinpoint the birth second","Thermal AI detects birth moment with 97% recall, no visible images","Heat-sensing AI records birth time to the second, preserving privacy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000932,"raw_usage":{"total_tokens":3961,"prompt_tokens":887,"completion_tokens":3074,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":3004}},"tokens_in":503,"tokens_out":3074,"duration_ms":20660,"temperature":1.0,"reasoning_tokens":3004,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:31:40.155011+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have two or more clinicians independently re-annotate the 25 test videos (and ideally a fresh set of births) without seeing the model output, and measure inter-annotator agreement. If the spread between human annotators is comparable to or larger than the reported 1.5-second median error, the system's accuracy against a true time of birth is not established; if the model matches human consensus better than individual humans match each other, the claim is strengthened.","supporting_citations":[{"cited_title":"Ad- vancing newborn care: Precise birth time detection using ai- driven thermal imaging with adaptive normalization,","cited_arxiv_id":null,"evidence_quote":"Provides the image-based ToB detector and GMM normalization baseline that the video system must beat."},{"cited_title":"Consent-driven, semi-automated data collec- tion during birth and newborn resuscitation: Insights from the NewbornTime study,","cited_arxiv_id":null,"evidence_quote":"Describes the consent-driven semiautomatic data collection that produced the 321 birth videos."},{"cited_title":"Digital health innovations, tools, and resources to support helping babies sur- vive programs,","cited_arxiv_id":null,"evidence_quote":"Supplies the Liveborn Observation App used to log manual ToB and save the 30-minute thermal videos."},{"cited_title":"Detection of time of birth and cord clamping using thermal video in the delivery room,","cited_arxiv_id":null,"evidence_quote":"Reports earlier thermal-video detection of time of birth and cord clamping that motivates the clinical application."},{"cited_title":"To- wards using thermal cameras in birth detection,","cited_arxiv_id":null,"evidence_quote":"Documents thermal-sensor distortions such as autocalibration that justify the GMM normalization."},{"cited_title":"Quo vadis, action recognition? a new model and the kinetics dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies the I3D backbone compared in Table 1."},{"cited_title":"X3d: Expanding architectures for efficient video recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the X3D backbone compared in Table 1."},{"cited_title":"Movinets: Mobile video networks for efficient video recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the MoViNet family from which the best-performing MoViNet-A2 is taken."}],"review_version":1}