{"id":"cb4bcf0e-3de7-4e2d-80f2-c8db7f64180f","arxiv_id":"2607.19796","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A caregiver-worn garment uses natural mouth-to-breast contact as a signal channel to estimate latch duration, infant heart rate, suck-swallow-breathe ratio, and milk intake during breastfeeding.","lead":"Mammal is a caregiver-worn tank top that measures a baby's heart rate, feeding rhythm, and milk intake during breastfeeding without attaching sensors to the baby. In a ten-family study it reported small errors on most metrics, but the milk-intake numbers rely on small fitted models and proxy ground truth.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Milk-intake estimate cannot be reproduced from the paper's own event counts and Eq. 8; predicted per-session intake is ~10× the reported ground truth, so the 15.76% MRE is unsupported as written.","rationale":"The reader identified the inter-body ECG contact dependency as the weakest assumption; that is a genuine and explicitly acknowledged deployment limitation (§9.2), but it does not contradict the reported numbers within the studied cohort. A more immediate correctness threat is the milk-intake model: as written, Eq. 8 combined with the paper's own event counts and durations would produce session-level intakes an order of magnitude above the observed test-weighing ground truth, making the 15.76% MRE impossible to reproduce. This is an internal inconsistency in the evaluation, not a disagreement with outside consensus. Since the paper is a feasibility demonstration and the other sensing axes may still be plausible, the appropriate disposition remains CONDITIONAL, with a specific condition that the intake model and constants be corrected, released, or independently validated.","tokens_in":34318,"tokens_out":11835,"duration_ms":123944,"concrete_test":"Recompute per-session predicted intake using Table 7's per-session suck counts, Table 5's latch durations, and Eq. 8's stated constants (ISI = 60·duration / count; m = (min(ISI,1300)-578)/776; intake = Σ m). Compare the predicted totals to the reported 61–125 g test-weight range and the 15.76% MRE. If the totals exceed the observed range by ~10×, then either obtain the exact preprocessing/units used (e.g., burst-level events, within-burst ISI only, or corrected constants) or rerun the evaluation with a reproducible model; if no correction reproduces the reported MRE, the intake claim should be removed or revalidated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing correctness issue is the milk-intake result in §8.4/§6.4. Table 7 reports 696–1,576 detected sucking events per session and Table 5 reports latch durations of 15.9–32.5 min. The mean inter-suck interval is therefore ~0.99–2.85 s (990–2,850 ms). Applying Eq. 8 with the stated constants (a=578, b=776, cap=1300 ms) gives m≈0.54–0.93 g per detected suck; summing over the detected suck counts yields predicted intakes of roughly 400–1,200 g per session. Observed test-weighing intakes are 61–125 g (§8.4). Thus the described pipeline would overestimate by ~10×, yet the paper reports a 15.76% MRE. Either the 'sucking events' counted for intake are not the same quantity as those in Table 7, the ISI is computed only within bursts rather than across successive events, or the constants/units in Eq. 8 are misreported. Without code/data, a reader cannot tell. This is not a minor calibration detail: milk intake is a headline metric. The age-calibrated model (§8.4) also refits those parameters on 4–5 participants and cannot validate the model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Mammal, a caregiver-worn nursing tank top that uses inter-body electrical and acoustic sensing to monitor breastfeeding. The system detects latch onset/duration, separates infant ECG from the mixed caregiver-infant signal, detects suck and swallow events from contact-microphone signals, and estimates infant heart rate, suck-swallow-breathe (SSB) ratio, and milk intake. In a study with 10 caregiver-infant dyads, the authors report a MAPE of 5.56% for latch duration, an MAE of 3.61 bpm for infant heart rate, an MAE of 0.12 for SSB ratio, and a mean relative error of 15.76% for milk intake, together with high comfort and wearability ratings. The central claim is that a caregiver-worn garment can passively derive these metrics without instrumenting the infant.","tokens_in":34702,"tokens_out":4982,"duration_ms":47343,"significance":"If the reported accuracy holds, Mammal would be a meaningful contribution to at-home breastfeeding monitoring, addressing a gap between clinical observation and burden-free sensing. The inter-body ECG and acoustic characterization experiments are thoughtful, and the per-session result tables and micro-benchmarks provide useful detail. However, the milk-intake result is internally inconsistent with the paper's own data and equations, and the age-calibrated model is fit on the evaluation cohort. These issues currently prevent verification of the headline claims, so the contribution cannot be assessed until they are resolved.","major_comments":[{"comment":"The reported 15.76% milk-intake error cannot be reproduced from the paper's own data. Table 7 lists 696–1,576 detected sucks per session and Table 5 lists latch durations of 15.9–32.5 min, implying mean inter-suck intervals of roughly 0.99–2.85 s. Applying Eq. (8) with a=578, b=776, cap=1300 ms gives m≈0.54–0.93 g per suck; summing over detected sucks yields 400–1,200 g per session, an order of magnitude above the 61–125 g test-weighing references in §8.4. The authors must clarify whether the 'sucking events' used for intake are the same quantity as in Table 7, how ISI is computed (across all successive sucks or within bursts), and whether the constants/units in Eq. (8) are correct. Reporting per-session estimated versus actual intake and releasing code/data would resolve this.","section":"§6.4, §8.4 (Eqs. 8–9; Tables 5 and 7)"},{"comment":"The age-calibrated intake model refits a, b, and cap on the same study cohort, split into N=4 (<6 months) and N=5 (≥6 months), and evaluates via leave-one-out. This is a post-hoc model search over the study data; the reported 4.76% and 2.91% MAPE values are therefore fitted results, not independent predictions of age dependence. Combined with the inconsistency in the preceding comment, these numbers cannot be interpreted as validation. Please present the fitted parameters, per-fold predictions, and either an independent test set or a clear statement that these are exploratory model-fitting results.","section":"§8.4 (age-calibrated intake model)"},{"comment":"The ECG channel depends on direct skin contact between the infant's bare hand or foot and the waist electrode. The paper acknowledges this in §9.2, but the evaluation does not report how often and for how long that contact occurred across the 10 sessions. The reported HR and SSB metrics are thus conditional on contact being maintained. Without quantifying contact availability and performance conditioned on contact versus no-contact periods, the 'passive monitoring' claim is stronger than the evidence supports. Please provide per-session contact duration and a breakdown of performance when contact is lost.","section":"§5.1, §6.1, §9.2"},{"comment":"Ground-truth swallowing events were annotated by a machine-learning model (YAMNet) applied to external audio, not by direct human observation, with only low-confidence events reviewed by a medical student. The reported swallowing F1=0.93 is therefore agreement with an algorithmic proxy rather than with human ground truth. The authors should validate the proxy against human annotation on at least a subset of data, or discuss how label noise affects the reported F1 and the downstream SSB ratio.","section":"§8.3"}],"minor_comments":[{"comment":"The heart-rate formula is written H = 60 / (1/10 * sum d_i). This is mathematically equivalent to 600/sum(d_i), but the notation is easy to misread; please add parentheses and define d_i explicitly as seconds per RR interval.","section":"Eq. (5)"},{"comment":"Table 6 reports SSB MAE as 'per session' in the header, while §8.4 states the MAE is per 10-second window. This is inconsistent; please align the units.","section":"Tables 6 and 8.4 text"},{"comment":"Duplicate word in the sentence 'the the theoretical caregiver ECG component' — please fix.","section":"§6.2"},{"comment":"The abstract says 'mean relative error of 15.76%' for milk intake, while §8.4 also uses 'mean relative error' but later introduces 'MAPE' for the age-calibrated model. Please use consistent terminology (e.g., MRE vs. MAPE) throughout.","section":"Abstract and §8.4"}],"recommendation":"major_revision","confidential_remarks":"The milk-intake inconsistency is the key blocker. If the authors can supply a corrected analysis or release code/data showing that the 15.76% MRE is reproducible from the described pipeline, the paper could become acceptable after a major revision. The rest of the sensing pipeline is promising and well presented. I recommend asking for a point-by-point response that includes the exact per-session intake computation and a clarification of the age-calibrated model's status."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clever, well-organized feasibility study for caregiver-worn breastfeeding monitoring using inter-body ECG and acoustics, but the milk-intake headline doesn't survive arithmetic against the paper's own numbers. That's a load-bearing issue.\n\nWhat's genuinely new: Mammal is the first to extend inter-body ECG to breastfeeding, adding an inter-body acoustic channel for suck/swallow and using mouth-to-breast coupling strength for latch detection. The garment design is thoughtful, and the two characterization experiments (adult dyads, silicone-breast benchtop) are reasonable first checks. The 10-dyad study is a real effort, with careful ablations on the ECG separation and event detection. The latch-duration, HR, and SSB results look plausible, and the authors are honestly flagging this as a feasibility demonstration.\n\nThe soft spot is milk intake. Using Table 7's suck counts and Table 5's session durations, the mean inter-suck interval is about 1–2 seconds. Plugging that into Eq. 8 (a=578, b=776, cap=1300 ms) gives roughly 0.5–0.9 g per suck, which sums to 400–1,200 g per session. The test-weigh ground truth is 61–125 g. So the 15.76% MRE cannot be reproduced from the described pipeline. The age-calibrated model in §8.4 is fit on the same cohort (N=4/5) with leave-one-out, which still doesn't validate the model class. Either the suck events used for intake are different from those in Table 7, the ISI is computed only within bursts, or the equation/units are misreported. Without code or data, a reader can't tell.\n\nOther limitations are real but acknowledged: the ECG channel needs bare skin contact with the waist electrode, and the ground truths are proxies (caregiver button presses, ankle PPG, external audio). None of these break the paper; they just limit the strength of the claims.\n\nWho's this for: people working on body-channel sensing, infant health monitoring, and computational garments. It deserves a serious referee, but the intake claim needs to be fixed or dropped, and the abstract toned down to match what's actually supported. If the intake pipeline is a reporting error, that's correctable in revision; if not, the paper should drop that metric and reframe as a feasibility study.","headline":"Mammal is a clever feasibility study with a strong core idea, but the milk-intake result as written is contradicted by its own equations and event counts; the intake claim needs correction or removal.","tokens_in":35194,"tokens_out":8825,"would_cite":true,"duration_ms":80704,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A caregiver-worn nursing garment can monitor an infant's latch, heart rate, feeding rhythm, and milk intake during breastfeeding without any sensors attached to the baby.","keywords":["breastfeeding monitoring","inter-body sensing","wearable ECG","computational garment","infant heart rate","suck-swallow-breathe ratio","milk intake estimation","latch detection"],"falsifier":"A controlled session in which the infant is fully clothed except for mouth contact, with no bare skin touching the waist electrode, should yield no separable infant ECG; if a usable ECG still appears, the electrical pathway model is wrong. Conversely, in a home-style feed with a swaddled infant, most sessions should fail—if they do not, the stated contact requirement is not actually limiting.","tokens_in":34215,"feed_emoji":"🍼","tokens_out":8061,"duration_ms":78411,"temperature":0.7,"pith_summary":"This paper tries to show that a caregiver-worn nursing garment can monitor the key quantities of a breastfeeding session—when the baby latched and for how long, the baby's heart rate during the feed, the suck–swallow–breathe ratio, and how much milk was transferred—without placing any sensor on the infant. The trick is that the baby's mouth on the breast creates a natural electrical and acoustic bridge: the infant's heart signal and feeding sounds travel through the contact and are picked up by electrodes and a contact microphone sewn into the caregiver's clothing. In a 10-dyad study the reported errors are 5.56% for latch duration, 3.61 bpm for heart rate, 0.12 for SSB ratio, and 15.76% for milk intake. If the method holds up, routine at-home breastfeeding monitoring could become a passive by-product of wearing a smart nursing top, opening a window on infant feeding competence and cardiovascular health. The main thing that has to be true for any of this to work is that the baby's bare skin stays in contact with a waist electrode while feeding.","feed_headline":"No baby sensors: nursing top tracks heart rate and milk intake","feed_subtitle":"Inter-body sensing reads the infant's ECG and suck-swallow sounds through the caregiver's own garment.","key_machinery":"Inter-body signal transmission across the mouth-to-breast contact is the load-bearing mechanism. The latch forms two coupled pathways: a low-impedance electrical pathway (oral moisture and soft tissue conduct the infant's ECG onto the caregiver's torso) and a mechanical/acoustic pathway (sucking and swallowing vibrations propagate through the nipple and breast tissue). The garment's textile electrodes on the waist and upper back capture the mixed ECG, and a contact microphone under the breast captures the acoustics. Latch is detected from the R-amplitude ratio between caregiver ECG and inter-body ECG (threshold 0.8); infant ECG is separated from the caregiver's by a wavelet-subband adaptive","core_discovery":"On the paper's own terms, Mammal establishes that the mouth-to-breast contact during breastfeeding is not just a feeding interface but a usable sensing channel. The oral seal has measurably lower electrical impedance than ordinary skin-to-skin contact (R-amplitude ratio 0.91±0.03 vs 0.64±0.07 in an adult surrogate experiment), so latch onset can be read from the coupling strength; infant ECG rides on the caregiver's body and can be separated from the caregiver's ECG; sucking and swallowing vibrations survive transmission to the breast and can be detected. Feeding these signal streams together yields the four target metrics at the reported accuracies in a 10-dyad study, and a preliminary NICU","pith_inferences":["A home deployment would likely expose much lower usable-data coverage than the lab study: because the ECG pathway requires the infant's bare skin on the waist electrode, feeds with swaddled or clothed infants would silently drop the heart-rate, breathing, and SSB outputs—the paper lists this as its key limitation, and quantifying the real-world dropout rate is the natural next experiment.","The milk-intake estimator inherits a 1982 relationship between inter-suck interval and milk transfer; the paper's own age-split results suggest the mapping shifts with infant age, so accuracy for newborns (under 2 months) is untested and may need a developmental recalibration rather than a single global curve.","If the inter-body channel is as reliable as reported, the same garment-based principle could extend to other contact-rich caregiving contexts—skin-to-skin holding, bottle feeding, or post-feed settling—where the electrode and acoustic placements would need to be re-derived for each contact site; Mammal does not claim these extensions.","A concrete improvement to test: replacing the exposed waist electrode with an actively shielded capacitive electrode could remove the bare-skin-contact requirement entirely (the paper names this as future work); verifying that in a swaddle condition would directly address the weakest assumption."],"forward_implications":["Latch onset and duration are recoverable from coupling strength alone: 102 latch onsets were detected with F1 0.93 and duration estimated within 5.56% MAPE, even in sessions with incidental skin-to-skin contact.","Infant heart rate during feeding is inferred from the separated inter-body ECG with 3.61 bpm MAE, and the paper reports that within-session HR and HRV trajectories track PPG-derived ground truth (r ≈ 0.95–0.96).","Suck and swallow events are detectable from the caregiver's body with F1 0.92–0.93, making the suck–swallow–breathe ratio measurable (0.12 MAE) as a marker of feeding coordination.","Milk intake is estimated from detected sucks and inter-suck intervals with 15.76% mean relative error, and an age-calibrated version of the model lowers the error to 2.91–4.76% in this small cohort.","A preliminary NICU clinician review suggests the resulting summaries could support monitoring during the NICU-to-home transition, a scenario the paper identifies as a likely first use."],"fun_headline_variants":["Smart nursing top reads baby's heart and milk intake","No infant sensors: this garment tracks nursing metrics","Mammal garment senses infant ECG and feeding sounds","Nursing top captures baby's ECG and milk intake","Wearable nursing monitor: latch, heart, suck-swallow"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"For the ECG channel to work, the infant's bare skin—typically a hand or foot—must rest against the garment's waist electrode while the mouth is on the breast; if the baby is swaddled, clothed, or positioned so no bare skin touches that electrode, the infant ECG, heart rate, breathing, and SSB estimates all disappear.","fun_headline_variants_meta":{"raw":{"variants":["Smart nursing top reads baby's heart and milk intake","No infant sensors: this garment tracks nursing metrics","Mammal garment senses infant ECG and feeding sounds","Nursing top captures baby's ECG and milk intake","Wearable nursing monitor: latch, heart, suck-swallow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1501,"prompt_tokens":745,"completion_tokens":756,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":677}},"tokens_in":489,"tokens_out":756,"duration_ms":9103,"temperature":1.0,"reasoning_tokens":677,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:42:00.264211+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled session in which the infant is fully clothed except for mouth contact, with no bare skin touching the waist electrode, should yield no separable infant ECG; if a usable ECG still appears, the electrical pathway model is wrong. Conversely, in a home-style feed with a swaddled infant, most sessions should fail—if they do not, the stated contact requirement is not actually limiting.","supporting_citations":[],"review_version":1}