{"id":"2737aa68-11b9-4d58-83eb-5e773f4de138","arxiv_id":"2504.20104","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"This preprint describes a longitudinal neonatal fingerprint collection of about 1,284 sessions from over 450 Brazilian infants, but does not release the data or report recognition experiments.","lead":"Researchers describe an ongoing collection of high-resolution fingerprint scans from more than 450 newborns in Brazil, with follow-up captures scheduled up to one year of age. The value for a general reader is that a shared longitudinal infant fingerprint dataset could improve biometric identification for vaccinations, nutrition tracking, and child protection.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4's session and video totals are internally inconsistent, so the database's core scale claim is not yet verifiable.","rationale":"The reader correctly notes the one-child visual inspection, the lack of quality metrics, and the dataset's unavailability. My stress-test found a more immediately load-bearing issue: the headline counts in Section 4 do not add up. Table 2's rows sum to 1,284, yet the stated 829 home visits plus 452 neonatal visits equal 1,281 (or the follow-up sum is 832, not 829). The 14-videos-per-session statement implies 17,976 videos, not 9,940. None of these discrepancies is fatal by itself—early sessions used still images, and rows may contain typographical errors—but they are exactly the kind of unverified numbers that a data paper must get right. The permanence claim in Section 4 is also not quantitatively established, but it is less central: the dataset's value does not require permanence, since the stated purpose is precisely to study growth and ageing. Therefore the most load-bearing concern is the internal incoherence of the dataset's reported statistics. A session-level manifest would resolve it. I recommend keeping the reader's conditional verdict: the paper can be accepted only if the authors release or supply the manifest (and ideally the underlying dataset or a public sample) and correct the inconsistent totals.","tokens_in":9183,"tokens_out":5038,"duration_ms":48908,"concrete_test":"Request a session-level manifest from the authors for all data through April 2025: one row per child-session with anonymized child ID, age milestone (0d/7d/14d/1m/2m/3m/6m/1y), capture modality (still/video), number of video files, and file sizes. Recompute: (i) total enrolled children ≥ 450; (ii) total sessions should match Table 2 sum 1,284; (iii) total non-birth sessions should equal 829 (or correct Table 2); (iv) total video files should equal the sum of per-session video counts, and total bytes should equal ~776.56 GB. If the manifest reconciles all values, the scale claim is verified; if not, the reported statistics must be corrected before the dataset description can be accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that a high-quality longitudinal neonatal fingerprint database exists with the reported scale. Section 4's own numbers do not cohere. Table 2 sums to 1,284 sessions, but the non-birth follow-up rows sum to 832, whereas the text reports 829 home-based re-collection sessions. More importantly, the text says 'Each biometric session includes 14 individual fingerprint videos per child' and reports 9,940 biometric video files; 1,284 sessions × 14 = 17,976 videos. Even restricting to the 829 home sessions gives 11,606, not 9,940. The early-protocol switch from still images to videos could explain a lower video count, but then the sentence claiming every session contains 14 videos is false, and no per-session breakdown is given. For a database descriptor, these are not cosmetic typos: the primary deliverable is a counted, curated collection. If the counts do not reconcile, the claim of 'one of the most comprehensive longitudinal neonatal biometric datasets to date' is currently unsupported. The one-child visual permanence observation is also weak evidence, but it is secondary to the unverified scale of the dataset itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes the design, protocol, and preliminary results of a longitudinal neonatal fingerprint collection effort in Brazil. The authors report enrollment of over 450 children, 1,284 biometric sessions across milestones from birth to 12 months, totaling 9,940 video files (roughly 776 GB), and claim this is one of the most comprehensive longitudinal neonatal biometric datasets. The stated scientific goal is to enable deep learning models that emulate growth-induced distortions of neonatal minutiae maps more faithfully than the scaling-factor models in the literature. Section 4 also presents a single-child visual comparison of fingerprints from birth to 6 months to support the permanence of ridge patterns under growth.","tokens_in":9370,"tokens_out":4650,"duration_ms":46897,"significance":"If the dataset is real and of the reported scale, it would fill a genuine gap: current public longitudinal neonatal fingerprint collections reach at most a few hundred subjects and typically only two sessions. The protocol's strengths include ethical approval through Plataforma Brasil, a multi-milestone follow-up schedule (7 days, 14 days, 1, 2, 3, 6, and 12 months), a 5000 ppi scanner, ten-finger capture plus repeated pinch-finger captures, and rich metadata such as gestational information, infant behavior, and environmental conditions. However, the paper currently provides no public access to the data, no objective quality metrics, and the reported counts do not reconcile internally. Because the paper is a database descriptor, these issues are central, not cosmetic. There is no mathematical derivation, fitted model, or machine-checked artifact; the credibility of the contribution rests entirely on the trustworthiness of self-reported counts and the reproducibility of the protocol, so the inconsistencies in the core numbers must be resolved.","major_comments":[{"comment":"The table's follow-up rows sum to 832 sessions (164 + 147 + 140 + 118 + 113 + 101 + 49), not the 829 home-based re-collection sessions stated in the text; the total of 1,284 sessions is consistent with 452 + 832, but only if the text's '829' is a typo. Additionally, the duplicated '1 month' row (140 and 118) makes the table ambiguous. Since the session total is the paper's primary quantitative claim, the authors should reconcile these numbers and explain whether the two rows correspond to distinct sub-cohorts or to an error.","section":"Section 4, Table 2"},{"comment":"The statement 'Each biometric session includes 14 individual fingerprint videos per child' is inconsistent with the reported total of 9,940 biometric video files: 1,284 sessions × 14 = 17,976, and even the 829 home sessions × 14 = 11,606. The protocol change from still images to video after the first few months can explain part of this discrepancy, but then the sentence is false for early sessions and no per-session or per-age breakdown of the video count is provided. The authors should report the actual number of videos per session and reconcile it with the stated total.","section":"Section 4"},{"comment":"The claim that 'the fingerprints do not change in their fundamental pattern, only in scale' is supported only by visual inspection of one child. Because the entire motivation of the database is to study non-linear growth distortions, and because Section 2 emphasizes the difficulty of obtaining sufficient minutiae quality in young children, the paper should include quantitative evidence—such as minutiae counts, NFIQ or other quality scores, or recognition rates across ages—for at least a subset of the cohort. Without such evidence, the 'high-quality' and 'permanence' claims are unsupported.","section":"Section 4, Figure 2"},{"comment":"The paper does not state how or when the database will be made available: there is no URL, repository identifier, or data-sharing plan. For a dataset-description paper, the absence of an access mechanism prevents readers from verifying any of the reported counts and limits the scientific reuse of the resource. Please add a data availability statement describing the intended distribution model (open, restricted, or by request) and the conditions under which the data would be shared.","section":"Sections 3–4"}],"minor_comments":[{"comment":"The caption states that 'each column represents one child's fingerprint at different ages' while the text says 'we present a series of images of one child'; this should be clarified to avoid ambiguity about whether the figure shows one child or multiple children.","section":"Section 4, Figure 2"},{"comment":"The columns 'People per ss' and 'Total images' mix values of different granularities across rows (e.g., '1', '308, 186, 123', '>= 45395'), and the footnotes define session, finger, and scanner abbreviations but not how per-session values are aggregated; please add a legend.","section":"Table 1"},{"comment":"Several references are incomplete or inconsistently formatted: [IN25] lacks a title, the citation '[J 19]' contains an extra space and appears in Table 1 as is, and the UNICEF/WHO reports ([UN23a], [UN23b], [WH23a], [WH23b]) would benefit from access dates.","section":"References"},{"comment":"The reference '[HGB19]' appears twice in the same sentence in the discussion of growth models; please remove the duplicate or indicate that two different works are intended.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a project progress report rather than a completed dataset paper, and the internal numeric inconsistencies in Section 4 are significant enough that the main claim cannot be verified as written. I found no evidence of deliberate misrepresentation, but the discrepancies must be addressed before the paper can be considered. I also note that the scanner vendor Infant.ID is both the supplier of the hardware and an explicitly acknowledged operational partner; the authors should make the nature of this relationship and any data-ownership or access agreements explicit, as this bears on how the dataset can be shared publicly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a data descriptor for a longitudinal neonatal fingerprint collection that, if it exists as described, fills a real gap in the literature. The protocol write-up is clear and honest about the shift from still images to videos, the related-work table is genuinely useful, and the collection size (450+ children, multiple sessions from birth to 12 months) goes beyond earlier longitudinal efforts like En21 or J19. I also appreciate that the authors describe ethical approval and metadata collection in some detail. That is the good part.\n\nThe soft spots are concentrated in Section 4, and they are not cosmetic. The text says 829 home re-collection sessions, but the follow-up rows in Table 2 sum to 832. The table also has two rows labelled \"1 month,\" which is presumably a typo for 2 months but should be fixed. More importantly, the paper states that each session includes 14 fingerprint videos, yet 1,284 sessions x 14 would be 17,976 videos, not the reported 9,940. Even restricting to the 829 home sessions gives 11,606. The early protocol switch to videos might explain the gap, but then the sentence claiming every session has 14 videos is false, and no per-session breakdown is given. For a database paper, these inconsistencies undermine the central claim that the repository is one of the most comprehensive to date.\n\nThe permanence observation is also thin: one child's prints shown visually across time is illustrative, not evidence. The deep-learning growth-emulation proposal is a hypothesis, not an evaluated result. There is also no data availability statement, no quality metrics, and no external audit, so the main artifact is not verifiable by readers. I want to stress that these are problems with the evidence as presented, not with the underlying collection effort, which could be valuable.\n\nCitation pattern looks fine. The authors engage the relevant prior work without inflating their contribution.\n\nWho this is for: people working on infant biometrics and fingerprint ageing will want to know this dataset exists and might use it once released. In its current form, I would not cite it as a source of data.\n\nRecommendation: send it to peer review rather than desk-reject, because the collection effort is substantial and the topic is underserved. But acceptance should be conditional on correcting the session/video arithmetic, fixing the table, and either releasing the data or providing per-session counts and quality statistics that let reviewers check the claimed scale.","headline":"A genuinely gap-filling longitudinal neonatal fingerprint collection, but the paper's core counts don't reconcile and no data is released, so treat it as a promising data descriptor in need of major revision.","tokens_in":9932,"tokens_out":2359,"would_cite":false,"duration_ms":25492,"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":"This paper reports the creation of a longitudinal neonatal fingerprint database in Brazil: more than 450 children, 1,284 collection sessions, and roughly 776 GB of high-resolution fingerprint video from birth to twelve months, intended to…","keywords":["neonatal biometrics","fingerprint ageing","longitudinal dataset","infant identification","minutiae map","deep learning growth emulation","fingerprint permanence","biometric database"],"falsifier":"Measure, on the collected sessions, the fraction of minutiae detected at birth that can be re-matched to the same infant's later captures with a standard matcher after only scale normalization; if the match rate is no better than chance or falls sharply by six months, the permanence assumption fails and growth-emulation models trained on this data cannot deliver the claimed benefit.","tokens_in":9007,"feed_emoji":"👶","tokens_out":8569,"duration_ms":74476,"temperature":0.7,"pith_summary":"The paper argues that the main barrier to better neonatal biometric identification is the absence of longitudinal fingerprint data, and it reports a project that is filling that gap: a growing collection of high-resolution (5,000 ppi) fingerprints from more than 450 newborns, captured repeatedly at eight developmental milestones from birth to twelve months. Between December 2023 and April 2025 the team ran 1,284 acquisition sessions and stored roughly 776 GB of fingerprint video. The authors' stated hypothesis is that a deep learning model fed the raw fingerprint, the ridge-and-valley segmentation, and the minutiae map can emulate growth-induced distortions more faithfully than the fixed or age-category scaling factors used in prior work. If the collection exists as described, it gives researchers the data needed to study fingerprint permanence in infancy and to train growth-emulation models.","feed_headline":"A 450-infant fingerprint database tracks growth from birth to year one","feed_subtitle":"Repeated high-resolution scans at eight ages give AI the data to model how infant fingerprints change.","key_machinery":"The load-bearing object is the longitudinal database itself: repeated, high-resolution captures of the same infants' ten fingers across developmental milestones, paired with metadata on age, gestational history, and collection conditions. The enabling premise is fingerprint permanence—the assumption that ridge and valley topology is stable during infancy and only scale changes—which is what makes the collected time series useful for training growth-emulation models. The protocol carries the argument: a 5,000 ppi infant-specific scanner, ten-finger capture plus pinch-finger video, and home visits at fixed milestones so that the same child is re-identified months later.","core_discovery":"The central claim is that a longitudinal neonatal fingerprint database of this scale can be produced in a hospital-to-home protocol: over 450 children enrolled at birth, with follow-up captures at roughly 7 days, 14 days, 1 month, 2 months, 3 months, 6 months, and 12 months, for a total of 1,284 sessions. The protocol calls for 14 fingerprint videos per infant per session—ten fingers plus repeated captures of the thumbs and index fingers—and the accumulated collection totals 9,940 biometric video files, about 776 GB. The paper's permanence hypothesis is illustrated by one child's prints from birth to six months: the overall ridge-and-valley pattern stays the same while the finger simply enlarges, so minutiae present at birth are still present later, spaced farther apart. On that basis the authors propose deep learning models that combine the raw image, segmentation, and minutiae to predict growth-related changes, moving beyond linear scaling.","pith_inferences":["The permanence claim currently rests on one child's visual example; the dataset itself could settle it quantitatively by measuring, across all 1,284 sessions, how many birth minutiae are re-detectable at later ages after scale normalization.","The paper does not describe a public release mechanism for the dataset; its practical value for the field depends on whether it is eventually shared or benchmarked publicly.","Because the captures use a 5,000 ppi infant-specific scanner, the findings may not transfer directly to lower-resolution or contactless devices used in field conditions; an extension would test the same permanence question on images downsampled to more common sensor resolutions.","If the permanence assumption fails for some infants, the same data could instead support template-updating strategies rather than static growth emulation."],"forward_implications":["If the database is real and complete as described, it becomes a benchmark for studying how infant fingerprints change over the first year of life.","It would allow deep learning models to be trained on raw images, ridge-and-valley structure, and minutiae jointly, and tested against scaling-factor baselines such as fixed and age-category scaling.","Reliable growth emulation would directly improve matching across time for vaccination tracking, nutrition program monitoring, and newborn identification in hospitals.","A successful permanence finding would support the design of national identity systems that enroll children at birth and re-identify them later without re-enrollment."],"supporting_citations":[{"why":"Establishes the feasibility of infant fingerprint capture and recognition and provides the scanner reference for ridge detail.","marker":"[En21]"},{"why":"Defines the fixed-scaling-factor baseline for emulating growth-induced minutiae distortion that this work aims to beat.","marker":"[Ja16b]"},{"why":"Shows age-category scaling on an infant and toddler fingerprint database and documents low-quality minutiae in young children.","marker":"[Ca17]"},{"why":"Proposes a two-factor fingerprint growth model that the paper identifies as not evaluated on younger individuals.","marker":"[HGB19]"},{"why":"Reports distinctive non-linear growth patterns in children's fingerprints that motivate abandoning simple linear scaling.","marker":"[Sc10]"},{"why":"Supplies the permanence and ageing-effect framework the paper uses to explain why fingerprints change with time.","marker":"[GHB18b]"},{"why":"Vendor reference for the 5,000 ppi scanner's ability to capture fine ridge detail on neonates.","marker":"[IN25]"}],"fun_headline_variants":["Brazil builds 450-baby fingerprint database to track infant growth","Neonatal fingerprint growth mapped across 450 infants in new database","450 newborns' fingerprints captured at 8 ages to train growth AI","Fingerprint database from 450 Brazilian infants tracks growth patterns","Infant print growth: 450 babies, 10k videos, 8 time points"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an infant's fingerprint ridge-and-valley pattern stays stable over the first months of life, changing mainly in size, so the same minutiae can be followed across sessions; the paper's support for this is visual inspection of one child, not quantitative comparison across the cohort.","fun_headline_variants_meta":{"raw":{"variants":["Brazil builds 450-baby fingerprint database to track infant growth","Neonatal fingerprint growth mapped across 450 infants in new database","450 newborns' fingerprints captured at 8 ages to train growth AI","Fingerprint database from 450 Brazilian infants tracks growth patterns","Infant print growth: 450 babies, 10k videos, 8 time points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1451,"prompt_tokens":974,"completion_tokens":477,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":383}},"tokens_in":590,"tokens_out":477,"duration_ms":5138,"temperature":1.0,"reasoning_tokens":383,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:58:38.574096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure, on the collected sessions, the fraction of minutiae detected at birth that can be re-matched to the same infant's later captures with a standard matcher after only scale normalization; if the match rate is no better than chance or falls sharply by six months, the permanence assumption fails and growth-emulation models trained on this data cannot deliver the claimed benefit.","supporting_citations":[],"review_version":1}