Pith. sign in

REVIEW 3 major objections 7 minor 56 references

Camera detects baby breathing stops, hits 90% with sensors

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-07 22:14 UTC pith:EJYCTRFE

load-bearing objection Camera-only COBE detection in pre-term infants is feasible; hybrid improvement claim is weak due to label-input circularity and tiny test set the 3 major comments →

arxiv 2607.05230 v1 pith:EJYCTRFE submitted 2026-07-06 cs.LG

Video-based detection of cessation of breathing in pre-term infants using machine learning

Dineo Serame , Lionel Tarassenko , Mauricio Villarroel This is my paper
classification cs.LG
keywords apnoea of prematuritycessation of breathingvideo-based respiratory monitoringnon-contact monitoringmachine learningmultimodal data fusionneonatal intensive careimpedance pneumography
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper investigates whether a simple video camera can detect when a pre-term infant stops breathing (a cessation of breathing event, or COBE). The authors extracted two signals from video of infants in a neonatal intensive care unit: a frame-difference signal capturing overall torso motion, and a pixel-intensity signal (PPGirr) capturing localised abdominal breathing motion. A deep neural network trained on the abdominal-motion signal alone correctly classified COBE events with 76.9% balanced accuracy on held-out infants. When the video signal was fused with impedance pneumography (IP) — the standard chest-electrode method for measuring breathing — balanced accuracy rose to 90.6%, surpassing either modality used alone. The central claim is that video-derived respiratory motion contains clinically meaningful information that is complementary to conventional contact-based sensors, and that combining the two makes detection more robust.

Core claim

The key finding is that a camera watching a pre-term infant's abdomen extracts respiratory information that standard physiological sensors do not fully capture. The pixel-intensity signal from a small abdominal region of interest outperformed a global frame-difference signal, suggesting that localised breathing motion is more informative than general body movement. When this camera signal was combined with impedance pneumography, the hybrid model achieved 90.6% balanced accuracy — higher than video alone (76.9%) and higher than contact-only machine-learning models on the same data (88.7%). The improvement from fusion indicates that the two modalities carry partially independent information:

What carries the argument

Two camera-derived signals feed the models. The frame-difference (FD) signal measures mean pixel-intensity change between consecutive frames across the torso, capturing gross body motion. The PPGirr signal measures mean pixel intensity within a 75×75-pixel region over the lower abdomen, capturing breathing-induced surface motion. A ResNet processes these one-dimensional time series for classification. For hybrid models, a ConvNeXt branch processes physiological signals (IP waveform, ECG-derived respiration, PPG-derived respiratory envelope) and late-fusion concatenates features from both branches before a shared classifier outputs the probability of a COBE event for each 20-second window.

Load-bearing premise

The COBE labels that define ground truth were derived using the same impedance pneumography signal that is later fed to the hybrid models as an input feature. This means the model is partly learning to detect patterns in a signal that was used to create the labels it is judged against, which could inflate the apparent benefit of fusion.

What would settle it

If an independent reference standard (e.g., airflow or polysomnography) were used for labelling instead of IP-derived criteria, and the hybrid PPGirr+IP model no longer outperformed IP-only models, the claim of complementary video information would be weakened.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Non-contact video monitoring could serve as a backup or redundancy layer in NICUs where adhesive electrodes fail, are displaced, or cause skin damage to fragile pre-term infants.
  • The approach could be particularly valuable in low-resource settings where specialised monitoring equipment is scarce but camera-based systems are comparatively accessible.
  • If validated on larger multi-centre cohorts, camera-based detection could reduce false alarms by providing visual context that helps distinguish true breathing pauses from motion artefacts in contact sensors.
  • The finding that indiscriminate fusion of all signals (84%) underperformed selective PPGirr+IP fusion (90.6%) suggests that principled modality selection matters more than maximising data inputs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The ground-truth COBE labels were derived partly from the same impedance pneumography signal later used as a model input, creating a potential circularity: if IP artefacts influenced annotation, the hybrid model's advantage may partly reflect the model learning labelling patterns rather than genuinely complementary physiology. An IP-independent reference standard (e.g., airflow or polysomnography)
  • The use of BlazePose, trained on adults, for neonatal landmark detection is a stopgap; a pose-estimation model trained on infant anatomy could substantially improve ROI tracking robustness, especially during repositioning or partial occlusion.
  • The 23-infant sample and single-site design limit claims about generalisability across incubator types, lighting conditions, and ethnicities — factors known to affect camera-based vital-sign extraction.
  • If the video signal genuinely captures respiratory information independent of IP, then camera-only performance should correlate with the degree to which a COBE event produces visible abdominal motion versus a purely central pause with no respiratory effort. The paper does not analyse this distinction, which would clarify what the camera is actually detecting.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. This manuscript presents a machine learning framework for detecting cessation of breathing (COBE) in pre-term infants using video camera-derived signals, both alone and in combination with physiological signals. Camera-derived respiratory motion signals (frame-difference and pixel-intensity PPGirr) are extracted from dynamically tracked torso/abdominal regions and fed to adapted 1D ResNet architectures. Hybrid models fuse camera signals with impedance pneumography (IP), ECG-derived respiration, and PPG envelopes via late fusion. The study uses data from 30 pre-term infants (23 retained after quality assessment), with infant-level train/test splitting (19 train / 4 test) and five-fold cross-validation. Camera-only models achieve 76.9% balanced accuracy on the test set; the best hybrid model (PPGirr + IP) achieves 90.6%, compared to 88.7% for contact-only ML models on the same dataset (ref 58).

Significance. The study addresses a clinically important problem—non-contact respiratory monitoring in the NICU—and the camera-only result (76.9% balanced accuracy) is a meaningful feasibility demonstration. The infant-level data splitting and five-fold cross-validation are methodologically appropriate. The systematic evaluation of multiple signal combinations (Tables 3, 5, S1, S2) provides useful empirical evidence on which modalities contribute discriminative information. The finding that indiscriminate fusion of all signals degrades performance (84.0% vs 90.6%) is a practically useful observation. However, the central claim that video provides complementary information beyond standard physiological signals rests on a comparison that is not statistically tested, and the label-input relationship between IP and the COBE annotations warrants explicit discussion.

major comments (3)
  1. §2.3 and §2.9: Label-input circularity. COBE labels are defined using RRip, which is computed directly from the IP waveform (§2.2–2.3), and the filtered IP waveform is used as a model input in the hybrid models (§2.9). The paper does not discuss this shared signal source. While the labels involve multi-signal human annotation (not a pure algorithmic function of IP), the IP/RRip signal is a primary input to the annotation process, and its direct use as a model feature creates a potential information path from label to input. The paper should explicitly acknowledge this limitation and discuss its implications for interpreting the hybrid model's performance. This is load-bearing because the complementary-information claim depends on the hybrid model's performance being attributable to video rather than to the IP branch exploiting label-correlated patterns.
  2. §3, Table 5 and Discussion: The claim that video provides complementary information rests on the hybrid model (PPGirr + IP, 90.6%) outperforming contact-only ML (88.7%, ref 58) by approximately 2 percentage points. With only 4 test infants (116 positive windows, 728 total windows), no confidence intervals or significance tests are reported for any result. A 2-point difference on this test set is well within plausible sampling variability. The paper should either provide bootstrap confidence intervals or a permutation test for the difference, or substantially temper the complementary-information claim. As stated, the abstract's assertion that the hybrid result indicates video provides 'respiratory information beyond standard physiological signals' is not supported by the evidence presented.
  3. §2.6.1, Table 2: The test set comprises only 4 infants (116 positive, 612 negative windows). The paper does not report how these 4 infants were selected or whether the test-set composition is representative. Given that the central quantitative claims (76.9% and 90.6%) are derived solely from this small test set, the absence of any uncertainty quantification makes it difficult to assess whether the reported numbers are stable. At minimum, the paper should report per-infant performance or bootstrap-based confidence intervals for the primary metrics.
minor comments (7)
  1. §2.4.1, Eq. (3): The EMA smoothing coefficient thresholds (2.55, 7 pixels) and corresponding α values (0.2, 0.6, 1.0) are described as 'selected empirically based on visual inspection.' A brief note on sensitivity to these choices would strengthen the analysis.
  2. §2.5.2: 29% of COBE segments (100/346) and 27% of normal segments (165/608) were excluded due to occlusion or poor video quality. This exclusion rate is substantial and could introduce selection bias if excluded segments differ systematically from retained ones. A brief discussion of this would be helpful.
  3. Table 5: The Cohen's κ value for PPGirr + IP is reported as 0.7 in Table S2 but 0.68 in Table 5. Please reconcile.
  4. §2.3: The inter-rater reliability for non-desaturation events (κ=0.57) is only moderate. Given that 620 non-desaturation segments were randomly sampled for annotation, the paper could comment on how label noise in this subset might affect model training.
  5. Abstract: 'outperforming either modality alone' — the paper does not report a standalone IP-only model trained with the same architecture and protocol. The 88.7% comparison comes from a different model (ref 58). The abstract should clarify that the comparison is to a previously published contact-only model, not to an IP-only branch within the current framework.
  6. §4, Discussion: The paper states that combining PPGirr with PPG resulted in 'only a marginal increase' over PPGirr alone (77.1% vs 76.9%), but this is a 0.2-point difference that is better described as 'no improvement.'
  7. References: Ref 58 (Serame et al., medRxiv 2026) is cited as 'Paper in preparation/archive.' If this is the companion contact-only study, its availability and peer-review status should be clarified, as the central comparison depends on it.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The three major comments raise related and valid concerns about (1) potential label-input circularity arising from the shared use of impedance pneumography (IP) in both COBE annotation and hybrid model inputs, (2) the absence of statistical testing or confidence intervals for the central complementary-information claim, and (3) the small test set (4 infants) and lack of uncertainty quantification. We agree with all three points and will revise the manuscript accordingly. Specifically, we will: (a) add an explicit discussion of the IP label-input relationship and its implications for interpreting hybrid model performance; (b) report bootstrap confidence intervals and a permutation test for the difference between the hybrid and contact-only models; and (c) report per-infant test-set performance and describe test-set selection criteria. We will also temper the complementary-information claim in the abstract and discussion to reflect the uncertainty in the evidence.

read point-by-point responses

  1. Referee: §2.3 and §2.9: Label-input circularity. COBE labels are defined using RRip, which is computed directly from the IP waveform (§2.2–2.3), and the filtered IP waveform is used as a model input in the hybrid models (§2.9). The paper does not discuss this shared signal source.

    Authors: The referee is correct that this shared signal source is not discussed in the manuscript and that it warrants explicit acknowledgement. We will add a paragraph to the Discussion (and a note in §2.9) addressing this issue directly. To clarify the nature of the concern: COBE labels were not derived algorithmically from RRip alone. The annotation process (§2.3, Appendix B) involved three independent annotators who reviewed RRip, SpO2, ECG, PPG, and the raw IP waveform within 5-minute contextual windows. A COBE event required either (i) RRip < 20 breaths/min for at least 20 seconds, or (ii) a shorter pause with concurrent bradycardia or desaturation. The final label reflected a consensus human decision incorporating multiple signals, not a thresholded function of IP. Nevertheless, we agree that RRip—derived from IP—is a primary input to the annotation process, and the filtered IP waveform is simultaneously used as a model feature in the hybrid models. This creates a potential information path from label to input that cannot be fully excluded. We will state this limitation explicitly and discuss its implications for interpreting the hybrid model's performance. Specifically, we will note that the hybrid model's performance cannot be unambiguously attributed to video providing complementary information, because the IP branch may exploit label-correlated patterns in the IP waveform. We will also note that the camera-only models (76.9%) do not use IP as input and are therefore not affected by this concern, so the feasibility claim for non-contact COBE detection remains valid. The complementary-information claim, however, will be appropriately qualified. revision: yes

  2. Referee: §3, Table 5 and Discussion: The claim that video provides complementary information rests on the hybrid model (PPGirr + IP, 90.6%) outperforming contact-only ML (88.7%, ref 58) by approximately 2 percentage points. With only 4 test infants (116 positive windows, 728 total windows), no confidence intervals or significance tests are reported for any result.

    Authors: The referee is correct. A 2-percentage-point difference on a test set of 4 infants (728 windows) is well within plausible sampling variability, and the absence of confidence intervals or significance tests is a genuine gap. We will address this in two ways. First, we will compute bootstrap confidence intervals (using infant-level resampling with 10,000 iterations) for the balanced accuracy of the camera-only model, the best hybrid model (PPGirr + IP), and the contact-only ML model from ref 58, evaluated on the same test set. Second, we will perform a permutation test for the difference in balanced accuracy between the hybrid and contact-only models. We will report these results in a revised Table 5 (or an accompanying table) and in the Discussion. If the confidence intervals are wide or the permutation test does not yield a statistically significant difference—as is plausible given the small test set—we will explicitly state that the complementary-information claim is not statistically supported and revise the abstract and discussion accordingly. We agree that the current abstract's assertion ('indicating that video provides respiratory information beyond standard physiological signals') overstates the evidence and will be softened to reflect the uncertainty. revision: yes

  3. Referee: §2.6.1, Table 2: The test set comprises only 4 infants (116 positive, 612 negative windows). The paper does not report how these 4 infants were selected or whether the test-set composition is representative. Given that the central quantitative claims (76.9% and 90.6%) are derived solely from this small test set, the absence of any uncertainty quantification makes it difficult to assess whether the reported numbers are stable.

    Authors: The referee raises a valid concern. We will address it by (1) describing the test-set selection criteria in §2.6.1, and (2) reporting per-infant performance breakdowns. Regarding selection: the 23 infants were partitioned into 19 training and 4 test infants to maintain demographic balance with respect to gestational age, sex, and the distribution of COBE and non-COBE windows, as stated in §2.6.1. We will make the specific demographic and clinical characteristics of the 4 test infants explicit (gestational ages, sex, number of COBE and normal segments per infant) so that readers can assess representativeness. Regarding per-infant performance: we will add a table reporting balanced accuracy, TPR, and FPR for each of the 4 test infants individually, for both the camera-only (PPGirr) and best hybrid (PPGirr + IP) models. This will allow readers to assess whether the aggregate metrics are driven by one or two infants or are reasonably consistent across the test set. We acknowledge that 4 test infants is a limitation inherent to the dataset size and will state this explicitly. We will also note that the five-fold cross-validation results (Tables 3, S1) provide additional evidence of stability across the training folds, though they do not substitute for a larger independent test set. revision: yes

Circularity Check

1 steps flagged

Label-input signal overlap between COBE annotation (RRip from IP) and hybrid model input (filtered IP waveform) creates partial methodological circularity, but labels involve expert multi-signal review and are not mechanically reducible to the IP input.

specific steps
  1. fitted input called prediction [Sections 2.2–2.3 (label definition) and Section 2.9 (hybrid model input)]
    "Section 2.2: 'respiratory rate was recomputed directly from the raw IP waveform (RRip).' Section 2.3: 'A COBE event was defined as either (i) a respiratory pause (RRip < 20 breaths/min) lasting at least 20 seconds, or (ii) a shorter respiratory pause of at least 10 seconds accompanied by bradycardia (HR < 100 beats/min) or oxygen desaturation (SpO2 < 80% lasting at least 10 seconds).' Section 2.9: 'The filtered IP waveform was used directly as an input respiratory signal.'"

    The COBE ground-truth labels are partly defined by RRip, which is derived from the IP waveform (Section 2.2). The hybrid model then receives the filtered IP waveform as a direct input feature (Section 2.9). This creates a structural overlap: the model has access to the same signal from which a key component of the label was derived. A model receiving the raw IP waveform can learn patterns corresponding to low RRip (the primary label criterion) without genuinely extracting complementary information. However, this is not a perfect by-construction circularity: (1) labels involve expert multi-signal review by three annotators using RRip, SpO2, ECG, PPG, and IP waveforms with inter-rater agreement requirements (κ=0.80 and 0.57); (2) the label definition includes disjunctive criteria (bradycard+

full rationale

The paper's central feasibility claim — that camera-only models can detect COBE (76.9% balanced accuracy) — is unaffected by circularity concerns, as camera-only models use no physiological inputs. The hybrid model's 90.6% accuracy does involve a label-input overlap: RRip (from IP) defines COBE labels, and the filtered IP waveform is a model input. This partially inflates the IP-containing branch's performance, but the reduction is not exact because labels require expert annotation across multiple signals and are not mechanically thresholded from the IP input alone. The self-citations (refs 31, 45, 58 — overlapping authors) provide methodology descriptions and a baseline comparison, not load-bearing theoretical premises that would force the conclusion. The ~2-point hybrid improvement over contact-only ML (88.7%, ref 58) is small and untested for significance, but this is a statistical power concern, not circularity by construction. Score 3 reflects the partial label-input overlap without rising to a forced result.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities, particles, or forces. The free parameters are standard machine learning and signal processing hyperparameters. The axioms are domain assumptions about the clinical validity of IP-based labelling and pose estimation on neonates, which are the primary epistemic risks.

free parameters (5)
  • EMA smoothing coefficients (αt) = 0.2, 0.6, 1.0
    Selected empirically based on visual inspection of ROI stability (Section 2.4.1, Eq. 3). Displacement thresholds (2.55, 7 pixels) also empirically chosen.
  • RR ROI size = 75×75 pixels
    Selected based on previous work (Section 2.4.2); not optimized on current data but is a design choice.
  • ResNet channel configuration = [32,32,64,64]
    Reduced from standard [64,128,256,512] following Carter et al.; chosen for 1D signal complexity.
  • Learning rate range = 9×10⁻⁵ to 1×10⁻⁴
    Range specified but exact final value per model not reported (Section 2.8.2).
  • Class weights for cross-entropy loss = Not specified
    Higher weights assigned to minority COBE class (Section 2.10) but exact values not reported.
axioms (4)
  • domain assumption BlazePose landmarks can be reliably detected on neonatal bodies despite the model being trained on adults.
    Section 2.4 states BlazePose is 'primarily trained on adult human data' but was applied to neonatal recordings for ROI localisation. No quantitative validation of landmark accuracy on neonates is provided.
  • domain assumption IP-derived respiratory rate (RRip) is a reliable reference for COBE labelling.
    Section 2.2 argues that monitor-provided RR is unreliable and recomputes RRip from raw IP. However, IP itself is susceptible to motion artefacts and cardiac oscillations (acknowledged in Section 1), and no independent gold standard (e.g., polysomnography) is used to validate RRip.
  • domain assumption Three-annotator majority agreement constitutes valid ground truth for COBE events.
    Section 2.3 uses three annotators (one clinician, two engineers) with Fleiss' κ=0.80 for desaturation events and κ=0.57 for non-desaturation events. The moderate agreement for non-desaturation events indicates label noise.
  • domain assumption 20-second analysis windows with 10-second overlap preserve clinically relevant COBE information.
    Section 2.6 states this window length was chosen to contain the shortest COBE episodes (10 seconds), but no sensitivity analysis of window size is provided.

pith-pipeline@v1.1.0-glm · 20718 in / 2956 out tokens · 722546 ms · 2026-07-07T22:14:44.632917+00:00 · methodology

0 comments
read the original abstract

Pre-term infants are susceptible to potentially harmful apnoea-related cessations of breathing due to immature respiratory control. However, reliable respiratory monitoring in the neonatal intensive care unit (NICU) remains challenging because motion artefacts, sensor displacement, and skin fragility can compromise contact-based measurements. Non-contact video monitoring offers a complementary approach that does not depend on adhesive sensors while providing additional respiratory information. We investigated whether camera-based signals can detect apnoea-related cessation of breathing (COBE) and provide complementary information to routinely acquired physiological signals. Using video and clinical recordings from 30 pre-term infants, respiratory motion was extracted from dynamically tracked torso regions to generate camera-derived time-series signals. Camera-only models were trained using residual network (ResNet) architectures, while hybrid models combined video-derived signals with impedance pneumography (IP), ECG-derived respiration (EDR), and the PPG-derived respiratory envelope. Camera-only models achieved a balanced accuracy of 76.9%, demonstrating the feasibility of non-contact COBE detection. Combining video-derived features with IP improved balanced accuracy to 90.6%, outperforming either modality alone and indicating that video provides respiratory information beyond standard physiological signals. These findings show that video-derived signals contain clinically relevant respiratory features and enhance COBE detection when combined with conventional physiological signals. This supports non-contact video as a complementary modality for automated COBE detection and highlights its potential to improve the robustness of neonatal respiratory monitoring.

Figures

Figures reproduced from arXiv: 2607.05230 by Dineo Serame, Lionel Tarassenko, Mauricio Villarroel.

Figure 1
Figure 1. Figure 1: Components of the data acquisition setup used for synchronised video and physiological monitoring in the NICU. a) Video camera positioned through a modified opening in the incubator canopy; b) Representative video frame acquired during monitoring; c) Representative Philips patient monitor with SpO2 measurement modules. 4/29 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: RR estimation using the breath counting method on a 5-minute segment of the IP signal containing a COBE episode (approximately 1.6–2.7 mins): a) Filtered IP waveform; b) comparison of the recomputed respiratory rate (RRip) and the monitor-provided respiratory rate (RRphilips) during a period containing COBE. RRphilips exhibits reduced short-term variability, consistent with temporal averaging. 5/29 [PITH_… view at source ↗
Figure 3
Figure 3. Figure 3: Dynamic ROI localisation using BlazePose landmarks. (a) Left and right shoulder and hip landmarks (blue circular markers) connected to form a torso quadrilateral used to estimate torso orientation; (b) The resulting torso ROI (blue). A 75×75 pixel respiratory ROI (red) was positioned over the lower abdominal region to capture respiratory-related motion. 8/29 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Camera-derived respiratory signals extracted from a neonatal video recording, showing an example segment containing periods of normal breathing and a COBE episode. a) Overlapping 20-second analysis win￾dows assigned binary class labels of COBE (red) or non-COBE (green). Black asterisks indicate the centres of consecutively labelled windows (10-s step size), while the coloured horizontal lines show the temp… view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the machine learning approaches evaluated for COBE detection. Camera-only models were trained using FD and PPGirr signals. Multimodal camera models combined FD and PPGirr, while hybrid models integrated camera-derived and physiological signals. ResNet and ConvNeXt architectures were evaluated for binary classification of COBE and non-COBE. All models were trained and evaluated using the same da… view at source ↗
Figure 6
Figure 6. Figure 6: Adapted one-dimensional ResNet architecture for camera-derived COBE detection. The network processes frame-difference (FD) and pixel-intensity (PPGirr) signals to produce a binary classification output. 2.8.2 Network training Three standard ResNet variants (ResNet-18, ResNet-34, and ResNet-50) were evaluated to assess the influence of network depth on camera-based COBE detection performance. Models were tr… view at source ↗
Figure 7
Figure 7. Figure 7: Late-fusion ResNet architecture for multimodal camera-based COBE detection. Frame-difference (FD) and PPGirr signals were processed independently using separate ResNet branches identical to the single-input architecture shown in figure 6. The resulting feature vectors were concatenated and passed to a shared multi-layer perceptron (MLP) classifier to predict COBE and non-COBE events. 13/29 [PITH_FULL_IMAG… view at source ↗
Figure 8
Figure 8. Figure 8: Hybrid multimodal architecture for COBE detection. Camera-derived features extracted using ResNet were fused with physiological features extracted using ConvNeXt. The resulting feature vectors were concatenated and processed by a fusion head before final classification. 14/29 [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Review workflow used for reference labelling of COBE events. Annotators examined synchronised respiratory rate, oxygen saturation, and waveform signals within extended contextual windows to determine whether a true cessation of breathing was present. C Supplementary methods: Camera-based respiratory rate estimation To support respiratory rate estimation during both normal breathing and COBE episodes, a mot… view at source ↗
Figure 10
Figure 10. Figure 10: Examples of respiratory signals extracted from synchronised physiological recordings and used for machine learning. a) IP waveform b) ECG-derived respiration (EDR), and c) the PPG-derived respiratory envelope. E COBE detection using a combination of camera and physiological signals Tables S1 and S2 show the performances of hybrid models under different input combinations. Each hybrid model combines camera… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages · 1 internal anchor

  1. [1]

    Licence: CC BY-NC-SA 3.0 IGO

    World Health Organization (WHO).Born too soon: decade of action on preterm birth(Geneva: World Health Organization, 2023). Licence: CC BY-NC-SA 3.0 IGO

    work page 2023

  2. [2]

    Koteswari, P.et al.Preterm birth: causes and complications observed in tertiary care hospitals.Cell. Mol. Biomed. Reports2, 202–2012 (2022). 3.Platt, M. Outcomes in preterm infants.Public health128, 399–403 (2014)

    work page 2012

  3. [3]

    Thompson, L., Werthammer, J. W. & Gozal, D. Apnea of prematurity and oxidative stress: Potential implications. Antioxidants13, 1304 (2024)

    work page 2024

  4. [4]

    Res.1–8 (2025)

    O’Shea, M.et al.Caffeine and preterm infants: multiorgan effects and therapeutic creep: scope to optimise dose and timing.Pediatr. Res.1–8 (2025). 6.Mohr, M. A.et al.Very long apnea events in preterm infants.J. Appl. Physiol.118, 558–568 (2015)

    work page 2025

  5. [5]

    V ., Fort, P., Miller, T

    Pergolizzi Jr, J. V ., Fort, P., Miller, T. L., LeQuang, J. A. & Raffa, R. B. The epidemiology of apnoea of prematurity.J. clinical pharmacy therapeutics47, 685–693 (2022)

    work page 2022

  6. [6]

    Sleep2, 1256078 (2023)

    Zou, D.et al.A new approach to streamline obstructive sleep apnea therapy access using peripheral arterial tone-based home sleep test devices.Front. Sleep2, 1256078 (2023)

    work page 2023

  7. [7]

    & Ogunsola, F

    Nwaneri, S., Ezenwa, B., Osuntoki, A., Ezeaka, V . & Ogunsola, F. A review of infant apnea monitor design.J. Clin. Sci.21, 93–98 (2024)

    work page 2024

  8. [8]

    A.et al.Medical devices for pediatric apnea monitoring and therapy: past and new trends.IEEE reviews biomedical engineering10, 199–212 (2017)

    Pullano, S. A.et al.Medical devices for pediatric apnea monitoring and therapy: past and new trends.IEEE reviews biomedical engineering10, 199–212 (2017)

    work page 2017

  9. [9]

    & Isaiah, A

    Bertoni, D. & Isaiah, A. Towards patient-centered diagnosis of pediatric obstructive sleep apnea—a review of biomedical engineering strategies.Expert. review medical devices16, 617–629 (2019)

    work page 2019

  10. [10]

    K., Miaskowski, C., Hu, X., Rodway, G

    Bawua, L. K., Miaskowski, C., Hu, X., Rodway, G. W. & Pelter, M. M. A review of the literature on the accuracy, strengths, and limitations of visual, thoracic impedance, and electrocardiographic methods used to measure respiratory rate in hospitalized patients.Annals Noninvasive Electrocardiol.26, e12885 (2021). 13.Massaroni, C.et al.Contact-based methods...

    work page 2021

  11. [11]

    Poets, C. F. Pulse oximetry vs. transcutaneous monitoring in neonates: practical aspects.www. bloodgas. com, Neonatol. Copenhagen: Radiom. Med. A/S(2003). 25/29

    work page 2003

  12. [12]

    Ostojic, D.et al.Reducing false alarm rates in neonatal intensive care: a new machine learning approach.Oxyg. Transp. to Tissue XLI285–290 (2020)

    work page 2020

  13. [13]

    Baharestani, M. M. An overview of neonatal and pediatric wound care knowledge and considerations.Os- tomy/wound management53, 34–6 (2007)

    work page 2007

  14. [14]

    Sharma, A.Chronological evaluation of functional changes in neonatal skin: A temporal evaluation of skin maturation. Ph.D. thesis, University of Southampton (2025)

    work page 2025

  15. [15]

    & Boylan, G

    Lloyd, R., Goulding, R., Filan, P. & Boylan, G. Overcoming the practical challenges of electroencephalography for very preterm infants in the neonatal intensive care unit.Acta paediatrica104, 152–157 (2015)

    work page 2015

  16. [16]

    & Lasenby, J

    Bonner, O., Beardsall, K., Crilly, N. & Lasenby, J. There were more wires than him: the potential for wireless patient monitoring in neonatal intensive care.BMJ innovations3(2017)

    work page 2017

  17. [17]

    G.et al.Inpatient care of small and sick newborns: a multi-country analysis of health system bottlenecks and potential solutions.BMC pregnancy childbirth15, S7 (2015)

    Moxon, S. G.et al.Inpatient care of small and sick newborns: a multi-country analysis of health system bottlenecks and potential solutions.BMC pregnancy childbirth15, S7 (2015)

    work page 2015

  18. [18]

    R., Julius, Z., Kosia, E

    Kamala, S. R., Julius, Z., Kosia, E. M. & Manzi, F. Availability and functionality of neonatal care units in healthcare facilities in mtwara region, tanzania: The quest for quality of in-patient care for small and sick newborns.Plos one17, e0269151 (2022)

    work page 2022

  19. [19]

    In2014 IEEE Workshop on Biometric Measurements and Systems for Security and Medical Applications (BIOMS) Proceedings, 67–73 (IEEE, 2014)

    Cattani, L.et al.A wire-free, non-invasive, low-cost video processing-based approach to neonatal apnoea detec- tion. In2014 IEEE Workshop on Biometric Measurements and Systems for Security and Medical Applications (BIOMS) Proceedings, 67–73 (IEEE, 2014)

    work page 2014

  20. [20]

    & Ono, N

    Yang, C., Cheung, G., Stankovic, V ., Chan, K. & Ono, N. Sleep apnea detection via depth video and audio feature learning.IEEE Transactions on Multimed.19, 822–835 (2016)

    work page 2016

  21. [21]

    & Elphick, H

    Muhammad, U., Evans, R., Saatchi, R., Kingshott, R. & Elphick, H. P016 using non-invasive thermal imaging for apnoea detection.BMJ Open Respir. Res.6, A9–A10 (2019)

    work page 2019

  22. [22]

    & Murthy, J

    Fei, J., Pavlidis, I. & Murthy, J. Thermal vision for sleep apnea monitoring. InMedical Image Computing and Computer-Assisted Intervention–MICCAI 2009: 12th International Conference, London, UK, September 20-24, 2009, Proceedings, Part II 12, 1084–1091 (Springer, 2009). 26.Ishrak, M. S.et al.Doppler radar remote sensing of respiratory function.Front. Phys...

    work page 2009

  23. [23]

    & Villarroel, M

    Michaelis, T., Jorge, J., Bijlani, N. & Villarroel, M. VitalCSI: Contactless respiratory rate estimation using consumer-grade wi-fi channel state information.Sensors (Basel, Switzerland)26, 225 (2025). 26/29

    work page 2025

  24. [24]

    Vitazkova, D.et al.Advances in respiratory monitoring: a comprehensive review of wearable and remote technologies.Biosensors14, 90 (2024)

    work page 2024

  25. [25]

    & Chawla, D

    Maurya, L., Mahapatra, P. & Chawla, D. Non-contact breathing monitoring by integrating RGB and thermal imaging via RGB-thermal image registration.Biocybern. Biomed. Eng.41, 1107–1122 (2021)

    work page 2021

  26. [26]

    E., Visser, G

    Geertsema, E. E., Visser, G. H., Sander, J. W. & Kalitzin, S. N. Automated non-contact detection of central apneas using video.Biomed. Signal Process. Control.55, 101658 (2020)

    work page 2020

  27. [27]

    NPJ digital medicine2, 1–18 (2019)

    Villarroel, M.et al.Non-contact physiological monitoring of preterm infants in the neonatal intensive care unit. NPJ digital medicine2, 1–18 (2019)

    work page 2019

  28. [28]

    reports10, 18529 (2020)

    Villarroel, M.et al.Non-contact vital-sign monitoring of patients undergoing haemodialysis treatment.Sci. reports10, 18529 (2020)

    work page 2020

  29. [29]

    & Taati, B

    Akbarian, S., Montazeri Ghahjaverestan, N., Yadollahi, A. & Taati, B. Distinguishing obstructive versus central apneas in infrared video of sleep using deep learning: Validation study.J. Med. Internet Res.22, e17252 (2020)

    work page 2020

  30. [30]

    Pinilla, L., Chai-Coetzer, C. L. & Eckert, D. J. Diagnostic modalities in sleep disordered breathing: Current and emerging technology and its potential to transform diagnostics.Respirology30, 286–302 (2025)

    work page 2025

  31. [31]

    Jorge, J.et al.Non-contact physiological monitoring of post-operative patients in the intensive care unit.NPJ digital medicine5, 4 (2022)

    work page 2022

  32. [32]

    Coleman, J.et al.Assessment of neonatal respiratory rate variability.J. Clin. Monit. Comput.36, 1869–1879 (2022)

    work page 2022

  33. [33]

    38.Lee, H.et al.A new algorithm for detecting central apnea in neonates.Physiol

    Adjei, T.et al.New method to measure interbreath intervals in infants for the assessment of apnoea and respiration.BMJ open respiratory research8, e001042 (2021). 38.Lee, H.et al.A new algorithm for detecting central apnea in neonates.Physiol. measurement33, 1 (2011)

    work page 2021

  34. [34]

    R.Video respiration monitoring: Towards remote apnea detection in the clinic

    Lorato, I. R.Video respiration monitoring: Towards remote apnea detection in the clinic. Ph.D. thesis, Eindhoven University of Technology (2021)

    work page 2021

  35. [35]

    Zhao, J., Gonzalez, F. & Mu, D. Apnea of prematurity: from cause to treatment.Eur. journal pediatrics170, 1097–1105 (2011)

    work page 2011

  36. [36]

    & Zibolen, M

    Haskova, K., Javorka, K., Javorka, M., Matasova, K. & Zibolen, M. Apnea in preterm newborns: determinants, pathophysiology, effects on cardiovascular parameters and treatment.Acta Medica Martiniana13, 5 (2014). 27/29

    work page 2014

  37. [37]

    & Allegaert, K

    Joosten, K., de Goederen, R., Pijpers, A. & Allegaert, K. Sleep related breathing disorders and indications for polysomnography in preterm infants.Early Hum. Dev.113, 114–119 (2017)

    work page 2017

  38. [38]

    pulmonology 52, 84–90 (2017)

    Levy, J.et al.Impact of hands-on care on infant sleep in the neonatal intensive care unit.Pediatr. pulmonology 52, 84–90 (2017)

    work page 2017

  39. [39]

    Selvaraju, V .et al.Continuous monitoring of vital signs using cameras: A systematic review.Sensors22, 4097 (2022)

    work page 2022

  40. [40]

    Jorge, J.et al.Assessment of signal processing methods for measuring the respiratory rate in the neonatal intensive care unit.IEEE J. Biomed. Heal. Informatics23, 2335–2346 (2019). DOI 10.1109/JBHI.2019.2898273

    work page doi:10.1109/jbhi.2019.2898273 2019

  41. [41]

    Chaichulee, S.Non-contact vital sign monitoring of pre-term infants. Ph.D. thesis, University of Oxford (2018)

    work page 2018

  42. [42]

    & Inoue, S

    Honda, J., Murakawa, M. & Inoue, S. Effect of averaging time and respiratory pause time on the measurement of acoustic respiration rate monitoring.JA Clin. Reports9, 61 (2023)

    work page 2023

  43. [43]

    Lu, W.et al.A semi-automatic method for peak and valley detection in free-breathing respiratory waveforms. Med. physics33, 3634–3636 (2006)

    work page 2006

  44. [44]

    V ., Fort, P., Miller, T

    Pergolizzi Jr, J. V ., Fort, P., Miller, T. L., LeQuang, J. A. & Raffa, R. B. The limited management options for apnoea of prematurity.J. Clin. Pharm. Ther.47, 396–401 (2022)

    work page 2022

  45. [45]

    Meas.37, 564–579 (2016)

    Joshi, R.et al.Pattern discovery in critical alarms originating from neonates under intensive care.Physiol. Meas.37, 564–579 (2016)

    work page 2016

  46. [46]

    BlazePose: On-device Real-time Body Pose tracking

    Bazarevsky, V .et al.Blazepose: On-device real-time body pose tracking.arXiv preprint arXiv:2006.10204 (2020). 52.Villarroel, M.Non-contact vital sign monitoring in the clinic. Ph.D. thesis, University of Oxford (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  47. [47]

    & Tarassenko, L

    Carter, J., Jorge, J., Venugopal, B., Gibson, O. & Tarassenko, L. Deep learning-enabled sleep staging from vital signs and activity measured using a near-infrared video camera. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5940–5949 (2023)

    work page 2023

  48. [48]

    M.et al.Comparison of techniques for respiratory rate extraction from electrocardiogram and photoplethysmogram.Sensors25, 5136 (2025)

    Ponsiglione, A. M.et al.Comparison of techniques for respiratory rate extraction from electrocardiogram and photoplethysmogram.Sensors25, 5136 (2025)

    work page 2025

  49. [49]

    & Asl, B

    Zarei, A. & Asl, B. M. Automatic classification of apnea and normal subjects using new features extracted from hrv and ecg-derived respiration signals.Biomed. Signal Process. Control.59, 101927 (2020). 28/29

    work page 2020

  50. [50]

    H.et al.Breathing rate estimation from the electrocardiogram and photoplethysmogram: A review

    Charlton, P. H.et al.Breathing rate estimation from the electrocardiogram and photoplethysmogram: A review. IEEE reviews biomedical engineering11, 2–20 (2017)

    work page 2017

  51. [51]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11976–11986 (2022)

    Liu, Z.et al.A convnet for the 2020s. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11976–11986 (2022)

    work page 2022

  52. [52]

    & Tarassenko, L

    Serame, D., Villarroel, M. & Tarassenko, L. Deep learning-based detection of apnoea-related respiratory pauses in pre-term infants.medRxiv(2026). DOIPaper in preparation/archive

    work page 2026

  53. [53]

    Li, Q., Mark, R. G. & Clifford, G. D. Robust heart rate estimation from multiple asynchronous noisy sources using signal quality indices and a kalman filter.Physiol. measurement29, 15 (2007)

    work page 2007

  54. [54]

    M., Holsti, L

    Ahani, S., Niknafs, N., Lavoie, P. M., Holsti, L. & Dumont, G. A. Video-based respiratory rate estimation for infants in the nicu.IEEE J. Transl. Eng. Heal. Medicine(2024)

    work page 2024

  55. [55]

    & Mark, R

    Zong, W., Heldt, T., Moody, G. & Mark, R. An open-source algorithm to detect onset of arterial blood pressure pulses. InComputers in Cardiology, 2003, 259–262 (IEEE, 2003)

    work page 2003

  56. [56]

    & Tompkins, W

    Pan, J. & Tompkins, W. J. A real-time qrs detection algorithm.IEEE transactions on biomedical engineering 230–236 (1985). 29/29

    work page 1985