Pith. sign in

REVIEW 3 major objections 7 minor 300 references

CITRUS claims that motion-corrupted heart-rate estimation from ring PPG is best solved by scoring a dense set of frequency candidates, decoding them with a causal Viterbi path, and reporting only the windows the model trusts, cutting motion

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 · deepseek-v4-flash

2026-08-01 14:13 UTC pith:YQJ4CBUO

load-bearing objection A careful, honest engineering thesis on motion-robust wearable HR; the full-coverage estimator gains are real, but the 6.2 BPM headline is a test-set-quantile operating point, not a calibrated deployable result. the 3 major comments →

arxiv 2607.18818 v1 pith:YQJ4CBUO submitted 2026-07-21 cs.HC

CITRUS: Candidate Inference and Temporal-tracking for Reliable, Unobtrusive Sensing of Wearable Heart Rate under Motion

Yi Wang This is my paper
classification cs.HC
keywords photoplethysmographyheart rate estimationmotion artifactsViterbi decodingselective predictionwearable sensingcandidate scoringreliability-aware reporting
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 thesis tackles the single failure mode that makes wearable heart-rate monitoring unreliable: motion corrupts PPG signals so the cardiac frequency is hard to pick out. The author proposes a two-layer system, CITRUS, that first turns each time window into roughly 140 candidate heart-rate values, scores them with a small learned model, and traces a physiologically smooth path through them with a causal Viterbi decoder. A second reporting layer estimates reliability and learns an accept/hold/reject policy, so the device can stay silent when uncertain. The payoff is a large drop in motion-window error—from 10.8 to 6.2 BPM mean absolute error when only the most confident 50% of motion windows are reported—while the estimator stays under 100k parameters and fits a microcontroller-class compute budget. A sympathetic reader would see this as a practical blueprint for making wrist or ring HR sensing trustworthy during everyday movement.

Core claim

The central claim is that motion-robust heart-rate estimation should be decomposed into four explicit stages—candidate construction, candidate scoring, temporal decoding, and reliability-aware reporting—rather than solved end-to-end. The author shows that a DSP-generated candidate set of about 140 plausible BPM values per window, scored by a small MLP that weighs agreement among independent estimators and harmonic structure, and then decoded by a causal Viterbi algorithm with a heart-rate-level-dependent transition penalty, reduces full-coverage motion MAE from 14.39 BPM (LSTM baseline) to 10.91 BPM on a ring PPG dataset. Adding a reliability model and a learned finite-state reporting policy

What carries the argument

The central object is the candidate set C_t, roughly 140 frequency candidates per analysis window derived from Welch spectral peaks, autocorrelation peaks, and peak-interval estimates on both PPG channels, with harmonic variants and physiological range filtering. Around that set sit three coupled components: a learned candidate scorer (a small MLP mapping source, channel, spectral power, consensus, harmonic, and temporal features to candidate logits, trained with a soft target centered on the reference HR), a causal Viterbi decoder that selects a minimum-cost path through candidate graphs with a transition penalty that scales with heart-rate level and candidate probability, and a reporting l

Load-bearing premise

The learned candidate scorer and reliability model are trained on reference heart-rate labels from the same capture protocol and evaluated leave-one-subject-out on a single 31-subject ring subset with only 267 motion windows, so the reported error reductions depend on those training labels being representative of the motion conditions the system will actually face.

What would settle it

Run the same estimation and reporting pipeline on a held-out motion protocol that includes activity types not in the deep-squat/striding/shaking-head set (e.g., cycling or rowing), and check whether the 50%-coverage motion MAE still falls from ~10.8 to ~6.2 BPM; if the reliability model's rankings are no better than random chance on unseen motion types, the selective-reporting gain would disappear, which would directly falsify the central claim that the learned reliability score transfers across motion conditions.

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

If this is right

  • If the central claim holds, wearable rings and wrist devices can offer heart-rate estimates that stay accurate during motion by simply refusing to report when uncertain, trading coverage for error in a principled, policy-learned way.
  • The demonstration that a causal, past-only Viterbi decoder matches offline full-sequence decoding (10.66 vs 10.88 BPM motion MAE) means real-time, on-device deployment can capture most of the temporal-consistency benefit without waiting for future windows.
  • The ablation finding that the raw-waveform CNN encoder adds no measurable accuracy once candidate features are present suggests that future wearable HR models can shrink to roughly 28k parameters and still retain the motion robustness, easing memory and power constraints.
  • The explicit separation of estimation from reporting gives wearable designers a coverage–error dial controlled by a rejection cost, rather than a single fixed accuracy number, which is how reliability should be stated for real-world use.
  • On an independent wrist dataset (PPG-DaLiA), the same estimation layer improves full-coverage motion MAE from 14.22 to 12.62 BPM, indicating the candidate-plus-decoding mechanism transfers beyond the original ring sensor, even if the reporting calibration must be redone per setting.

Where Pith is reading between the lines

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

  • Editorial inference: The paper's bottleneck—candidate identifiability at 1-of-140 being only modestly above chance—suggests the next big gain will come from better candidate generation (e.g., fusing a motion-adaptive spectral cleaning step) rather than from better scoring or decoding, since the ceiling of the current pipeline is set by whether the true BPM is even in the candidate set.
  • Editorial inference: The observed over-smoothing in the 3–5 BPM transition bin (decoding increasing MAE there by ~3 BPM) implies a fixed quadratic transition penalty is too rigid; a reliability- or motion-adaptive transition cost, as the author lists for future work, could recover those errors and likely push full-coverage motion MAE below 10 BPM.
  • Editorial inference: The selective-reporting results are an argument that the field should report coverage–error curves rather than single MAE numbers, because a wearable that can silently wait out a motion burst is more clinically useful than one that always emits a noisy value; this could be tested by simulated patient-monitoring scenarios that penalize both error and silence.
  • Editorial inference: Since ACC features helped candidate generation but not reliability scoring, a sensor-fusion strategy that consumes motion data early (for segment selection) but not late (for confidence) may be a general design principle for motion-tolerant PPG, though this is the author's observation and my generalization.

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 paper presents CITRUS, a two-layer system for heart-rate estimation from motion-corrupted ring PPG. The estimation layer generates ~140 frequency candidates per window via DSP methods, scores them with a small MLP, and applies causal Viterbi decoding; the reporting layer uses an HGB reliability model and a learned accept/hold/reject policy. On a 31-subject ring subset with 267 motion windows, the full-coverage causal Viterbi estimator reduces motion MAE from 12.94 to 10.91 BPM relative to independent per-window scoring, and selective reporting of the most reliable 50% of motion windows yields 6.2 BPM. On PPG-DaLiA, the same candidate/Viterbi estimator (retrained on DaLiA) lowers full-coverage motion MAE from 14.22 to 12.62 BPM. The paper also reports multi-vital benchmark results for two ring variants across HR, RR, SpO2, and BP, plus ablations of candidate source, ACC use, raw-waveform encoder, decoder latency, and reporting policy.

Significance. If the headline operating point is reproducible, the paper offers a transparent, low-compute alternative to end-to-end deep models, with the unusual virtue of reporting a negative result (Table 10) and naming its own bottleneck (candidate identifiability). The ablations separating candidate-source structure, ACC role, raw-waveform encoder redundancy, and causal-vs-offline decoding are genuinely informative, and the explicit limitation table is a strength. However, the central selective-reporting result is computed with a test-set-derived threshold, lacks error bars, and is not produced by the deployable finite-state policy; the temporal decoder is shown to hurt a substantial subset of windows; and the paper's own limitation statement calls candidate identifiability 'the open bottleneck' without providing a chance-corrected selection metric. These issues must be resolved before the claimed contribution is established.

major comments (3)
  1. [§5.2, Eq. (12)–(13)] The 50% coverage operating point (10.8 → 6.2 BPM) is obtained by setting the threshold at the empirical quantile q_{1−α}(ρ) of reliability scores computed on the same test windows used to evaluate MAE(S_α). No validation-based threshold-selection procedure is described, and no confidence interval is reported for the 6.2 BPM figure. This is an in-sample selection of the reporting set, so the headline is not a reproducible operating point. Moreover, Table 6, the deployable finite-state policy, reports motion-coverage rows 0.87 and 0.36 with no 0.50 row, so the headline number is not generated by the learned policy. Please provide a threshold-calibration protocol on a separate validation split, seed-level error bars, and a policy-based coverage–MAE curve that includes a 50% motion-coverage point.
  2. [§6.3, Table 10] The paper's own stratified analysis shows that causal Viterbi decoding increases motion-window MAE in the adjacent-window reference-change bin of 3–5 BPM (n=120; MAE 8.55 → 11.53, gain −2.98), while improving the other bins. This is a substantial negative effect on a nontrivial subset of windows. The abstract and contribution list present temporal decoding as a main component without this caveat. Since the decoder is a core layer, the paper should either introduce a decoder variant that avoids over-smoothing moderate changes or explicitly bound the claimed aggregate benefit with this over-smoothing cost separated; at minimum, the aggregate gain should be reported together with the stratification so readers can weigh the trade-off.
  3. [§6.3, Limitation 1; §4.2] The limitations state that candidate identifiability is '1-of-≈140 selection, only modestly above chance' and call this the open bottleneck. Yet the paper does not report a chance-corrected top-k selection accuracy, so it is unclear how much of the 10.91 BPM motion MAE comes from the learned scorer versus from the density of the candidate grid. The uniform-grid control in §4.2 shows that a 160-state grid alone achieves 0.32 BPM coverage MAE (better than the DSP set's 0.6 BPM), which underscores the need for a selection-quality metric. Please add a top-k selection metric with a random-selection baseline as a primary evaluation of the estimation layer; this is directly requested by the authors' own limitation statement.
minor comments (7)
  1. [§4.4] Section 4.4 reports 'full model 10.73' motion MAE, while Table 5 lists 'DSP candidates IR/red/ACC' as 10.91±0.73. Please reconcile the numbers or state that the 2×2 ablation uses a different feature set or seed configuration.
  2. [Tables 6 and 9] The finite-state policy results are reported as point estimates from three-seed runs without standard deviations, unlike the five-seed mean±SD format used elsewhere. Please add seed-level variability.
  3. [Abstract; §6.1] The phrase 'independent wrist-BVP cohort' for PPG-DaLiA could be misread as cross-dataset transfer of the ring-trained model. Section 6.1 states the estimator is retrained on DaLiA with LOSO and no ring-trained weights are used; please clarify this in the abstract.
  4. [§4.1] Candidate-set coverage MAE (0.6 BPM) is compared to a 160-state uniform grid (0.32 BPM) without noting that coverage depends on grid density. Consider reporting coverage as a function of grid density or explicitly stating that this metric is not a deployable-quality metric.
  5. [§5.5] The deployment section reports 'estimated runtime' and 'fits a microcontroller-class budget' without on-device measurements. The future-work section already mentions hardware validation; please state explicitly in §5.5 that these are analytical estimates.
  6. [References] The reference list contains an inconsistent capitalization: 'Gao, h.-a.' in Wang et al. 2025. Please normalize.
  7. [Eq. (8)] The transition cost equation has many constants (s0, h0, amin, amax, σmin, σmax, Ωmax, dmax, astep, λjump, λtr). A table of symbols and values would improve reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the estimation and reporting systems are trained on labels but evaluated under LOSO and on independent PPG-DaLiA; the headlined 6.2 BPM is an explicit coverage operating point, not a parameter-free prediction.

full rationale

The paper's derivation chain is not circular. The candidate scorer (Eq. 5) and reliability model (Eq. 10) are trained against reference HR labels, but the central evaluations use leave-one-subject-out splits and the estimation layer is additionally evaluated on the independent PPG-DaLiA wrist dataset with subject-disjoint training. The candidate-coverage metric is explicitly controlled by comparing DSP candidates against uniform-grid candidates and shuffled references, so candidate coverage is not conflated with selection quality. The reported 6.2 BPM figure at 50% coverage comes from Eq. (12)-(13), where the threshold is the empirical reliability quantile on the test windows; this is a transparently stated reporting operating point rather than a calibrated deployment prediction, and the paper itself warns that low-coverage rows are operating points and that cross-cohort calibration remains future work. The τ-Ring hardware/dataset citation is a self-citation, but it supplies the sensing platform and data, not the algorithmic result, so it is not load-bearing circularity. No fitted parameter is renamed as a prediction, no uniqueness theorem from the authors is invoked, and no ansatz is smuggled in via citation. The honest limitations section further weakens any circularity concern by identifying candidate identifiability and temporal over-smoothing as open bottlenecks.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entity or conserved quantity. The central claim depends on a series of tuned parameters (Viterbi constants, policy costs, soft-target temperature) and domain assumptions (accuracy of clinical references, representativeness of the 267-window motion subset, validity of the quadratic transition penalty). The uniform-grid control is an honest attempt to discount the candidate-coverage confound, but the candidate set itself remains an artifact of the DSP source functions.

free parameters (4)
  • Candidate-set size / grid 160 states over 35–220 BPM = 160 states; DSP candidates ≈140 per window
    The uniform-grid control (Eq. 3) fixes the number of candidate states to 160, and the DSP candidate set produces ≈140 candidates per window; candidate-set size affects coverage MAE and the scorer's selection difficulty.
  • Viterbi transition parameters = s0=10, h0=80, amin=0.75, amax=1.8, σmin=6, σmax=30, Ωmax=9, dmax=25, astep=0.8, λjump=0.25
    These constants in Eq. 8 are fixed by hand (except λtr, which is selected on the training fold); they control the transition penalty shape and affect both temporal smoothing and the over-smoothing cost in the 3–5 BPM bin.
  • Reliability temperature τr and soft-target temperature τ = not stated explicitly in text
    The reliability target r_t = exp(-e_t/τr) in Eq. 10 and the soft candidate target q_t,k in Eq. 5 depend on temperature scales that are not reported; these affect ranking calibration and the reported coverage–MAE trade-off.
  • Policy rejection cost λrej and hold penalty β = λrej swept from 30 to 2; β not reported
    The finite-state policy operating point is selected by tuning λrej (Tables 6, 9); the entire coverage–error curve is a function of this hand-set cost, not of a measured external constraint.
axioms (5)
  • domain assumption Reference HR from finger oximeter, respiratory band, BP monitor, and VivaLNK ECG patch is accurate enough to serve as the training target and evaluation ground truth for ring HR/PPG.
    The training losses in Eqs. 5 and 6 and the reliability target in Eq. 10 are defined against y_t; the benchmark's validity depends on these clinical references being synchronized and accurate during motion. Entered in Sections 3.2–3.4.
  • domain assumption Motion windows in the ring subset (n=267) are representative of the motion conditions the system will face in deployment.
    The headline motion improvements rest on a small 267-window subset from deep-squat, striding, and shaking-head protocols; the paper itself cites candidate identifiability as the open bottleneck. Entered in Section 3.4.
  • domain assumption Physiological heart-rate transitions are well modeled by the specified quadratic transition penalty with saturated bounds (Eq. 8).
    The causal Viterbi decoder's improvement depends on this smoothness prior; the paper's own Table 10 shows the penalty over-smooths for 3–5 BPM reference changes, so the prior is only partially valid.
  • standard math Leave-one-subject-out evaluation with five seeds gives a fair performance estimate of the learned scorer/reliability model.
    The scorer, reliability model, and policy are evaluated LOSO; this assumes each subject is independent and that the 31-subject, 2009-window split is sufficient for unbiased estimates.
  • domain assumption The DSP candidate generation (Welch peaks, autocorrelation peaks, peak-interval estimates, harmonic variants) covers the true HR in each motion window.
    The candidate-set coverage MAE (0.6 BPM in the paper) is measured against the same labels used to train the scorer; the paper uses uniform-grid controls to bound this, but the DSP candidate coverage is fundamental to the estimation layer's ceiling. Entered in Section 4.1, Eq. 2.

pith-pipeline@v1.3.0-alltime-deepseek · 35 in / 9495 out tokens · 66737 ms · 2026-08-01T14:13:50.733412+00:00 · methodology

0 comments
read the original abstract

Wearable photoplethysmography (PPG) provides continuous heart-rate measurements, but its accuracy degrades under motion. In the ring-platform benchmark, the best supervised baseline reaches 5.33 BPM mean absolute error (MAE) on the overall heart-rate task. In the motion-focused ring-only audit, a supervised LSTM baseline reaches $14.39 \pm 0.47$ BPM MAE on motion windows, and simple smoothing and ACC priors reduce this only to $13.00 \pm 0.41$ BPM. This thesis addresses motion-corrupted HR estimation through three connected stages: candidate-frequency identification, temporal decoding, and reliability-aware reporting. The study first evaluates two wearable PPG ring variants against clinical references for heart rate, respiration, SpO$_2$, and blood pressure in 54 participants. The proposed system then uses two layers. The estimation layer converts each window into approximately 140 frequency candidates, scores candidates using agreement among independent estimators, and applies causal Viterbi decoding with a physiological transition penalty. The reporting layer estimates reliability and applies a learned accept/hold/reject policy. Reporting only the most-confident 50% of motion windows reduces motion MAE from 10.8 to 6.2 BPM. The heart-rate estimator uses fewer than 100k parameters and runs within a microcontroller-class compute budget. Additional PPG-DaLiA experiments evaluate the same candidate-selection and temporal-decoding estimator on an independent wrist-BVP cohort.

Figures

Figures reproduced from arXiv: 2607.18818 by Yi Wang.

Figure 1
Figure 1. Figure 1: Ring design and sensor layout. (A) Physical struc￾ture of the ring. (D) Internal components including battery and chip. (B) and (E) Sensor configuration for Ring 1 (reflec￾tive). (C) and (F) Sensor configuration for Ring 2 (transmis￾sive). 3.3 Benchmark and Multi-Vital Results [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] 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

300 extracted references · 14 canonical work pages · 1 internal anchor

  1. [1]

    Pyphysio:

    Bizzego, Andrea and Battisti, Alessandro and Gabrieli, Giulio and Esposito, Gianluca and Furlanello, Cesare , year =. Pyphysio:. SoftwareX , volume =. doi:10.1016/j.softx.2019.100287 , langid =

    arXiv 2019

  2. [2]

    IEEE BSN , year=

    Non-Contact Health Monitoring During Daily Personal Care Routines , author=. IEEE BSN , year=

  3. [3]

    Scientific Data , volume=

    A high altitude respiration and spo2 dataset for assessing the human response to hypoxia , author=. Scientific Data , volume=. 2024 , publisher=

    2024

  4. [4]

    Scientific Data , volume=

    Understanding occupants’ behaviour, engagement, emotion, and comfort indoors with heterogeneous sensors and wearables , author=. Scientific Data , volume=. 2022 , publisher=

    2022

  5. [5]

    arXiv preprint arXiv:2505.04172 , year=

    A Dataset and Toolkit for Multiparameter Cardiovascular Physiology Sensing on Rings , author=. arXiv preprint arXiv:2505.04172 , year=. doi:10.48550/arXiv.2505.04172 , url=

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.04172

  6. [6]

    and Leabman, Michael and Shumate, Joe and Tompkins, Kim L

    Mastrototaro, John J. and Leabman, Michael and Shumate, Joe and Tompkins, Kim L. , year =. Performance of a. JMIR Formative Research , volume =. doi:10.2196/54256 , urldate =

    work page doi:10.2196/54256

  7. [7]

    Breteler, Martine J. M. and Leigard, Ellen and Hartung, Lisa C. and Welch, John R. and Brealey, David A. and Fritsch, Sebastian J. and Konrad, David and Hertzberg, Daniel and Bell, Max and Rienstra, Heleen and Rademakers, Frank E. and Kalkman, Cor J. , year =. Reliability of an All-in-One Wearable Sensor for Continuous Vital Signs Monitoring in High-Risk ...

  8. [8]

    Detection of

    Kwon, Soonil and Hong, Joonki and Choi, Eue-Keun and Lee, Byunghwan and Baik, Changhyun and Lee, Euijae and Jeong, Eui-Rim and Koo, Bon-Kwon and Oh, Seil and Yi, Yung , year =. Detection of. Journal of Medical Internet Research , volume =. doi:10.2196/16443 , urldate =

    work page doi:10.2196/16443

  9. [9]

    Continuous

    Haveman, Marjolein E and. Continuous. 2022 , month = jan, journal =. doi:10.2196/30863 , urldate =

    work page doi:10.2196/30863 2022

  10. [10]

    American heart journal , volume=

    The circulatory effects of squatting , author=. American heart journal , volume=. 1962 , publisher=

    1962

  11. [11]

    Anesthesia & Analgesia , volume=

    Effects of acute, profound hypoxia on healthy humans: implications for safety of tests evaluating pulse oximetry or tissue oximetry performance , author=. Anesthesia & Analgesia , volume=. 2017 , publisher=

    2017

  12. [12]

    Sensors , volume =

    Joshi, Jitesh and Wang, Katherine and Cho, Youngjun , year =. Sensors , volume =. doi:10.3390/s23198244 , copyright =

    work page doi:10.3390/s23198244

  13. [13]

    and Pensabene, Virginia and Actis, Paolo , year =

    Benedetto, Simone and Caldato, Christian and Bazzan, Elia and Greenwood, Darren C. and Pensabene, Virginia and Actis, Paolo , year =. Assessment of the. PLOS ONE , volume =. doi:10.1371/journal.pone.0192691 , urldate =

    work page doi:10.1371/journal.pone.0192691

  14. [14]

    2023 , month = mar, journal =

    A Benchmark for Machine-Learning Based Non-Invasive Blood Pressure Estimation Using Photoplethysmogram , author =. 2023 , month = mar, journal =. doi:10.1038/s41597-023-02020-6 , urldate =

    doi:10.1038/s41597-023-02020-6 2023

  15. [15]

    Analysing

    van Gent, Paul and Farah, Haneen and van Nes, Nicole and van Arem, Bart , year =. Analysing. Journal of Open Research Software , volume =. doi:10.5334/jors.241 , urldate =

    work page doi:10.5334/jors.241

  16. [16]

    Behavior research methods , pages=

    NeuroKit2: A Python toolbox for neurophysiological signal processing , author=. Behavior research methods , pages=. 2021 , publisher=

    2021

  17. [17]

    The Quest for Raw Signals: A Quality Review of Publicly Available Photoplethysmography Datasets , shorttitle =

    Wolling, Florian and Van Laerhoven, Kristof , year =. The Quest for Raw Signals: A Quality Review of Publicly Available Photoplethysmography Datasets , shorttitle =. Proceedings of the. doi:10.1145/3419016.3431485 , urldate =

    arXiv

  18. [18]

    Progress in brain research , volume=

    Cardiorespiratory coupling: common rhythms in cardiac, sympathetic, and respiratory activities , author=. Progress in brain research , volume=. 2014 , publisher=

    2014

  19. [19]

    doi:10.1016/j.trf.2019.09.015 , urldate =

    2019 , month = oct, journal =. doi:10.1016/j.trf.2019.09.015 , urldate =

    doi:10.1016/j.trf.2019.09.015 2019

  20. [20]

    and Zhang, Y.T

    Zheng, D.C. and Zhang, Y.T. , year =. A Ring-Type Device for the Noninvasive Measurement of Arterial Blood Pressure , booktitle =. doi:10.1109/IEMBS.2003.1280819 , urldate =

    Pith/arXiv arXiv 2003

  21. [21]

    and Park, Seung Woo , year =

    Kim, Jihoon and Chang, Sung-A. and Park, Seung Woo , year =. First-in-. Journal of Korean Medical Science , volume =. doi:10.3346/jkms.2024.39.e18 , urldate =

    work page doi:10.3346/jkms.2024.39.e18 2024

  22. [22]

    Anesthesia & Analgesia , volume=

    Takuo Aoyagi: discovery of pulse oximetry , author=. Anesthesia & Analgesia , volume=. 2007 , publisher=

    2007

  23. [23]

    Morgado Areia, Carlos and Santos, Mauro and Vollam, Sarah and Pimentel, Marco and Young, Louise and Roman, Cristian and Ede, Jody and Piper, Philippa and King, Elizabeth and Gustafson, Owen and Harford, Mirae and Shah, Akshay and Tarassenko, Lionel and Watkinson, Peter , year =. A. Journal of Medical Internet Research , volume =. doi:10.2196/27547 , urldate =

    work page doi:10.2196/27547

  24. [24]

    and Sargent, Charli and Roach, Gregory D

    Miller, Dean J. and Sargent, Charli and Roach, Gregory D. , year =. A. Sensors , volume =. doi:10.3390/s22166317 , urldate =

    work page doi:10.3390/s22166317

  25. [25]

    and Ulman, Hana K

    Stone, Jason D. and Ulman, Hana K. and Tran, Kaylee and Thompson, Andrew G. and Halter, Manuel D. and Ramadan, Jad H. and Stephenson, Mark and Finomore, Victor S. and Galster, Scott M. and Rezai, Ali R. and Hagen, Joshua A. , year =. Assessing the. Frontiers in Sports and Active Living , volume =. doi:10.3389/fspor.2021.585870 , urldate =

    arXiv 2021

  26. [26]

    Fiore, Matteo and Bianconi, Alessandro and Sicari, Gaia and Conni, Alice and Lenzi, Jacopo and Tomaiuolo, Giulia and Zito, Flavia and Golinelli, Davide and Sanmarchi, Francesco , year =. The. Applied Sciences , volume =. doi:10.3390/app142310778 , urldate =

    work page doi:10.3390/app142310778

  27. [27]

    Scientific Data , volume =

    Ho, Matthew Yiwen and Pham, Hung Manh and Saeed, Aaqib and Ma, Dong , year =. Scientific Data , volume =. doi:10.1038/s41597-025-04453-7 , urldate =

    work page doi:10.1038/s41597-025-04453-7

  28. [28]

    , year =

    Jarchi, Delaram and Casson, Alexander J. , year =. Description of a. Data , volume =. doi:10.3390/data2010001 , urldate =

    doi:10.3390/data2010001

  29. [29]

    Pervasive

    Kalanadhabhatta, Manasa and Min, Chulhong and Montanari, Alessandro and Kawsar, Fahim , editor =. Pervasive. 2022 , pages =. doi:10.1007/978-3-030-99194-4_14 , isbn =

    work page doi:10.1007/978-3-030-99194-4_14 2022

  30. [30]

    and Chung, H

    Lee, H. and Chung, H. and Lee, J. , year =. Motion. IEEE Sensors Journal , volume =. doi:10.1109/JSEN.2018.2879970 , urldate =

    arXiv 2018

  31. [31]

    Advances in Neural Information Processing Systems , volume =

    Meier, Manuel and Demirel, Berken Utku and Holz, Christian , year =. Advances in Neural Information Processing Systems , volume =

  32. [32]

    Scientific Data , volume =

    Montanari, Alessandro and Ferlini, Andrea and Balaji, Ananta Narayanan and Mascolo, Cecilia and Kawsar, Fahim , year =. Scientific Data , volume =. doi:10.1038/s41597-023-02762-3 , urldate =

    work page doi:10.1038/s41597-023-02762-3

  33. [33]

    Reiss, Attila and Indlekofer, Ina and Schmidt, Philip and Van Laerhoven, Kristof , year =. Deep. Sensors , volume =. doi:10.3390/s19143079 , urldate =

    doi:10.3390/s19143079

  34. [34]

    Introducing

    Schmidt, Philip and Reiss, Attila and Duerichen, Robert and Marberger, Claus and Van Laerhoven, Kristof , year =. Introducing. Proceedings of the 20th. doi:10.1145/3242969.3242985 , urldate =

    arXiv

  35. [35]

    Computing with Smart Rings: A Systematic Literature Review

    Wang, Zeyu and Yu, Ruotong and Wang, Xiangyang and Ding, Jiexin and Tang, Jiankai and Fang, Jun and He, Zhe and Li, Zhuojun and R. Computing with. 2025 , month = feb, number =. doi:10.48550/arXiv.2502.02459 , urldate =. 2502.02459 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.02459 2025

  36. [36]

    IEEE Transactions on Biomedical Engineering , volume =

    Zhang, Zhilin and Pi, Zhouyue and Liu, Benyuan , year =. IEEE Transactions on Biomedical Engineering , volume =. doi:10.1109/TBME.2014.2359372 , urldate =. 1409.5181 , primaryclass =

    arXiv 2014

  37. [37]

    Artificial Intelligence Review , volume=

    Handling imbalanced medical datasets: review of a decade of research , author=. Artificial Intelligence Review , volume=. 2024 , publisher=

    2024

  38. [38]

    Computational Intelligence in Healthcare Informatics , pages=

    Learning from imbalanced data in healthcare: State-of-the-art and research challenges , author=. Computational Intelligence in Healthcare Informatics , pages=. 2024 , publisher=

    2024

  39. [39]

    IEEE DataPort , year=

    A Dataset of Synchronized Signals from Wearable Cardiovascular Monitoring Sensors , author=. IEEE DataPort , year=

  40. [40]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  41. [41]

    Data Mining and Knowledge Discovery , volume=

    Inceptiontime: Finding alexnet for time series classification , author=. Data Mining and Knowledge Discovery , volume=. 2020 , publisher=

    2020

  42. [42]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  43. [43]

    arXiv preprint arXiv:2312.00752 , year=

    Mamba: Linear-time sequence modeling with selective state spaces , author=. arXiv preprint arXiv:2312.00752 , year=

    Pith/arXiv arXiv

  44. [44]

    Scientific data , volume=

    MIMIC-III, a freely accessible critical care database , author=. Scientific data , volume=. 2016 , publisher=

    2016

  45. [45]

    Nature , volume=

    Synchronization and rhythmic processes in physiology , author=. Nature , volume=. 2001 , publisher=

    2001

  46. [46]

    2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) , pages=

    Mmpd: Multi-domain mobile video physiology dataset , author=. 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) , pages=. 2023 , organization=

    2023

  47. [47]

    Advances in Neural Information Processing Systems , volume=

    rppg-toolbox: Deep remote ppg toolbox , author=. Advances in Neural Information Processing Systems , volume=

  48. [48]

    ACL 24 , year=

    Longbench: A bilingual, multitask benchmark for long context understanding , author=. ACL 24 , year=

  49. [49]

    AI Health Summit 23 , year=

    ALPHA: AnomaLous Physiological Health Assessment Using Large Language Models , author=. AI Health Summit 23 , year=

  50. [50]

    Neural Networks , year=

    Spiking-PhysFormer: Camera-Based Remote Photoplethysmography with Parallel Spike-driven Transformer , author=. Neural Networks , year=

  51. [51]

    2024 IEEE International Symposium on Biomedical Imaging (ISBI) , pages=

    A Comprehensive Dataset and Automated Pipeline for Nailfold Capillary Analysis , author=. 2024 IEEE International Symposium on Biomedical Imaging (ISBI) , pages=. 2024 , organization=

    2024

  52. [52]

    CHI'24 Workshop PhysioCHI'24 , year=

    Camera-Based Remote Physiology Sensing for Hundreds of Subjects Across Skin Tones , author=. CHI'24 Workshop PhysioCHI'24 , year=

  53. [53]

    Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=

    The earsavas dataset: Enabling subject-aware vocal activity sensing on earables , author=. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=. 2024 , publisher=

    2024

  54. [54]

    The 21st IEEE International Conference on Ubiquitous Intelligence and Computing (UIC 2024) , year=

    Summit Vitals: Multi-Camera and Multi-Signal Biosensing at High Altitudes , author=. The 21st IEEE International Conference on Ubiquitous Intelligence and Computing (UIC 2024) , year=

    2024

  55. [55]

    arXiv preprint arXiv:2411.15283 , year=

    A Plug-and-Play Temporal Normalization Module for Robust Remote Photoplethysmography , author=. arXiv preprint arXiv:2411.15283 , year=

    Pith/arXiv arXiv

  56. [56]

    The 21st IEEE International Conference on Ubiquitous Intelligence and Computing (UIC 2024) , year=

    Camera-Based HRV Prediction for Remote Learning Environments , author=. The 21st IEEE International Conference on Ubiquitous Intelligence and Computing (UIC 2024) , year=

    2024

  57. [57]

    arXiv preprint arXiv:2503.01699 , year=

    Camera Measurement of Blood Oxygen Saturation , author=. arXiv preprint arXiv:2503.01699 , year=

    Pith/arXiv arXiv

  58. [58]

    arXiv preprint arXiv:2503.23930 , year=

    Exploring Reliable PPG Authentication on Smartwatches in Daily Scenarios , author=. arXiv preprint arXiv:2503.23930 , year=

    Pith/arXiv arXiv

  59. [59]

    arXiv preprint arXiv:2504.01774 , year=

    Memory-efficient Low-latency Remote Photoplethysmography through Temporal-Spatial State Space Duality , author=. arXiv preprint arXiv:2504.01774 , year=

    Pith/arXiv arXiv

  60. [60]

    ACM Computing Surveys (CSUR) , volume=

    Digital digits: A comprehensive survey of finger augmentation devices , author=. ACM Computing Surveys (CSUR) , volume=. 2015 , publisher=

    2015

  61. [61]

    NPJ Digital Medicine , volume=

    Biometrics of complete human pregnancy recorded by wearable devices , author=. NPJ Digital Medicine , volume=. 2024 , publisher=

    2024

  62. [62]

    Proceedings of the 2018 CHI conference on human factors in computing systems , pages=

    Observations on typing from 136 million keystrokes , author=. Proceedings of the 2018 CHI conference on human factors in computing systems , pages=

    2018

  63. [63]

    The 34th Annual ACM Symposium on User Interface Software and Technology , pages=

    GestuRING: A web-based tool for designing gesture input with rings, ring-like, and ring-ready devices , author=. The 34th Annual ACM Symposium on User Interface Software and Technology , pages=

  64. [64]

    Distributed, Ambient, and Pervasive Interactions: First International Conference, DAPI 2013, Held as Part of HCI International 2013, Las Vegas, NV, USA, July 21-26, 2013

    Subtle, Natural and Socially Acceptable Interaction Techniques for Ringterfaces—Finger-Ring Shaped User Interfaces , author=. Distributed, Ambient, and Pervasive Interactions: First International Conference, DAPI 2013, Held as Part of HCI International 2013, Las Vegas, NV, USA, July 21-26, 2013. Proceedings 1 , pages=. 2013 , organization=

    2013

  65. [65]

    Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems , pages=

    Notiring: A comparative study of notification channels for wearable interactive rings , author=. Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems , pages=

  66. [66]

    Journal of healthcare informatics research , volume=

    A survey of smartwatches in remote health monitoring , author=. Journal of healthcare informatics research , volume=. 2018 , publisher=

    2018

  67. [67]

    Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies , volume=

    Sensing with earables: A systematic literature review and taxonomy of phenomena , author=. Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies , volume=. 2022 , publisher=

    2022

  68. [68]

    Journal of Sports Sciences , volume=

    A validation study of the WHOOP strap against polysomnography to assess sleep , author=. Journal of Sports Sciences , volume=. 2020 , publisher=

    2020

  69. [69]

    I Could Wear It All of the Time, Just Like My Wedding Ring:

    “I Could Wear It All of the Time, Just Like My Wedding Ring:” Insights into Older People’s Perceptions of Smart Rings , author=. Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems , pages=

    2023

  70. [70]

    Proceedings of the CHI Conference on Human Factors in Computing Systems , pages=

    MouseRing: Always-available Touchpad Interaction with IMU Rings , author=. Proceedings of the CHI Conference on Human Factors in Computing Systems , pages=

  71. [71]

    Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems , pages=

    Enabling voice-accompanying hand-to-face gesture recognition with cross-device sensing , author=. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems , pages=

    2023

  72. [72]

    Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=

    DRG-Keyboard: Enabling subtle gesture typing on the fingertip with dual IMU rings , author=. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=. 2023 , publisher=

    2023

  73. [73]

    Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=

    DualRing: Enabling subtle and expressive hand interaction with dual IMU rings , author=. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=. 2021 , publisher=

    2021

  74. [74]

    Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=

    Qwertyring: Text entry on physical surfaces using a ring , author=. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=. 2020 , publisher=

    2020

  75. [75]

    Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=

    Keep the phone in your pocket: Enabling smartphone operation with an imu ring for visually impaired people , author=. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=. 2020 , publisher=

    2020

  76. [76]

    Proceedings of the 32nd annual ACM symposium on user interface software and technology , pages=

    Accurate and low-latency sensing of touch contact on any surface with finger-worn IMU sensor , author=. Proceedings of the 32nd annual ACM symposium on user interface software and technology , pages=

  77. [77]

    Proceedings of the ACM on Human-Computer Interaction , volume=

    How does sleep tracking influence your life? experiences from a longitudinal field study with a wearable ring , author=. Proceedings of the ACM on Human-Computer Interaction , volume=. 2022 , publisher=

    2022

  78. [78]

    Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems , pages=

    New insights into user-defined smart ring gestures with implications for gesture elicitation studies , author=. Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems , pages=

    2023

  79. [79]

    Proceedings of the 23rd International Conference on Mobile Human-Computer Interaction , pages=

    ARO: Exploring the design of smart-ring interactions for encumbered hands , author=. Proceedings of the 23rd International Conference on Mobile Human-Computer Interaction , pages=

  80. [80]

    Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=

    Efring: Enabling thumb-to-index-finger microgesture interaction through electric field sensing using single smart ring , author=. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=. 2023 , publisher=

    2023

Showing first 80 references.