Pith. sign in

REVIEW 2 major objections 6 minor 46 references

Reassessing the Feasibility of PPG-Based Non-Invasive Blood Glucose Level Estimation

T0 review · 2 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper argues that published PPG-based blood glucose estimates have not been shown to generalize: five reproduced methods degrade to a mean-prediction baseline once participants are split out of training, while Clarke Error Grid zones s

desk verdict A useful negative result, but the collapse of the five published methods is only as strong as the re-implementations. read the letter →

arxiv 2608.01820 v1 pith:JCQO3KH3 submitted 2026-08-03 cs.HC cs.LG

Supraja Ramesh , Markus Neufeld , Michael Küttner , Tobias Röddiger , Michael Beigl This is my paper
classification cs.HCcs.LG
keywords bloodglucosemonitoringphotoplethysmographynon-invasivedataleakageevaluationmethodologyClarkeErrorGridmodelgeneralizationwearablehealth
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that current published results for non-invasive blood glucose estimation from PPG signals are driven by evaluation setup rather than real prediction skill. It reimplements five representative methods, runs them on two public datasets under three split protocols, and shows that the favorable numbers under random window-level splits collapse to near-zero or negative R² under participant-aware and leave-some-participants-out splits — about what a model that always predicts the training mean achieves. The same predictions still land over 90% of the time in Clarke Error Grid zones A and B, the same "clinically acceptable" zones prior papers cite as evidence of success. A sympathetic reader should care because this calls into question a decade of optimistic claims about a wearable glucose monitor that does not require needles, and because it offers a reusable evaluation protocol for testing future claims.

What carries the argument

The evaluation pipeline is the load-bearing object: a standardized, reproducible setup that aligns lab glucose timestamps with 16-minute PPG segments, builds five methods from their published descriptions, and evaluates each under three data-split protocols — random window-level (train and test may share windows), participant-aware (one session of a target participant is seen in training, one held out), and leave-some-participants-out (no participant overlap). The two comparison instruments are R² (whether predictions track glucose variation) and the Clarke Error Grid (whether predictions are clinically safe). The mechanism the paper exposes is that these instruments disagree: predictions th

What would settle it

If the original authors release their code and any of the five methods retains a clearly positive R² (say above 0.3) under participant-independent LSPO splits on the same public datasets, the paper's central collapse claim would be weakened to a claim about this particular reimplementation. A simpler check: train any PPG-based model on VitalDB with LSPO and see whether its R² clearly exceeds the mean-prediction baseline; the paper predicts it will not.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that two evaluation practices together have been masking failure in PPG-based blood glucose estimation. First, random window-level train/test splits let segments from the same recording or the same person appear in both training and test sets, so models learn subject- and recording-specific context rather than a glucose signal; under participant-aware and LSPO splits, all five reimplemented methods fall to near-zero or negative R², indistinguishable from a mean-prediction baseline. Second, Clarke Error Grid zone analysis reports that over 90% of predictions fall into the clinically acceptable A and B zones even for the baseline, so zone-bas

Load-bearing premise

The reimplementations of the five published methods must faithfully match the original models; since no original code was released, any divergence in training or tuning could be what collapses under strict splits rather than a property of the originals.

Editorial extensions

If this is right

  • If the paper is right, any PPG-based BGL method that reports only random-split results should be regarded as unvalidated until it passes a participant-independent split.
  • Clinical acceptability claims from Clarke Error Grid zones alone are not evidence of glucose estimation; future work should require a mean-baseline comparison before citing zone A+B percentages.
  • Because participant-aware splits (one session from the target person in training) performed no better than fully unseen LSPO splits, one-session personalization is not a rescue for current models.
  • Among the reproduced methods, the three built on handcrafted features retained small positive R² under LSPO while raw-signal deep networks did not, suggesting feature representations deserve continued investigation.
  • Datasets skewed toward the normal glucose range make the mean-prediction trap worse; balancing the glucose distribution, augmentation, or upweighting non-mean samples are concrete next steps the authors propose.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If this pattern holds, the same critique likely applies to other PPG-based continuous health estimates, such as blood pressure or hydration, that use window-level random splits and safety-zone-style metrics; those systems should be tested with the same participant-independent splits and baseline comparisons.
  • The over-90% zone A+B baseline suggests that clinical acceptance criteria built on error-grid zones may need an explicit no-signal baseline benchmark: a device must beat a constant mean predictor on unseen participants before claiming clinical acceptability.
  • A testable extension is to apply the same pipeline to multi-wavelength, in-ear, or smartphone-camera PPG datasets; if any modality keeps positive R² under LSPO, the bottleneck is single-channel fingertip PPG rather than glucose estimation itself.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper proposes a reproducible evaluation pipeline for PPG-based blood glucose level (BGL) estimation and uses it to reassess five published methods (Zeynali et al., Alghlayini et al., Prabha et al., Islam et al., Gupta et al.) on two public datasets, VitalDB and MUST. Models are evaluated under three split protocols—random window-level, participant-aware, and leave-some-participants-out (LSPO)—plus an out-of-distribution (OOD) transfer to MUST. The central empirical finding is that random splits yield optimistic MAE/R2 values, while participant-aware and LSPO splits cause nearly all models to collapse to near-zero or negative R2, comparable to a mean-prediction baseline. At the same time, over 90% of predictions from every model and split, including the baseline, fall into Clarke Error Grid zones A+B. The authors conclude that random-split evaluation substantially overestimates generalization and that CEG-based clinical metrics can conceal model failure.

Significance. If the findings hold, they make an important methodological contribution to PPG-based BGL research: the paper provides a public, extensible benchmarking pipeline; it compares models against a mean-prediction baseline; it applies three clearly defined split protocols; and it demonstrates a concrete blind spot in CEG-based clinical validation. The baseline comparison and the splitting experiments are well designed and the core observation—that random splits overestimate generalization and CEG zones remain high even for a mean predictor—is supported by the reported numbers. The main caveat is that the five methods are re-implemented from text descriptions because none of the original code was released, so the precise attribution of the collapse to the original methods is less certain than the abstract implies. Nevertheless, the methodological conclusions about leakage and metric blind spots are robust to this caveat.

major comments (2)
  1. [§5.1, §7.1, Table 2] The load-bearing attribution of the performance collapse to the original published methods is weakened by reimplementation uncertainty. The paper states that none of the five original works published code and that 'variation in results can occur due to differences in model training and tuning across datasets and gaps in the descriptions' (§7.1). For Zeynali et al., the only method whose original dataset is available, the reproduction uses a different split (80-20 rather than the reported 70-15-15) and yields materially different metrics (MAE 13.97, R2 0.60 vs. reported MAE 18.54, R2 0.40). As a result, the abstract's claim that 'five representative PPG-based BGL methods' collapse is not fully supported; what is demonstrated is that the authors' re-implementations collapse. I recommend adding a reproduction sanity-check table that compares each re-implementation with the original reported
  2. [§5, Participant-aware split; Table 2] The participant-aware split is restricted to only 76 participants with multiple recordings, and the description does not specify how the two retained recordings were selected when a participant had more than two. The fold-to-fold standard deviations are large (e.g., Zeynali R2 = -0.08 ± 0.22, MAE = 36.18 ± 4.51), reflecting the small, specially selected test set. The paper acknowledges this noise, but the prose ('prior exposure to the target individual yields no reliable improvement') is stronger than the evidence from this protocol alone. The conclusion is also supported by the LSPO and OOD results, so I do not see this as fatal; however, the participant-aware analysis should either report exact test-set sizes, glucose ranges, and confidence intervals, or be explicitly framed as a pilot/personalization probe rather than a definitive quantitative comparison.
minor comments (6)
  1. [§5, Participant-aware split] The sentence 'exactly two recordings per participant were retained' is ambiguous: does this apply to all participants or only to those assigned to the test set? Please clarify the selection procedure and how many recordings were used for participants with more than two sessions.
  2. [§4, Data] VitalDB is described as having 4,872 cases with 4,626 unique participants. Please clarify whether 'cases' are surgical cases and whether multiple cases per participant are allowed; this is relevant to the leakage analysis.
  3. [§6, Table 2] The table lists standard deviations for all splits, but the OOD row has no standard deviation. Since OOD evaluation uses a single model trained on all VitalDB data, please state explicitly that no cross-validation was used and whether hyperparameters were tuned on VitalDB or MUST.
  4. [§6, Figures 1 and 2] Only two of the five models are illustrated with CEG plots. Adding a supplementary figure with all five models, or explicitly stating that the two are representative, would help the reader assess the generality of the 'horizontal band' pattern.
  5. [Abstract and Introduction] The phrase 'first reproducible, extensible evaluation pipeline' is strong. The paper should temper it to 'first, to our knowledge' or add a brief search statement to justify the novelty claim, since the related work section does not systematically review existing benchmarks.
  6. [References] Reference [2] is a GitHub repository. For reproducibility, please include a version tag, commit hash, or DOI, and state the license of the pipeline code.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the evaluation is an external empirical check with genuine baseline controls; the only self-citation is a code repository and is not load-bearing.

full rationale

The paper's central claims are empirical rather than definitional. Five published methods are re-implemented from textual descriptions (Section 5.1) and evaluated on two public datasets under three externally defined split protocols (random, participant-aware, LSPO). The conclusion that random splits overestimate generalization is supported by comparing each model's test-set R^2 and MAE to a mean-prediction baseline; the baseline is a genuine control, and the strict splits are designed to remove the leakage the paper criticizes. The conclusion that Clarke Error Grid zones conceal failure is demonstrated by the baseline itself falling >90% in zones A+B, which is an observed fact about the metric, not a consequence of model construction. The paper explicitly discloses its main uncertainty in Section 7.1: the models were reproduced from descriptions and 'variation in results can occur due to differences in model training and tuning across datasets and gaps in the descriptions.' This is a fidelity/validity limitation, not a circular step: the reproduction could in principle have matched the original papers' reported results, and the negative finding was not guaranteed by the pipeline definitions. The only self-citation is reference [2], the authors' own GitHub pipeline; it is cited as the location of the reproducible code and does not supply any of the paper's evidential premises. Thus no load-bearing argument reduces to its own inputs, and the derivation chain is self-contained against external benchmarks and controls.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

No global parameters are fitted by this study. The reproduced models inherit hyperparameters from the original papers (e.g., ResNet-34 depth, XGBoost settings), but the study's conclusion does not depend on tuning these internally. The split ratios and the 'two recordings per participant' rule in the participant-aware protocol are design choices, not fitted parameters. The paper introduces no new physical entities; the only new artifact is a software pipeline (reference [2]).

assumptions (5)
  • domain assumption The five re-implementations faithfully capture the corresponding original methods.
    Models were built from descriptions because original code is unavailable (Section 5.1). The paper acknowledges reproduction uncertainty in Section 7.1, but the central negative conclusion about the original literature depends on this fidelity.
  • domain assumption Random window-level splitting with overlapping windows is representative of the evaluation protocols used in prior PPG-based BGL work.
    The paper asserts that many prior studies split at the segment level, allowing overlapping windows in train and test (Sections 2 and 5). If prior work did not actually overlap, the leakage critique would be less applicable.
  • domain assumption Participant-aware and LSPO splits are the appropriate tests of real-world generalization for this application.
    The paper selects these protocols, including a participant-aware design that retains only two recordings per participant from 76 participants (Section 5). Alternatively, longitudinal or session-level splits could be argued to be more realistic for a personalized wearable.
  • domain assumption VitalDB and MUST are adequate datasets for assessing PPG-based BGL feasibility.
    VitalDB is perioperative (under anesthesia), which the paper flags as a limitation (Section 7.1); MUST is small (23 participants) and has a narrow glucose range. The OOD conclusion relies on the two distributions being similar in mean and spread (Section 6).
  • standard math Standard definitions of R2, MARD, and Clarke Error Grid zones apply.
    The metrics are defined per cited references (Shalizi [37], Heinemann et al. [17], Clarke [9]) and used consistently; no new metric definitions are introduced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reassessing the Feasibility of PPG-Based Non-Invasive Blood Glucose Level Estimation." pith.science (2026). https://pith.science/paper/JCQO3KH3

@misc{pith2026260801820,
  author       = {Pith},
  title        = {Pith review of: Reassessing the Feasibility of PPG-Based Non-Invasive Blood Glucose Level Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JCQO3KH3}},
  note         = {Machine review of arXiv:2608.01820}
}
abstract

Non-invasive blood glucose level (BGL) estimation from photoplethysmography (PPG) holds great promise for wearable health monitoring, but results across studies are hard to compare due to inconsistent datasets, data leakage, and non-standardized evaluation metrics. We present the first reproducible, extensible evaluation pipeline and use it to reassess five representative PPG-based BGL methods on published datasets under three increasingly strict data-split protocols: random window-level, participant-aware, and leave-some-participants-out (LSPO). Models appeared competitive under random splitting but collapsed under participant-aware and LSPO evaluation, with nearly all yielding near-zero or negative R$^2$ values comparable to a mean-prediction baseline. Critically, across every model and split, over 90% of predictions fell within clinically acceptable zones (Clarke Error Grid A+B), including the baseline. This reveals a fundamental disconnect: clinical zone metrics systematically conceal model failure in this domain. Our findings demonstrate that random train-test splits substantially overestimate the generalization of PPG-based BGL models due to sample-level data leakage, and that robust ML evaluation must precede clinical validation to meaningfully assess real-world utility.

Figures

Figures reproduced from arXiv: 2608.01820 by the authors.

Figure 1
Figure 1. CEG analysis for Prabha et al. [34], pooled across all five folds; color indicates prediction density. Predictions concentrate around the training mean under all three splits, consistent with the near-zero R2 in [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 1 canonical work pages

  1. [1]

    Assessment of Quality of Life and Self-care Behaviors in Patients with Type 2 Diabetes Mellitus in Mashhad, Iran.Jundishapur Journal of Chronic Disease Care(2020)

    2020. Assessment of Quality of Life and Self-care Behaviors in Patients with Type 2 Diabetes Mellitus in Mashhad, Iran.Jundishapur Journal of Chronic Disease Care(2020). doi:10.5812/jjcdc.105910

    work page doi:10.5812/jjcdc.105910 2020

  2. [2]

    Blood Glucose Monitoring From PPG: Reproducible Benchmark Pipeline

    2026. Blood Glucose Monitoring From PPG: Reproducible Benchmark Pipeline. https://github.com/R-Supraja/bgl-evaluation-pipeline.git

    2026

  3. [3]

    Gökhan Adigüzel, Ümit Şentürk, and Kemal Polat. 2024. Blood glucose level estimation using photoplethysmography (PPG) signals with explainable artificial intelligence techniques.Open Journal of Nano9, 1 (2024), 45–62

    2024

  4. [4]

    Saifeddin Alghlayini, Mohammed Azmi Al-Betar, Mohamed Atef, and Ghazi Al-Naymat. 2024. Bayesian Optimization-Based CNN Model for Blood Glucose Estimation Using Photoplethysmography Signals. InThe International Conference on Innovations in Computing Research. Springer, 142–152

    2024

  5. [5]

    Saifeddin Alghlayini, Asmaa Hosni, and Mohamed Atef. 2023. Photoplethysmography based blood glucose estimation using convolutional neural networks. In2023 Advances in Science and Engineering Technology International Conferences (ASET). IEEE, 1–5

    2023

  6. [6]

    Nurul A Mohd Asarani, Andrew N Reynolds, Sara E Boucher, Martin de Bock, and Benjamin J Wheeler. 2020. Cutaneous complications with continuous or flash glucose monitoring use: systematic review of trials and observational studies.Journal of diabetes science and technology14, 2 (2020), 328–337

    2020

  7. [7]

    Luis Alfredo Castro-Pimentel, Adriana del Carmen Téllez-Anguiano, Oscar Iván Coronado-Reyes, and Jorge Luis Diaz-Huerta. 2023. Three- wavelength PPG and support vector machine for non-invasive estimation of blood glucose.Optical and Quantum Electronics55, 8 (2023), 708

    2023

  8. [8]

    Justin Chu, Wen-Tse Yang, Wei-Ru Lu, Yao-Ting Chang, Tung-Han Hsieh, and Fu-Liang Yang. 2021. 90% accuracy for photoplethysmography-based non-invasive blood glucose prediction by deep learning with cohort arrangement and quarterly measured hba1c.Sensors21, 23 (2021), 7815

    2021

Show all 46 references
  1. [9]

    William L Clarke. 2005. The original Clarke error grid analysis (EGA).Diabetes technology & therapeutics7, 5 (2005), 776–779

    2005

  2. [10]

    Sarah D’Haese, Maxim Verboven, Lize Evens, Dorien Deluyker, Ivo Lambrichts, BO Eijnde, Dominique Hansen, and Virginie Bito. 2023. Moderate-and high-intensity endurance training alleviate diabetes-induced cardiac dysfunction in rats.Nutrients15, 18 (2023), 3950

    2023

  3. [11]

    Md Asaf-uddowla Golap, SM Taslim Uddin Raju, Md Rezwanul Haque, and MMA Hashem. 2021. Hemoglobin and glucose level estimation from PPG characteristics features of fingertip video using MGGP-based model.Biomedical Signal Processing and Control67 (2021), 102478

    2021

  4. [12]

    Shantanu Sen Gupta, Shifat Hossain, Chowdhury Azimul Haque, and Ki-Doo Kim. 2020. In-vivo estimation of glucose level using PPG signal. In 2020 International Conference on Information and Communication Technology Convergence (ICTC). IEEE, 733–736

    2020

  5. [13]

    Shantanu Sen Gupta, Tae-Ho Kwon, Shifat Hossain, and Ki-Doo Kim. 2021. Towards non-invasive blood glucose measurement using machine learning: An all-purpose PPG system design.Biomedical Signal Processing and Control68 (2021), 102706

    2021

  6. [14]

    Shraddha Habbu, Manisha Dale, and Rajesh Ghongade. 2019. Estimation of blood glucose by non-invasive method using photoplethysmography. S¯adhan¯a44, 6 (2019), 135

    2019

  7. [15]

    Ghena Hammour and Danilo P Mandic. 2023. An in-ear PPG-based blood glucose monitor: A proof-of-concept study.Sensors23, 6 (2023), 3319

    2023

  8. [16]

    Vera Hartmann, Haipeng Liu, Fei Chen, Qian Qiu, Stephen Hughes, and Dingchang Zheng. 2019. Quantitative comparison of photoplethysmographic waveform characteristics: Effect of measurement site.Frontiers in physiology10 (2019), 198. Reassessing the Feasibility of PPG-Based Non-...

    2019

  9. [17]

    Lutz Heinemann, Michael Schoemaker, Günther Schmelzeisen-Redecker, Rolf Hinzmann, Adham Kassab, Guido Freckmann, Florian Reiterer, and Luigi Del Re. 2020. Benefits and limitations of MARD as a performance parameter for continuous glucose monitoring in the interstitial space. J...

    2020

  10. [18]

    Shamima Hossain, Bidya Debnath, Sabyasachi Biswas, Md Junaed Al-Hossain, Adrita Anika, and Syed Khaled Zaman Navid. 2019. Estimation of blood glucose from PPG signal using convolutional neural network. In2019 IEEE international conference on biomedical engineering, computer an...

    2019

  11. [19]

    Tanvir Tazul Islam, Md Sajid Ahmed, Md Hassanuzzaman, Syed Athar Bin Amir, and Tanzilur Rahman. 2021. Blood glucose level regression for smartphone ppg signals using machine learning.Applied Sciences11, 2 (2021), 618

    2021

  12. [20]

    Yulong Jia, Yu Zheng, Zhong Wang, and Yuee Li. 2025. Non-invasive blood glucose based on dual wavelength transmission PPG signal.IEEE Sensors Journal(2025)

    2025

  13. [21]

    Shi-En Jian, Yu-Lung Lo, Yun-Tzu Chuang, and Shu-Han Kuo. 2025. Using Machine learning to predict blood glucose level based on Photoplethys- mography.Measurement253 (2025), 117421

    2025

  14. [22]

    Hui Jiang, Tianliang Yao, and Cheng Ding. 2025. PPG-based glucose sensors: a review.Artificial Intelligence Review58, 12 (2025), 1–33

    2025

  15. [23]

    Sayash Kapoor and Arvind Narayanan. 2023. Leakage and the reproducibility crisis in machine-learning-based science.Patterns4, 9 (2023)

    2023

  16. [24]

    Abdolamir Karbalaie, Farhad Abtahi, and Charlotte K Häger. 2026. Participant-Aware Model Validation for Repeated-Measures Data: Comparative Cross-Validation Study.JMIR AI5 (2026), e87728

    2026

  17. [25]

    Ali Kermani and Hossein Esmaeili. 2023. The dataset of photoplethysmography signals collected from a pulse sensor to measure blood glucose level. Mendeley Data3 (2023), 10–17632

    2023

  18. [26]

    Khadija Khan, Laraib Malik, Abdul Qadeer Khan, Saadullah Farooq Abbasi, and Theodoros N Arvanitis. 2026. Non invasive blood glucose estimation using green light photoplethysmography and machine learning.Frontiers in digital health8 (2026), 1705086

    2026

  19. [27]

    Maurício Krause and Giuseppe De Vito. 2023. Type 1 and type 2 diabetes mellitus: commonalities, differences and the importance of exercise and nutrition. 4279 pages

    2023

  20. [28]

    Hyung-Chul Lee, Yoonsang Park, Soo Bin Yoon, Seong Mi Yang, Dongnyeok Park, and Chul-Woo Jung. 2022. VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients.Scientific Data9, 1 (2022), 279

    2022

  21. [29]

    Po-Lei Lee, Kuo-Wei Wang, and Chen-Yuan Hsiao. 2023. A noninvasive blood glucose estimation system using dual-channel PPGs and pulse-arrival velocity.IEEE Sensors Journal23, 19 (2023), 23570–23582

    2023

  22. [30]

    Enric Monte-Moreno. 2011. Non-invasive estimate of blood glucose and blood pressure from a photoplethysmograph by means of machine learning techniques.Artificial intelligence in medicine53, 2 (2011), 127–138

    2011

  23. [31]

    Franklin R Muntis, Abbie E Smith-Ryan, Jamie Crandell, Kelly R Evenson, David M Maahs, Michael Seid, Saame R Shaikh, and Elizabeth J Mayer- Davis. 2023. A high protein diet is associated with improved glycemic control following exercise among adolescents with type 1 diabetes.N...

    2023

  24. [32]

    Woon May Ong, Siew Siang Chua, and Chirk Jenn Ng. 2014. Barriers and facilitators to self-monitoring of blood glucose in people with type 2 diabetes using insulin: a qualitative study.Patient preference and adherence(2014), 237–246

    2014

  25. [33]

    SN Parsons, SD Luzio, JN Harvey, SC Bain, WY Cheung, Alan Watkins, and DR Owens. 2019. Effect of structured self-monitoring of blood glucose, with and without additional TeleCare support, on overall glycaemic control in non-insulin treated Type 2 diabetes: the SMBG Study, a 12...

    2019

  26. [34]

    Anju Prabha, Jyoti Yadav, Asha Rani, and Vijander Singh. 2022. Intelligent estimation of blood glucose level using wristband PPG signal and physiological parameters.Biomedical Signal Processing and Control78 (2022), 103876

    2022

  27. [35]

    Florian Reiterer, Philipp Polterauer, Michael Schoemaker, Guenther Schmelzeisen-Redecker, Guido Freckmann, Lutz Heinemann, and Luigi Del Re

  28. [36]

    Christian Salamea, Erick Narvaez, and Melisa Montalvo. 2019. Database proposal for correlation of glucose and photoplethysmography signals. In The International Conference on Advances in Emerging Trends and Technologies. Springer, 44–53

    2019

  29. [37]

    2013.Advanced data analysis from an elementary point of view

    Cosma Shalizi. 2013.Advanced data analysis from an elementary point of view. Citeseer

    2013

  30. [38]

    Trisha Shang, Jennifer Y Zhang, Andreas Thomas, Mark A Arnold, Beatrice N Vetter, Lutz Heinemann, and David C Klonoff. 2022. Products for monitoring glucose levels in the human body with noninvasive optical, noninvasive fluid sampling, or minimally invasive technologies.Journa...

    2022

  31. [39]

    Gašper Slapničar, Nejc Mlakar, and Mitja Luštrek. 2019. Blood pressure estimation from photoplethysmogram using a spectro-temporal deep neural network.Sensors19, 15 (2019), 3420

    2019

  32. [40]

    Liu Tang, Shwu Jen Chang, Ching-Jung Chen, and Jen-Tsai Liu. 2020. Non-invasive blood glucose monitoring technology: a review.Sensors20, 23 (2020), 6925

    2020

  33. [41]

    Mutian Wang, Xuelei Liu, Wenyi Han, Xinyu Lin, Xin Chen, Shun Zhao, Zhiqiang Zhuang, Leian Zhang, and Peiqiang Su. 2025. Design of a wearable finger PPG-based blood glucose monitor.Annals of Biomedical Engineering53, 10 (2025), 2580–2593

    2025

  34. [42]

    WHO. 2019. CLASSIFICATION OF DIABETES MELLITUS. (2019). https://iris.who.int/server/api/core/bitstreams/2cb3ab68-a52a-402e-ad47- 8bc5a4edc834/content

    2019

  35. [43]

    Luke Young, Danling Wang, and Na Gong. 2025. Low-cost wearable edge-ai device for diabetes management. InProceedings of the Great Lakes Symposium on VLSI 2025. 238–244. 12 Ramesh et al

    2025

  36. [44]

    Mahdi Zeynali, Khalil Alipour, Bahram Tarvirdizadeh, and Mohammad Ghamari. 2025. Non-invasive blood glucose monitoring using PPG signals with various deep learning models and implementation using TinyML.Scientific Reports15, 1 (2025), 581

    2025

  37. [45]

    Gaobo Zhang, Zhen Mei, Yuan Zhang, Xuesheng Ma, Benny Lo, Dongyi Chen, and Yuanting Zhang. 2020. A noninvasive blood glucose monitoring system based on smartphone PPG signal processing and machine learning.IEEE Transactions on Industrial Informatics16, 11 (2020), 7209–7218

    2020

  38. [2017]

    Significance and reliability of MARD for the accuracy of CGM systems.Journal of diabetes science and technology11, 1 (2017), 59–67

    2017

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.