Noise-Driven AI Sensors: Secure Healthcare Monitoring with PUFs

Abhijit Sarkar; A. Lynn Abbott; Christiana Chamon

REVIEW 4 major objections 4 minor 15 references

Noise-Driven AI Sensors: Secure Healthcare Monitoring with PUFs

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a wearable ECG monitor can harvest the same hardware noise that normally degrades signals to simultaneously improve a lightweight classifier's anomaly-detection accuracy to 92% (from 85%) and generate…

desk verdict A simulation-only pitch that packages familiar blocks; the 'same noise' premise conflates dynamic and static randomness, and the headline numbers have no independent check. read the letter →

arxiv 2506.05135 v1 pith:TFY2KDST submitted 2025-06-05 eess.SP

Christiana Chamon , Abhijit Sarkar , A. Lynn Abbott This is my paper

classification eess.SP

keywords hardwarenoisephysicalunclonablefunctionsECGanomalydetectionrandomforestclassifiernoise-drivensignalprocessinglow-powerwearablesensorsatrialfibrillationprematureventricularcontractions

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

This paper tries to establish that noise in a healthcare sensor is not a problem to filter away but a resource to exploit twice. It proposes an ECG wearable platform in which Gaussian sensor noise sampled during analog-to-digital conversion is added to the feature set of a random forest classifier, raising PVC and AF detection accuracy from 85% to 92%, while the same underlying noise notionally drives a ring-oscillator PUF that produces device-unique 64-bit keys with 98% inter-device uniqueness. The entire simulated system is said to run on 50 µW, making the combination attractive for battery-powered wearables and implants. A sympathetic reader would care because it promises a single low-power mechanism that simultaneously addresses three known pain points: noise, security, and energy.

What carries the argument

The central object is the dual-use noise path. Noise is sampled during analog-to-digital conversion with a dedicated low-power circuit rather than filtered, and the same noise is asserted to drive both the ML feature augmentation and the ring-oscillator PUF that produces a 64-bit signature, stabilized by SHA-256 hashing and BCH error correction. The ML side rests on a random forest classifier with 100 trees and 15 features—RR-interval statistics, QRS-width statistics, and 12 wavelet coefficients from a 5-level Daubechies transform—tuned by grid search. The power budget (50 µW) and the claimed accuracy and security figures depend on this shared-noise architecture.

What would settle it

Take a prototype of the 65 nm CMOS sensor described in the paper's future work and measure, on the same die, the ADC's real noise statistics, the random forest accuracy when trained and run with that real noise, and the PUF's inter-device Hamming distance and intra-device bit error rate when the ML and PUF paths draw on the same noise source. If real noise deviates from white Gaussian in a way that drops classification below the 92% claim, or if bit error rate rises above 0.5% under simultaneous use, the central dual-purpose-noise claim fails.

Watch

Extended reading notes

Core claim

The core claim is that hardware noise can be a dual-purpose resource: it is treated as an information-bearing stochastic feature for ML rather than filtered out, and it is the entropy source for PUF-based authentication. In simulation, a 250 Hz synthetic ECG stream (10,000 samples, 10% PVC/AF anomalies) is augmented with white Gaussian noise (0.1–1 mV, mean 0), and a 15-feature random forest (RR intervals, QRS-width statistics, and 12 Daubechies wavelet coefficients) reaches 92% accuracy and 8% higher F1 than traditional filtering. A ring-oscillator PUF evaluated by Monte Carlo over 1,000 virtual devices yields 98% uniqueness, <0.5% bit error rate after BCH correction, and >99% stability under ±10°C and ±5% voltage variation, while power is budgeted at 5 µW for the PUF, 30 µW for ML inference, and 15 µW for noise processing, totaling 50 µW. The paper argues this unifies functions that are normally separate, reducing complexity and power for wearable and implantable ECG monitoring.

Load-bearing premise

The design stands on the assumption that the white Gaussian noise added to synthetic ECG signals in simulation behaves exactly like real hardware noise, and that the same physical noise can simultaneously feed the ML features and the PUF key generator without corrupting the ECG signal or the key.

Editorial extensions

If this is right

If the noise-augmented classifier really gains 8% over filtered training, sensor noise can be repurposed as a data-augmentation source for lightweight models, removing the need for expensive noise-filtering front ends.
If the same noise source feeds the PUF, device authentication keys can be generated without a separate randomness source, saving silicon area and power.
If the 50 µW budget holds, continuous ECG anomaly detection with security becomes plausible on coin-cell-powered wearables and implants.
The reported <10 ms inference time and <1 ms key generation would support real-time alerting and per-packet authentication in telemedicine.
Because the classifier is a 100-tree random forest running on a 32-bit microcontroller, the approach is positioned to scale beyond ECG to other physiological signals without deep-learning power costs.

Reading between the lines

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

Editorial inference: the reported 8% gain is most naturally read as a data-augmentation effect—more diverse training samples make the random forest generalize better—rather than proof that hardware noise itself carries diagnostic information.
Editorial inference: the architecture's viability in silicon hinges on an explicit isolation mechanism preventing the noise tapped for ML features from corrupting the ECG path and preventing PUF readout from draining the shared node; the paper does not specify that mechanism.
Editorial inference: a direct next experiment is to train the same classifier on recorded ADC noise from a real prototype instead of synthetic Gaussian noise; if accuracy degrades, the dual-use claim is simulation-specific.
Editorial inference: the same recipe could be carried to other physiological sensors, such as PPG or glucose monitors, where noise is also treated as a nuisance rather than a feature.

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.

Desk Editor's Note

read the letter

First read: the one genuinely new framing--using a single hardware noise source for both ML augmentation and PUF keys--is attractive as a pitch, but the paper never shows a mechanism that makes it work. The rest is a collection of familiar components: a noise-augmented random forest, a ring-oscillator PUF, and an ECG classifier, and all results come from the authors' own simulation, with no code, data, or hardware check.

What earns credit: the authors identify a real problem--low-power secure wearables are a real gap, and the block-level architecture is clearly laid out. They are honest that this is a simulation study with hardware as future work, and the 50 µW power budget is transparently itemized (5+30+15 µW). The ML comparison is simple and the accuracy gain (92% vs 85%) is presented without overclaiming.

The soft spot is not soft. The claim that 'the same noise drives a PUF module' conflates two distinct randomness sources. A PUF key must be reproducible across power cycles and environmental corners; that reproducibility comes from static manufacturing variation such as ring-oscillator frequency mismatch. ML noise augmentation, as the paper describes it, uses dynamic Gaussian noise, which is by design non-reproducible. If the same dynamic noise were literally used, the intra-device Hamming distance would sit near 50%, and no stable key would be possible. If instead the PUF uses frequency mismatch, then the two uses are separate and the novelty collapses into two independent, known blocks. The paper gives no circuit-level explanation of how one physical noise source serves both purposes without corrupting the signal or key. That is not a missing detail; it is the load-bearing assumption that makes the integrated 50 µW budget meaningful.

The rest of the simulation is likewise unsupported. The 98% uniqueness is a Monte Carlo result under unstated variation assumptions; the 50 µW budget is a sum of three estimated numbers; side-channel resistance is asserted without any attack model. Citation hygiene is poor too--a PUF paper is cited for market projections, a market report for PUF physics.

Who is this for? A speculative idea at a workshop, perhaps. As a citable contribution, it doesn't hold up. The central flaw is fundamental, and I don't see how a referee could approve it without a complete rework of the noise-sharing premise. Recommendation: desk reject. If the authors come back with a concrete circuit architecture that genuinely shares a noise source, or a revised claim that the two uses are separate, it might become a decent workshop paper. As it stands, not ready for peer review.

Referee Report

4 major / 4 minor

Summary. The paper proposes an integrated wearable ECG monitoring platform in which hardware noise is used as a dual-purpose resource: it is sampled during analog-to-digital conversion to augment machine-learning features for arrhythmia detection, and it is also claimed to drive a ring-oscillator physical unclonable function (PUF) for secure authentication. The platform is evaluated entirely in simulation: a synthetic ECG dataset with 10% anomalous beats, a random-forest classifier, a Monte Carlo model of 1,000 virtual PUF devices, and component-level power estimates. The headline claims are 92% anomaly-detection accuracy versus 85% with traditional filtering, 98% PUF uniqueness, and a 50 µW total power budget. The manuscript states that hardware implementation in 65 nm CMOS is future work (Section V-B).

Significance. If substantiated, the platform would be a meaningful step toward low-power secure wearable monitoring, and the explicit integration of noise augmentation with PUF-based security is an interesting idea. The paper gives clear architecture-level descriptions, a quantified power breakdown, and comparisons against CNN and AES-128 baselines (Table III). However, the central evidence is entirely self-simulation: there is no released code, no hardware measurement, no error bars, and no real ECG validation. Moreover, the dual-use noise premise is not physically developed. The significance therefore rests on a load-bearing assumption that the manuscript does not establish.

major comments (4)

[Section III-A] The statement 'The same noise drives a PUF module' conflates two incompatible randomness sources. A ring-oscillator PUF key must be reproducible across power cycles and environmental corners, which requires static manufacturing mismatch; the Gaussian noise used for ML augmentation is dynamic and non-reproducible by design. If the PUF entropy were literally the same dynamic noise, the raw intra-device Hamming distance would be near 50%, making BCH error correction insufficient for stable key generation. If instead the PUF relies on static ring-oscillator frequency mismatch, then the claimed dual-use of one physical noise source is unsupported, and no circuit-level coupling between the ADC noise path and the RO-PUF is described anywhere in the paper. This is load-bearing because the title, abstract, and architecture all depend on the dual-purpose noise claim.
[Section III-C and IV-A] The ML results are generated from a self-contained simulation on synthetic ECG 'inspired by the MIT-BIH Arrhythmia Database' rather than on the actual database, with no error bars, no cross-validation details, and no released code. The dataset is also heavily imbalanced (10% anomalies), so a 92% accuracy figure can be misleading without precision-recall and ROC analysis. The claimed 8% improvement from noise augmentation may therefore be an artifact of the specific synthetic noise level (0.1–1 mV) and random-forest hyperparameters, and it has not been shown to transfer to real wearable sensor data. This undermines the central claim that noise improves ML accuracy.
[Section IV-C] The 50 µW power budget is a sum of estimated component powers (5 µW PUF + 30 µW ML inference + 15 µW noise processing) with no measured data, no energy model for feature extraction, and no memory or throughput analysis to justify that a 100-tree random forest with 15 features and a 5-level wavelet transform can run on a 32-bit microcontroller at 250 Hz within 30 µW. Section V-B explicitly lists hardware implementation as future work, so the paper does not provide evidence that the central power claim is achievable.
[Section IV-B] The PUF security results are based on an undocumented Monte Carlo model: no distribution for ring-oscillator frequency variation, no temperature/voltage noise model beyond saying ±10°C and ±5%, no definition of the simulated side-channel attacks, and no attack results. A claimed '98% uniqueness' and '<0.5% BER after BCH' cannot be assessed without these assumptions. The statement that the PUF 'withstands simulated side-channel attacks' is presented as a result but no attack methodology or quantitative outcome is reported, so the security claim is not supported by the evidence in the manuscript.

minor comments (4)

[References] Multiple in-text citations do not match the reference list: the healthcare-market statement at the start of Section I cites [1] but reference [1] is a PUF paper, the ring-oscillator PUF design in Section II-B cites [9] but reference [9] is a market report, and the MIT-BIH database in Section III-A cites [13] rather than the MIT-BIH reference [7].
[Table II] Table II is labeled a confusion matrix but reports only per-class true-positive and false-negative rates; a full confusion matrix or at least explicit false-positive rates should be provided to support the claimed precision and specificity.
[Section IV-A and Figure 4] The figure axis labels say 'With Noise' and 'Without Noise' while the text compares 'noise-augmented features' against 'traditional filtering', which are not necessarily the same comparison; the figure and text should use consistent terminology.
[Section IV-D] The sentence 'This analysis highlights our system’s advantage in Table III' is duplicated, and Table III's CNN power range of 100–500 µW and AES-128 key-generation time of 5 ms are cited without a source in the reference list.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: all headline figures are simulation outputs, not fitted inputs; the dual-use noise premise is physically unsupported but not a reduction of output to input.

full rationale

The paper's headline metrics (92% accuracy, 98% PUF uniqueness, 50 uW) are computed from its own simulations rather than derived from the assumptions by construction. The 92% accuracy is the test-set performance of a random forest trained on synthetic ECG data, and the 85% baseline is a comparison pipeline; neither is a fitted parameter renamed as a prediction. The 98% uniqueness is a Monte Carlo average over 1,000 virtual devices, and the 50 uW budget is the sum of component estimates (5+30+15 uW). These are evaluations of an in-house simulator, not reductions of output to input. The paper does not rely on self-citations: the cited references on PUFs, fuzzy extractors, and ECG processing are external works, and no uniqueness theorem from the authors' prior work is invoked. The claim that 'the same noise drives a PUF module' (Section III-A) conflates dynamic sensor noise with static manufacturing variation and is a serious physical-realism gap, but it is an unsupported assumption rather than a circular derivation; it does not make the accuracy or uniqueness figures equivalent to their inputs by construction. Consequently, no circular step can be quoted with the required specificity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claims rest on a small set of hand-chosen simulation settings and unverified hardware assumptions; no independent benchmarks are used.

free parameters (4)

Synthetic ECG noise level = 0.1-1 mV Gaussian
Chosen by hand (Section III-A) to mimic sensor noise; directly shapes the ML accuracy comparison.
Random forest hyperparameters = 100 trees, depth via grid search
Tuned on a 15% validation split (Section III-B); reported 92% test accuracy depends on these choices.
Power estimates = PUF 5 µW, ML 30 µW, noise processing 15 µW
Stated per-component estimates (Section IV-C); their sum defines the 50 µW headline budget, which is not measured.
PUF device count and error correction = 1,000 virtual devices, BCH, 64-bit
Simulation setup (Section III-C); uniqueness/reliability numbers are outputs of this assumed model.

assumptions (3)

domain assumption Gaussian white noise 0.1-1 mV models real hardware sensor noise
Section III-A uses this to simulate sensor imperfections; the claim that hardware noise can be harvested rests on this equivalence.
ad hoc to paper The same hardware noise can simultaneously drive ADC-based ML features and a ring oscillator PUF without degrading either
Section III-A states 'The same noise drives a PUF module' with no circuit-level justification; this is a core architectural premise.
domain assumption PUF uniqueness measured as average inter-device Hamming distance with 98% is sufficient for security
Section IV-B reports 98% uniqueness as a security metric without an attack model or entropy analysis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Noise-Driven AI Sensors: Secure Healthcare Monitoring with PUFs." pith.science (2026). https://pith.science/paper/TFY2KDST

@misc{pith2026250605135,
  author       = {Pith},
  title        = {Pith review of: Noise-Driven AI Sensors: Secure Healthcare Monitoring with PUFs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TFY2KDST}},
  note         = {Machine review of arXiv:2506.05135}
}

read the original abstract

Wearable and implantable healthcare sensors are pivotal for real-time patient monitoring but face critical challenges in power efficiency, data security, and signal noise. This paper introduces a novel platform that leverages hardware noise as a dual-purpose resource to enhance machine learning (ML) robustness and secure data via Physical Unclonable Functions (PUFs). By integrating noise-driven signal processing, PUFbased authentication, and ML-based anomaly detection, our system achieves secure, low-power monitoring for devices like ECG wearables. Simulations demonstrate that noise improves ML accuracy by 8% (92% for detecting premature ventricular contractions (PVCs) and atrial fibrillation (AF)), while PUFs provide 98% uniqueness for tamper-resistant security, all within a 50 uW power budget. This unified approach not only addresses power, security, and noise challenges but also enables scalable, intelligent sensing for telemedicine and IoT applications.

Figures

Figures reproduced from arXiv: 2506.05135 by the authors.

**Figure 2.** ECG signals with normal, anomalous (PVC), and noisy versions, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** Comparison of ML accuracy with noise-augmented features versus [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 3.** Figure 3: Power spectral density (PSD) of normal and noisy ECG signals, [PITH_FULL_IMAGE:figures/full_fig_p003_3.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 14 canonical work pages

[1]

Silicon physical random functions,

B. Gassend, D. Clarke, M. van Dijk, and S. Devadas, “Silicon physical random functions,” inProc. 9th ACM Conf. Comput. Commun. Secur., Washington, DC, USA, Nov. 2002, pp. 148–160

work page 2002
[2]

Fuzzy extractors: How to generate strong keys from biometrics and other noisy data,

Y . Dodis, R. Ostrovsky, L. Reyzin, and A. Smith, “Fuzzy extractors: How to generate strong keys from biometrics and other noisy data,” SIAM J. Comput., vol. 38, no. 1, pp. 97–139, 2008

work page 2008
[3]

A survey on lightweight entity authentication with strong PUFs,

J. Delvaux, D. Gu, D. Schellekens, and I. Verbauwhede, “A survey on lightweight entity authentication with strong PUFs,”ACM Com- put. Surveys, vol. 48, no. 2, article 26, 42 pages, Nov. 2015. doi: 10.1145/2818186

work page doi:10.1145/2818186 2015
[4]

Maes,Physically Unclonable Functions: Constructions, Properties and Applications

R. Maes,Physically Unclonable Functions: Constructions, Properties and Applications. Berlin, Germany: Springer, 2013

work page 2013
[5]

Cardiologist-level arrhythmia detec- tion and classification in ambulatory electrocardiograms using a deep neural network,

A. Y . Hannun, P. Rajpurkar, M. Haghpanahi, G. H. Tison, C. Bourne, M. P. Turakhia, and A. Y . Ng, “Cardiologist-level arrhythmia detec- tion and classification in ambulatory electrocardiograms using a deep neural network,”Nat. Med., vol. 25, no. 1, pp. 65–69, Jan. 2019. doi: 10.1038/s41591-018-0268-3

doi:10.1038/s41591-018-0268-3 2019
[6]

A deep convolutional neural network model to classify normal, atrial fibrillation and noisy ECG signals,

U. R. Acharya, S. L. Oh, Y . Hagiwara, J. H. Tan, M. Adam, A. Gertych, and R. S. Tan, “A deep convolutional neural network model to classify normal, atrial fibrillation and noisy ECG signals,”Comput. Methods Programs Biomed., vol. 144, pp. 189–198, Jun. 2017

work page 2017
[7]

The impact of the MIT-BIH arrhythmia database,

G. B. Moody and R. G. Mark, “The impact of the MIT-BIH arrhythmia database,”IEEE Eng. Med. Biol. Mag., vol. 20, no. 3, pp. 45–50, May/Jun. 2001

work page 2001
[8]

Comprehensive survey of com- putational ECG analysis: Databases, methods and applications,

E. Merdjanovska and A. Rashkovska, “Comprehensive survey of com- putational ECG analysis: Databases, methods and applications,”Expert Syst. Appl., vol. 203, p. 117206, Oct. 2022

work page 2022

Show all 15 references

[9]

Healthcare IoT market by component, ap- plication, end user, and region—Global forecast to 2025,

MarketsandMarkets, “Healthcare IoT market by component, ap- plication, end user, and region—Global forecast to 2025,” Rep.,

2025
[10]

Deep learning-based cardiovascular image diagnosis: A promising challenge,

K. K. L. Wong, G. Fortino, and D. Abbott, “Deep learning-based cardiovascular image diagnosis: A promising challenge,”Future Gener. Comput. Syst., vol. 110, pp. 802–811, Sep. 2020

2020
[11]

Automatic Detection of Atrial Fibrillation from Single-Lead ECG Using Deep Learning of the Car- diac Cycle,

A. Dubatovka and J. M. Buhmann, “Automatic Detection of Atrial Fibrillation from Single-Lead ECG Using Deep Learning of the Car- diac Cycle,”BME Frontiers, vol. 2022, p. 9813062, Apr. 2022, doi: 10.34133/2022/9813062

2022 doi
[12]

Physical unclon- able functions and applications: A tutorial,

C. Herder, M.-D. Yu, F. Koushanfar, and S. Devadas, “Physical unclon- able functions and applications: A tutorial,”Proc. IEEE, vol. 102, no. 8, pp. 1126–1141, Aug. 2014

2014
[13]

Automated detection of atrial fibrillation using R-R intervals and multivariate-based classification,

A. Kennedy, D. Finlay, D. Guldenring, R. Bond, K. Moran, and J. McLaughlin, “Automated detection of atrial fibrillation using R-R intervals and multivariate-based classification,”J. Electrocardiol., vol. 49, no. 6, pp. 871–876, Nov./Dec. 2016

2016
[14]

Federated machine learning: Concept and applications,

Q. Yang, Y . Liu, T. Chen, and Y . Tong, “Federated machine learning: Concept and applications,”ACM Trans. Intell. Syst. Technol., vol. 10, no. 2, pp. 1–19, Feb. 2019

2019
[2020]

Available: https://www.marketsandmarkets.com/Market- Reports/iot-healthcare-market-160082804.html

[Online]. Available: https://www.marketsandmarkets.com/Market- Reports/iot-healthcare-market-160082804.html

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Silicon physical random functions,

B. Gassend, D. Clarke, M. van Dijk, and S. Devadas, “Silicon physical random functions,” inProc. 9th ACM Conf. Comput. Commun. Secur., Washington, DC, USA, Nov. 2002, pp. 148–160

work page 2002

[2] [2]

Fuzzy extractors: How to generate strong keys from biometrics and other noisy data,

Y . Dodis, R. Ostrovsky, L. Reyzin, and A. Smith, “Fuzzy extractors: How to generate strong keys from biometrics and other noisy data,” SIAM J. Comput., vol. 38, no. 1, pp. 97–139, 2008

work page 2008

[3] [3]

A survey on lightweight entity authentication with strong PUFs,

J. Delvaux, D. Gu, D. Schellekens, and I. Verbauwhede, “A survey on lightweight entity authentication with strong PUFs,”ACM Com- put. Surveys, vol. 48, no. 2, article 26, 42 pages, Nov. 2015. doi: 10.1145/2818186

work page doi:10.1145/2818186 2015

[4] [4]

Maes,Physically Unclonable Functions: Constructions, Properties and Applications

R. Maes,Physically Unclonable Functions: Constructions, Properties and Applications. Berlin, Germany: Springer, 2013

work page 2013

[5] [5]

Cardiologist-level arrhythmia detec- tion and classification in ambulatory electrocardiograms using a deep neural network,

A. Y . Hannun, P. Rajpurkar, M. Haghpanahi, G. H. Tison, C. Bourne, M. P. Turakhia, and A. Y . Ng, “Cardiologist-level arrhythmia detec- tion and classification in ambulatory electrocardiograms using a deep neural network,”Nat. Med., vol. 25, no. 1, pp. 65–69, Jan. 2019. doi: 10.1038/s41591-018-0268-3

doi:10.1038/s41591-018-0268-3 2019

[6] [6]

A deep convolutional neural network model to classify normal, atrial fibrillation and noisy ECG signals,

U. R. Acharya, S. L. Oh, Y . Hagiwara, J. H. Tan, M. Adam, A. Gertych, and R. S. Tan, “A deep convolutional neural network model to classify normal, atrial fibrillation and noisy ECG signals,”Comput. Methods Programs Biomed., vol. 144, pp. 189–198, Jun. 2017

work page 2017

[7] [7]

The impact of the MIT-BIH arrhythmia database,

G. B. Moody and R. G. Mark, “The impact of the MIT-BIH arrhythmia database,”IEEE Eng. Med. Biol. Mag., vol. 20, no. 3, pp. 45–50, May/Jun. 2001

work page 2001

[8] [8]

Comprehensive survey of com- putational ECG analysis: Databases, methods and applications,

E. Merdjanovska and A. Rashkovska, “Comprehensive survey of com- putational ECG analysis: Databases, methods and applications,”Expert Syst. Appl., vol. 203, p. 117206, Oct. 2022

work page 2022