pith. machine review for the scientific record. sign in

arxiv: 2604.18067 · v2 · submitted 2026-04-20 · 💻 cs.LG

Recognition: unknown

Towards Real-Time ECG and EMG Modeling on μNPUs

Josh Millar , Ashok Samraj Thangarajan , Soumyajit Chatterjee , Hamed Haddadi
Authors on Pith no claims yet

Pith reviewed 2026-05-10 05:52 UTC · model grok-4.3

classification 💻 cs.LG
keywords ECG analysisEMG signalslightweight modelsmicrocontroller NPUswavelet filter banksedge inferencephysiological signalsTransformer alternatives
0
0 comments X

The pith

PhysioLite matches Transformer accuracy on ECG and EMG signals while fitting in under 400KB for microcontroller NPUs.

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

The paper introduces PhysioLite, a compact neural network architecture for processing electrocardiogram and electromyogram signals on tiny hardware. It replaces the dynamic attention of large Transformers with learnable wavelet filter banks, offloaded encoding, and hardware-specific layers to reach comparable benchmark performance at roughly one-tenth the size after quantization. This setup is profiled for latency on two concrete micro-NPU chips, showing it can run locally. A sympathetic reader would care because it opens the door to private, real-time physiological analysis inside battery-powered wearables that cannot support cloud-scale models.

Core claim

PhysioLite is a lightweight, NPU-compatible model architecture and training framework for ECG/EMG signal analysis. Using learnable wavelet filter banks, CPU-offloaded positional encoding, and hardware-aware layer design, PhysioLite reaches performance comparable to state-of-the-art Transformer-based foundation models on ECG and EMG benchmarks, while being <10% of the size (~370KB with 8-bit quantization). The authors also profile its component-wise latency and resource consumption on the MAX78000 and HX6538 μNPUs to show its viability for signal analysis on constrained, battery-powered hardware.

What carries the argument

Learnable wavelet filter banks combined with CPU-offloaded positional encoding and hardware-aware layers that replace dynamic attention to enable efficient inference on μNPUs.

If this is right

  • Enables near-real-time, offline physiological signal analysis directly on wearable hardware.
  • Supports privacy-preserving inference by keeping all processing local without data offload.
  • Reduces model footprint to ~370KB after 8-bit quantization while matching benchmark scores.
  • Demonstrates concrete latency and resource profiles on MAX78000 and HX6538 μNPUs.

Where Pith is reading between the lines

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

  • The same wavelet-plus-hardware design might transfer to other time-series bio-signals such as EEG.
  • Hardware vendors could add native support for wavelet banks to further speed up similar models.
  • If the accuracy holds under patient-specific noise, it could support always-on monitoring features in consumer devices.
  • Releasing the models and training code allows independent checks on additional public datasets.

Load-bearing premise

The combination of learnable wavelet banks, CPU-offloaded encoding, and hardware-aware layers will preserve accuracy across real-world noisy signals and varying hardware without dynamic attention.

What would settle it

A test set of noisy real-world ECG or EMG recordings where PhysioLite's accuracy falls substantially below that of the compared Transformer models.

Figures

Figures reproduced from arXiv: 2604.18067 by Ashok Samraj Thangarajan, Hamed Haddadi, Josh Millar, Soumyajit Chatterjee.

Figure 1
Figure 1. Figure 1: Comparison of avg. F1 on ECG benchmarks (§5.1) with PhysioLite and a number of ECG foundation models. directly support dynamic attention or large intermediate activation buffers. Consequently, state-of-the-art physiological foundation models rely on server- or mobile-grade hardware, limiting their applicability for continuous, private, and real-time inference on wearable devices. We introduce PhysioLite, a… view at source ↗
Figure 2
Figure 2. Figure 2: , below, outlines the general design and layout of a µNPU, composed of a systolic array of processing elements (PEs) [46]. Each PE holds its own multiply-accumulate units and, importantly, its own dedicated weight SRAM; this avoids memory contention during network inference and exploits the inherent parallelism of CNN layers and dense matrix multiplications. The array of PEs is linked with a global buffer … view at source ↗
Figure 3
Figure 3. Figure 3: Performance of representative hardware accelera [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: PhysioLite’s architecture mapped to CPU and NPU execution. Multi-branch temporal convolutions pro￾vide wavelet-like feature extraction. 3.2 CPU-Based Positional Encoding Continuous learnable embeddings, relying on high-precision floating-point arithmetic, are incompatible with µNPU instruction sets. We instead offload positional encoding to the CPU: positional information is generated as deterministic sinu… view at source ↗
Figure 5
Figure 5. Figure 5: Performance comparison across ECG benchmark. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Latency breakdown on the MAX78000 for ECG input [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
read the original abstract

The miniaturisation of neural processing units (NPUs) and other low-power accelerators has enabled their integration into microcontroller-scale wearable hardware, supporting near-real-time, offline, and privacy-preserving inference. Yet physiological signal analysis has remained infeasible on such hardware; recent Transformer-based models show state-of-the-art performance but are prohibitively large for resource- and power-constrained hardware and incompatible with $\mu$NPUs due to their dynamic attention operations. We introduce PhysioLite, a lightweight, NPU-compatible model architecture and training framework for ECG/EMG signal analysis. Using learnable wavelet filter banks, CPU-offloaded positional encoding, and hardware-aware layer design, PhysioLite reaches performance comparable to state-of-the-art Transformer-based foundation models on ECG and EMG benchmarks, while being <10% of the size ($\sim$370KB with 8-bit quantization). We also profile its component-wise latency and resource consumption on both the MAX78000 and HX6538 WE2 $\mu$NPUs, demonstrating its viability for signal analysis on constrained, battery-powered hardware. We release our model(s) and training framework at: https://github.com/j0shmillar/physiolite.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper introduces PhysioLite, a lightweight model architecture and training framework for ECG and EMG signal analysis on microcontroller-scale NPUs. It combines learnable wavelet filter banks, CPU-offloaded positional encoding, and hardware-aware layer design to claim performance comparable to state-of-the-art Transformer-based foundation models, while achieving a model size of ~370KB under 8-bit quantization and demonstrating viability through latency and resource profiling on the MAX78000 and HX6538 μNPUs. The models and framework are released publicly.

Significance. If the performance claims hold, the work is significant for enabling real-time, offline physiological signal processing on battery-powered wearables with strong privacy guarantees. The explicit hardware profiling on two specific μNPUs and the public release of models plus training code are clear strengths that support reproducibility and practical deployment.

major comments (1)
  1. [Abstract] Abstract: the claim that PhysioLite 'reaches performance comparable to state-of-the-art Transformer-based foundation models on ECG and EMG benchmarks' is load-bearing for the central contribution yet provides no quantitative metrics, named datasets, baseline scores, or statistical tests; without these the size-performance tradeoff cannot be evaluated.
minor comments (1)
  1. [Abstract] Abstract: the phrase 'near-real-time' is used without a concrete latency target (e.g., ms per inference) that would allow readers to judge the hardware profiling results.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and positive recommendation for minor revision. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that PhysioLite 'reaches performance comparable to state-of-the-art Transformer-based foundation models on ECG and EMG benchmarks' is load-bearing for the central contribution yet provides no quantitative metrics, named datasets, baseline scores, or statistical tests; without these the size-performance tradeoff cannot be evaluated.

    Authors: We agree that the abstract would benefit from explicit quantitative support for the central claim. The manuscript body already contains the supporting experimental results, including performance metrics on named ECG and EMG benchmarks with direct comparisons to Transformer baselines. For the revision we will expand the abstract to include key quantitative metrics, dataset names, and baseline scores (while remaining within length limits) so that the size-performance tradeoff is immediately evaluable from the abstract alone. No changes to the underlying results or claims are required. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper introduces PhysioLite as an engineering architecture combining learnable wavelet filter banks, CPU-offloaded positional encoding, and hardware-aware layers to achieve compact size and μNPU compatibility. All performance claims are asserted via direct comparison to external Transformer baselines on standard ECG/EMG benchmarks, with code release for reproduction. No equations, derivations, or first-principles results are presented that reduce by construction to fitted inputs or self-citations; the central claims remain empirically grounded rather than tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on standard supervised learning assumptions and the representativeness of the cited ECG/EMG benchmarks; no new physical entities or ad-hoc constants are introduced beyond learned parameters.

pith-pipeline@v0.9.0 · 5523 in / 1006 out tokens · 42146 ms · 2026-05-10T05:52:44.740688+00:00 · methodology

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

59 extracted references · 41 canonical work pages · 2 internal anchors

  1. [1]

    Manal Alghieth. 2025. DeepECG-Net: a hybrid transformer-based deep learning model for real-time ECG anomaly detection.Scientific Reports15, 1 (2025), 20714. doi:10.1038/s41598-025-07781-1

    work page doi:10.1038/s41598-025-07781-1 2025

  2. [2]

    Analog Devices Inc. 2024. ai8x-synthesis: Quantization and Synthesis for MAX78000/MAX78002 Edge AI Devices. https://github.com/analogdevicesinc/ ai8x-synthesis. Accessed: 2026-04-02

    2024

  3. [3]

    Arm Ltd. 2024. Arm Ethos-U85 NPU. https://www.arm.com/products/silicon-ip- cpu/ethos/ethos-u85. Accessed: 2025-11-08

    2024

  4. [4]

    and NXP Semiconductors

    Arm Ltd. and NXP Semiconductors. 2024. ethos-u-vela: Neural Network Compiler for Arm Ethos-U NPUs. https://github.com/nxp-imx/ethos-u-vela. Accessed: 2026-04-02

    2024

  5. [5]

    Attia, Suraj Kapa, Francisco Lopez-Jimenez, Paul M

    Zachi I. Attia, Suraj Kapa, Francisco Lopez-Jimenez, Paul M. McKie, Dorothy J. Ladewig, Gaurav Satam, Patricia A. Pellikka, Maurice Enriquez-Sarano, Peter A. Noseworthy, Thomas M. Munger, Samuel J. Asirvatham, Christopher G. Scott, Rickey E. Carter, and Paul A. Friedman. 2019. Screening for cardiac contractile dysfunction using an artificial intelligence–...

    work page doi:10.1038/s41591-018-0240-2 2019

  6. [6]

    Benalcázar, Lorena Barona, Leonardo Valdivieso, Xavier Aguas, and Jonathan Zea

    Marco E. Benalcázar, Lorena Barona, Leonardo Valdivieso, Xavier Aguas, and Jonathan Zea. 2020. EMG-EPN-612 Dataset [Dataset]. Zenodo. doi:10.5281/ zenodo.4421500

    2020

  7. [7]

    Benalcázar, Andrés G

    Marco E. Benalcázar, Andrés G. Jaramillo, Jonathan, A. Zea, Andrés Páez, and Víctor Hugo Andaluz. 2017. Hand gesture recognition using machine learning and the Myo armband. In2017 25th European Signal Processing Conference (EUSIPCO). 1040–1044. doi:10.23919/EUSIPCO.2017.8081366

    work page doi:10.23919/eusipco.2017.8081366 2017

  8. [8]

    Yanlong Chen, Mattia Orlandi, Pierangelo Maria Rapa, Simone Benatti, Luca Benini, and Yawei Li. 2025. PhysioWave: A Multi-Scale Wavelet-Transformer for Physiological Signal Representation. arXiv:2506.10351 [cs.LG] https://arxiv.org/ abs/2506.10351

    work page arXiv 2025

  9. [9]

    Yanlong Chen, Mattia Orlandi, Pierangelo Maria Rapa, Simone Benatti, Luca Benini, and Yawei Li. 2025. WaveFormer: A Lightweight Transformer Model for sEMG-based Gesture Recognition. arXiv:2506.11168 [cs.CV] https://arxiv.org/ abs/2506.11168

    work page arXiv 2025

  10. [10]

    Subject-aware contrastive learning for biosignals

    Joseph Y. Cheng, Hanlin Goh, Kaan Dogrusoz, Oncel Tuzel, and Er- drin Azemi. 2020. Subject-Aware Contrastive Learning for Biosignals. arXiv:2007.04871 [cs.LG] https://arxiv.org/abs/2007.04871

    work page arXiv 2020

  11. [11]

    Romain Cosentino and Behnaam Aazhang. 2020. Learnable Group Transform for Time-Series. InProceedings of the 37th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 119). PMLR, 2164–2173. https://proceedings.mlr.press/v119/cosentino20a.html

    2020

  12. [12]

    Robert David, Jared Duke, Advait Jain, Vijay Janapa Reddi, Nat Jeffries, Jian Li, Nick Kreeger, Ian Nappier, Meghna Natraj, Shlomi Regev, Rocky Rhodes, Tiezhen Wang, and Pete Warden. 2021. TensorFlow Lite Micro: Embedded Machine Learning on TinyML Systems. arXiv:2010.08678 [cs.LG] https://arxiv.org/abs/ 2010.08678

    work page arXiv 2021

  13. [13]

    De Luca, L

    Carlo J. De Luca, L. Donald Gilmore, Mikhail Kuznetsov, and Serge H. Roy

  14. [14]

    An improved NExT-DMD for efficient automated operational modal analysis.Applied Mathematical Modelling2026,156, 116823

    Filtering the surface EMG signal: Movement artifact and baseline noise contamination.Journal of Biomechanics43, 8 (2010), 1573–1579. doi:10.1016/j. jbiomech.2010.01.027

    work page doi:10.1016/j 2010

  15. [15]

    Alberto Dequino, Luca Bompani, Luca Benini, and Francesco Conti. 2025. Op- timizing BFloat16 Deployment of Tiny Transformers on Ultra-Low Power Ex- treme Edge SoCs.Journal of Low Power Electronics and Applications15, 1 (2025). doi:10.3390/jlpea15010008

    work page doi:10.3390/jlpea15010008 2025

  16. [16]

    Ethan Eddy, Erik J Scheme, and Scott Bateman. 2023. A Framework and Call to Action for the Future Development of EMG-Based Input in HCI. InProceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI ’23). ACM, 1–23. doi:10.1145/3544548.3580962

    work page doi:10.1145/3544548.3580962 2023

  17. [17]

    Adams, Matthew Mattina, and Paul N

    Igor Fedorov, Ryan P. Adams, Matthew Mattina, and Paul N. Whatmough. 2019. SpArSe: Sparse Architecture Search for CNNs on Resource-Constrained Micro- controllers. arXiv:1905.12107 [cs.LG] https://arxiv.org/abs/1905.12107

    work page arXiv 2019

  18. [18]

    Sebastian Frey, Marco Guermandi, Simone Benatti, Victor Kartsch, Andrea Cosset- tini, and Luca Benini. 2023. BioGAP: A 10-Core FP-Capable Ultra-Low Power IoT Processor with Medical-Grade AFE and BLE Connectivity for Wearable Biosignal Processing.arXiv preprint arXiv:2307.01619(2023)

    work page arXiv 2023

  19. [19]

    Sebastian Frey, Giusy Spacone, Andrea Cossettini, Marco Guermandi, Philipp Schilk, Luca Benini, and Victor Kartsch. 2025. BioGAP-Ultra: A Modular Edge-AI Platform for Wearable Multimodal Biosignal Acquisition and Processing.arXiv preprint arXiv:2508.13728(2025)

    work page arXiv 2025

  20. [20]

    Brian Gow, Tom Pollard, Larry A Nathanson, Alistair Johnson, Benjamin Moody, Chrystinne Fernandes, Nathaniel Greenbaum, Jonathan W Waks, Parastou Eslami, Tanner Carbonati, Ashish Chaudhari, Elizabeth Herbst, Dana Moukheiber, Seth Berkowitz, Roger Mark, and Steven Horng. 2023. MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset.PhysioNet(Sept. 2023)....

    work page doi:10.13026/4nqg- 2023

  21. [21]

    Ahmed Imtiaz Humayun, Shabnam Ghaffarzadegan, Zhe Feng, and Taufiq Hasan

  22. [22]

    In2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)

    Learning Front-end Filter-bank Parameters using Convolutional Neural Networks for Abnormal Heart Sound Detection. In2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 1408–1411. doi:10.1109/embc.2018.8512578

    work page doi:10.1109/embc.2018.8512578 2018

  23. [23]

    Victor J. B. Jung, Alessio Burrello, Moritz Scherer, Francesco Conti, and Luca Benini. 2024. Optimizing the Deployment of Tiny Transformers on Low-Power MCUs. arXiv:2404.02945 [cs.LG] https://arxiv.org/abs/2404.02945

    work page arXiv 2024

  24. [24]

    Bhabesh Kalita, Nabamita Deb, and Daisy Das. 2024. AnEEG: leveraging deep learning for effective artifact removal in EEG data.Scientific Reports14, 1 (2024), 24234. doi:10.1038/s41598-024-75091-z

    work page doi:10.1038/s41598-024-75091-z 2024

  25. [25]

    Victor Javier Kartsch, Simone Benatti, Fiorenzo Artoni, Silvestro Micera, and Luca Benini. 2025. Open-source RISC-V platforms for embedded medical grade EMG processing: Are we there yet?Microprocessors and Microsystems(2025), 105239

    2025

  26. [26]

    Albert, Joel Xue, Aarya Parekh, Reza Sameni, Matthew A

    Zuzana Koscova, Qiao Li, Chad Robichaux, Valdery Moura Junior, Manohar Ghanta, Aditya Gupta, Jonathan Rosand, Aaron Aguirre, Shenda Hong, David E. Albert, Joel Xue, Aarya Parekh, Reza Sameni, Matthew A. Reyna, M. Brandon Westover, and Gari D. Cli- ford. 2024. The Harvard-Emory ECG Database.medRxiv(2024). arXiv:https://www.medrxiv.org/content/early/2024/10...

    work page doi:10.1101/2024.09.27.24314503 2024

  27. [27]

    Krilova, I

    N. Krilova, I. Kastalskiy, V. Kazantsev, V. Makarov, and S. Lobov. 2018. EMG Data for Gestures [Dataset]. UCI Machine Learning Repository. doi:10.24432/C5ZP5C

    work page doi:10.24432/c5zp5c 2018

  28. [28]

    Jun Li, Aaron Aguirre, Junior Moura, Che Liu, Lanhai Zhong, Chenxi Sun, Gari Clifford, Brandon Westover, and Shenda Hong. 2025. An Electrocardiogram Foundation Model Built on over 10 Million Recordings with External Evaluation across Multiple Domains. arXiv:2410.04133 [cs.LG] https://arxiv.org/abs/2410. 04133

    work page arXiv 2025

  29. [29]

    Yinan Liang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu

  30. [30]

    arXiv:2310.16898 [cs.CV] https://arxiv.org/abs/2310.16898

    MCUFormer: Deploying Vision Transformers on Microcontrollers with Limited Memory. arXiv:2310.16898 [cs.CV] https://arxiv.org/abs/2310.16898

    work page arXiv

  31. [31]

    Edgar Liberis, Łukasz Dudziak, and Nicholas D. Lane. 2021. 𝜇NAS: Constrained Neural Architecture Search for Microcontrollers. InProceedings of the 1st Work- shop on Machine Learning and Systems(Online, United Kingdom)(EuroML- Sys ’21). Association for Computing Machinery, New York, NY, USA, 70–79. doi:10.1145/3437984.3458836

    work page doi:10.1145/3437984.3458836 2021

  32. [32]

    Ng, et al

    Cheng Liu, E.Y.-K. Ng, et al. 2018. An Open Access Database for Evaluating the Algorithms of the 1st China Physiological Signal Challenge 2018. Proceedings of the 1st China Physiological Signal Challenge (CPSC) 2018, 7th International Conference on Biomedical Engineering and Biotechnology (ICBEB 2018). doi:10. 1166/jmihi.2018.2442

    work page arXiv 2018

  33. [33]

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. 2022. A ConvNet for the 2020s. arXiv:2201.03545 [cs.CV] https://arxiv.org/abs/2201.03545

    work page arXiv 2022

  34. [34]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. arXiv:1711.05101 [cs.LG] https://arxiv.org/abs/1711.05101

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  35. [35]

    2021.MAX78000 Microcontroller with Ultra-Low-Power Con- volutional Neural Network Accelerator

    Maxim Integrated. 2021.MAX78000 Microcontroller with Ultra-Low-Power Con- volutional Neural Network Accelerator. Maxim Integrated. https://www.analog. com/media/en/technical-documentation/data-sheets/MAX78000.pdf Data Sheet Rev 1; 19-100868; 5/21

    2021

  36. [36]

    Kaden McKeen, Sameer Masood, Augustin Toma, Barry Rubin, and Bo Wang. 2025. ECG-FM: An Open Electrocardiogram Foundation Model. arXiv:2408.05178 [cs.LG] https://arxiv.org/abs/2408.05178

    work page arXiv 2025

  37. [37]

    Gabriel Michau, Gaetan Frusque, and Olga Fink. 2022. Fully learnable deep wavelet transform for unsupervised monitoring of high-frequency time series. Proceedings of the National Academy of Sciences119, 8 (Feb. 2022). doi:10.1073/ pnas.2106598119

    2022

  38. [38]

    Josh Millar, Yushan Huang, Sarab Sethi, Hamed Haddadi, and Anil Madhavapeddy

  39. [39]

    arXiv:2503.22567 [cs.LG] https: //arxiv.org/abs/2503.22567

    Benchmarking Ultra-Low-Power 𝜇NPUs. arXiv:2503.22567 [cs.LG] https: //arxiv.org/abs/2503.22567

    work page arXiv

  40. [40]

    Monsoon Solutions Inc. 2023. High Voltage Power Monitor. https://www.msoon. com/high-voltage-power-monitor Accessed: 2026-04-02

    2023

  41. [41]

    Arthur Moss, Hyunjong Lee, Lei Xun, Chulhong Min, Fahim Kawsar, and Alessandro Montanari. 2023. Ultra-Low Power DNN Accelerators for IoT: Re- source Characterization of the MAX78000. InProceedings of the 20th ACM Con- ference on Embedded Networked Sensor Systems(Boston, Massachusetts)(Sen- Sys ’22). Association for Computing Machinery, New York, NY, USA, ...

    work page doi:10.1145/3560905.3568300 2023

  42. [42]

    Oliver and F

    N. Oliver and F. Flores-Mangas. 2006. HealthGear: a real-time wearable system for monitoring and analyzing physiological signals. InInternational Workshop on Towards Real-Time ECG and EMG Modeling onµNPUs SenSys ’26, May 11–14, 2026, Saint Malo, France Wearable and Implantable Body Sensor Networks (BSN’06). 4 pp.–64. doi:10.1109/ BSN.2006.27

    2006

  43. [43]

    Stefano Pizzolato, Luca Tagliapietra, Matteo Cognolato, Monica Reggiani, Hen- ning Müller, and Manfredo Atzori. 2017. Comparison of six electromyography acquisition setups on hand movement classification tasks.PLOS ONE12, 10 (2017), e0186132. doi:10.1371/journal.pone.0186132

    work page doi:10.1371/journal.pone.0186132 2017

  44. [44]

    Rajashekar Reddy, Vivian Dsouza, Ashok Samraj Thangarajan, Przemysław Pawełczak, Fahim Kawsar, and Alessandro Montanari

    C. Rajashekar Reddy, Vivian Dsouza, Ashok Samraj Thangarajan, Przemysław Pawełczak, Fahim Kawsar, and Alessandro Montanari. 2025. BioPulse: Towards Enabling Perpetual Vital Signs Monitoring Using a Body Patch. InProceedings of the 26th International Workshop on Mobile Computing Systems and Applications (HotMobile ’25). Association for Computing Machinery,...

    work page doi:10.1145/3708468.3711891 2025

  45. [45]

    Matthew A Reyna, Nadi Sadr, Erick A Perez Alday, Annie Gu, Amit J Shah, Chad Robichaux, Ali Bahrami Rad, Andoni Elola, Salman Seyedi, Sardar Ansari, Hamid Ghanbari, Qiao Li, Ashish Sharma, and Gari D Clifford. 2021. Will Two Do? Varying Dimensions in Electrocardiography: The PhysioNet/Computing in Cardiology Challenge 2021. In2021 Computing in Cardiology ...

    work page doi:10.23919/cinc53138.2021.9662687 2021

  46. [46]

    Manuele Rusci, Alessandro Capotondi, and Luca Benini. 2020. Memory-Driven Mixed Low Precision Quantization for Enabling Deep Network Inference on Microcontrollers. InProceedings of Machine Learning and Systems, I. Dhillon, D. Papailiopoulos, and V. Sze (Eds.), Vol. 2. 326–335. https://proceedings.mlsys. org/paper_files/paper/2020/file/0c8abcf158ed12d0dd94...

    2020

  47. [47]

    Swapnil Sayan Saha, Sandeep Singh Sandha, and Mani Srivastava. 2022. Machine Learning for Microcontroller-Class Hardware: A Review.IEEE Sensors Journal 22, 22 (2022), 21362–21390. doi:10.1109/JSEN.2022.3210773

    work page doi:10.1109/jsen.2022.3210773 2022

  48. [48]

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang- Chieh Chen. 2019. MobileNetV2: Inverted Residuals and Linear Bottlenecks. arXiv:1801.04381 [cs.CV] https://arxiv.org/abs/1801.04381

    work page arXiv 2019

  49. [49]

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. 2023. RoFormer: Enhanced Transformer with Rotary Position Embedding. arXiv:2104.09864 [cs.CL] https://arxiv.org/abs/2104.09864

    work page internal anchor Pith review arXiv 2023

  50. [50]

    Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S. Emer. 2020. How to Evaluate Deep Neural Network Processors: TOPS/W (Alone) Considered Harmful. IEEE Solid-State Circuits Magazine12, 3 (2020), 28–41. doi:10.1109/MSSC.2020. 3002140

    work page doi:10.1109/mssc.2020 2020

  51. [51]

    Jack Tchimino, Rehne Lessmann Hansen, Peter Holmberg Jørgensen, Jakob Dideriksen, and Strahinja Dosen. 2024. Application of EMG feedback for hand prosthesis control in high-level amputation: a case study.Scientific Reports14, 1 (2024), 31676. doi:10.1038/s41598-024-80828-x

    work page doi:10.1038/s41598-024-80828-x 2024

  52. [52]

    GreenWaves Technologies. 2021. GAP9 Product Brief

    2021

  53. [53]

    Himax Technologies. 2025. WiseEye2 AI Processor. https://www.himax.com. tw/products/wiseeye-ai-sensing/wiseeye2-ai-processor/ Accessed: 2025-03-12

    2025

  54. [54]

    Trockman and J

    Asher Trockman and J. Zico Kolter. 2022. Patches Are All You Need? arXiv:2201.09792 [cs.CV] https://arxiv.org/abs/2201.09792

    work page arXiv 2022

  55. [55]

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Wojciech Samek, and Tobias Schaeffter. 2022. PTB-XL, a large publicly available electrocardiography dataset (version 1.0.3). PhysioNet. doi:10.13026/kfzx-aw45 RRID:SCR_007345

    work page doi:10.13026/kfzx-aw45 2022

  56. [56]

    Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. 2022. MetaFormer Is Actually What You Need for Vision. arXiv:2111.11418 [cs.CV] https://arxiv.org/abs/2111.11418

    work page arXiv 2022

  57. [57]

    Jianwei Zheng, Hangyuan Guo, and Huimin Chu. 2022. A large scale 12-lead electrocardiogram database for arrhythmia study (version 1.0.0). PhysioNet. doi:10.13026/wgex-er52 RRID:SCR_007345

    work page doi:10.13026/wgex-er52 2022

  58. [58]

    Norbert Ádám, Dávid Val’ko, Zoltán Balogh, Branislav Madoš, and Ján Hurtuk

  59. [59]

    Comparative evaluation of filtration techniques for ECG signal denoising with emphasis on stationary wavelet transform.Scientific Reports15, 1 (2025), 42514. doi:10.1038/s41598-025-26476-1 A Ablation Study We conduct ablation studies to quantify the contribution of key architectural and training components, including knowledge dis- tillation, positional e...

    work page doi:10.1038/s41598-025-26476-1 2025