Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Design of an Edge-based Portable EHR System for Anemia Screening in Remote Health Applications

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read An edge-based EHR with a fingernail anemia module reaches clinically useful screening accuracy while running fully offline on low-power hardware.

desk verdict A solid systems-integration paper whose clinical screening claim rests on possibly rebalanced test data—fixable, but needs referee scrutiny. read the letter →

arxiv 2507.15146 v1 pith:V2KQL2MM submitted 2025-07-20 cs.ET cs.AIcs.CVcs.CYcs.LGcs.SE

Sebastian A. Cruz Romero , Misael J. Mercado Hernandez , Samir Y. Ali Rivera , Jorge A. Santiago Fernandez , Wilfredo E. Lugo Beauchamp This is my paper
classification cs.ETcs.AIcs.CVcs.CYcs.LGcs.SE
keywords edgecomputingportableEHRanemiascreeningfingernailpallorhemoglobinestimationrandomforestINT8quantizationoffline-firstarchitecture
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 argues that the practical barriers to digital health in remote areas—no stable power, no internet, little IT support—can be addressed by treating them as design requirements rather than obstacles. To show this, the authors built a portable EHR system that runs entirely on-device on a roughly 10-watt embedded computer, encrypts records with AES-256, and integrates an anemia screening module that estimates hemoglobin from fingernail photographs. On a publicly available 250-case dataset, a Random Forest model trained with KDE-balanced data reaches an RMSE of 1.969 g/dL and an MAE of 1.490 g/dL for hemoglobin, while a severity-based version reaches 79.2% sensitivity. An INT8-quantized YOLOv8n nail-bed detector holds mAP@0.5 at 0.995 while cutting inference latency to about 21.5 ms. If the claim holds, frontline workers in disconnected clinics could screen and track anemia without sending patient data or images to the cloud.

What carries the argument

The load-bearing mechanism is the two-stage joint-model screening pipeline. First, a YOLOv8n detector, quantized to INT8 by post-training quantization, extracts nail-bed regions; then color normalization and statistical features from RGB and L*a*b* spaces feed a Random Forest model. KDE-based rebalancing amplifies under-represented hemoglobin ranges and is what lets a small, imbalanced dataset (27% anemia prevalence) produce the reported sensitivity. The system is held together by a REST API, a local PostgreSQL database with record-level AES-256 encryption and role-based access control, and a ReactJS/FastAPI frontend, all running on a Jetson-class board under 10 watts.

What would settle it

Re-run the Random Forest evaluation on a test set drawn from the original, untouched 27%-prevalence distribution, or recruit a consecutive series of participants in a field clinic and compare the device's severity classifications with laboratory hemoglobin measurements. If severity sensitivity falls clearly below 79.2% or the RMSE exceeds roughly 2 g/dL in that setting, the claim of acceptable clinical screening performance on edge devices would be contradicted.

Watch

Extended reading notes

Core claim

The central claim is that a modest, constraint-driven system can deliver clinically useful anemia screening at the point of care. The system's nail-bed detector (YOLOv8n, quantized from FP32 to INT8) locates the region of interest, an automated preprocessing step normalizes color and derives RGB and L*a*b* statistics, and a Random Forest regressor or classifier maps those features to hemoglobin and severity. Trained on the 250-image dataset with KDE-based rebalancing and 7-fold cross-validation, the regressor achieves $ ext{RMSE} = 1.969$ g/dL and $ ext{MAE} = 1.490$ g/dL, and the severity classifier reaches 79.2% sensitivity. The paper treats these numbers, together with sub-7-watt power draw and record-level AES-256 encryption, as confirmation that acceptable clinical screening performance can be attained on low-power edge devices and that the architecture generalizes to other point-of-care diagnostics.

Load-bearing premise

The reported accuracy assumes the test examples were balanced the same way as the training examples; if they instead came from the natural population, where only about 27% of patients have anemia, the headline sensitivity and error may not reflect real-world screening.

Editorial extensions

If this is right

  • In a disconnected clinic, a worker can photograph a patient's fingernails and receive a hemoglobin estimate and severity class in tens of milliseconds, with the record stored encrypted on the device.
  • Because the severity model's 79.2% sensitivity comes with 65.7% specificity, the system is best used as a triage screen that flags suspected anemia for laboratory confirmation rather than as a replacement for lab testing.
  • INT8 quantization cuts detector latency from 46.96 ms to 21.50 ms while preserving mAP@0.5 at 0.995, showing that deep-learning components can be added to the pipeline without losing real-time edge performance.
  • The modular REST/EHR architecture lets new screening modules be added without redesigning storage or interfaces, extending the platform beyond anemia.

Reading between the lines

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

  • If the test set was rebalanced to match the training distribution, the headline numbers may not carry over to a natural population; the paper does not state how the original 27% prevalence was treated in the test split.
  • The most direct test of the platform would be a prospective field trial in a rural clinic with consecutive patients, comparing device predictions against standard laboratory hemoglobin; sensitivity at natural prevalence is the metric that would settle clinical utility.
  • The same two-stage skeleton (detector plus color-feature regressor) could plausibly be reused for other non-invasive visual screens such as jaundice or skin lesions, but each condition would require its own annotated dataset and clinical validation, which the paper does not provide.
  • Given the 21.5 ms per-image detector latency, a battery-powered device could acquire several images per patient and average the hemoglobin predictions; this is a cheap way to reduce measurement noise that the paper does not test.
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

3 major / 5 minor

Summary. The manuscript describes an offline-first, edge-deployed Electronic Health Record platform with an integrated non-invasive anemia screening module based on fingernail pallor analysis. The system combines a YOLOv8n nail-bed detector quantized to INT8, a Random Forest hemoglobin estimator trained on 250 public-dataset images with KDE-based rebalancing, and a severity-oriented classifier, together with AES-256 encrypted local PostgreSQL storage, RBAC, and an offline-capable React frontend. The authors report a test RMSE of 1.969 g/dL for hemoglobin estimation, 79.2% sensitivity for severity-based screening, a 54.2% reduction in detector inference latency after quantization, and sub-7W power consumption on Jetson hardware, and they argue that these results confirm acceptable clinical screening performance on low-power edge devices.

Significance. If the reported diagnostic metrics are representative of the population in which the system would be deployed, the paper is a useful systems demonstration: it combines a clinically motivated screening task, public data, explicit resource constraints, detector quantization, and a functional offline EHR stack. The strengths that deserve credit are the use of a public dataset, the concrete edge-deployment measurements, the maintained detection mAP after INT8 quantization, and the modular REST-based architecture. However, the central clinical claim currently rests on an evaluation protocol whose relationship to the natural 27% anemia prevalence is not specified. The paper's own limitations section acknowledges that the system has not undergone clinical field validation, so the Discussion's inference from benchmark metrics to clinical acceptability is not yet supported.

major comments (3)
  1. [III.A.2 and Tables II-III] The manuscript does not state whether KDE-based rebalancing was applied to the training folds only or also to the test folds. Table III reports sensitivity 0.792 and specificity 0.657 on "KDE balanced data by severity," and Figure 9 shows validation/test scatter plots "after KDE-based rebalancing." If the test set was rebalanced, these metrics are not estimates for the natural-prevalence screening population. Applying the reported severity-model sensitivity and specificity to the dataset's 27% anemia prevalence gives a positive predictive value of roughly 0.46, meaning more than half of positive screens would be false alarms. The claim in Section III.C that the results "confirm that acceptable clinical screening performance can be attained" is therefore not supported unless the authors report metrics on a natural-prevalence test set or explicitly recalibrate the operating threshold and report PPV/NPV.
  2. [II.C and IV.2] No confidence intervals or variance estimates are reported for any of the diagnostic metrics. With only 250 patient cases and 7-fold cross-validation, point estimates such as RMSE 1.969 g/dL and sensitivity 0.792 have nontrivial uncertainty, and the paper does not report the spread across folds. In addition, the Random Forest hyperparameters (100 trees, maximum depth 10), the KDE bandwidth/rebalancing parameters, and the severity classification threshold are introduced without justification or sensitivity analysis. Given that Section IV.2 explicitly states that the system "has yet to undergo clinical validation in real-world field deployments," the Discussion's inference from benchmark numbers to clinical acceptability needs to be substantially tempered or supported by additional validation evidence.
  3. [III.A.2 and Figures 8-9] The evaluation conflates regression and classification reporting. Sensitivity and specificity are listed in Tables II and III for regression models, but no definition is given of how a continuous hemoglobin prediction is converted into a binary or severity label, nor whether the threshold was pre-specified or chosen after inspecting validation results. The confusion matrices in Figures 8b and 9b are described as multiclass predictions on the test set, but class-wise sensitivity, specificity, and sample counts are not provided. Without a transparent statement of the thresholding rule and the exact train/validation/test split, the headline sensitivity of 79.2% cannot be interpreted or reproduced.
minor comments (5)
  1. [Abstract and III.A.2] The abstract reports 79.2% sensitivity for the severity-based model without mentioning the corresponding specificity of 0.657 or RMSE of 2.264 g/dL; reporting the full operating point would avoid a selective summary.
  2. [Table II] The caption says "Performance summary of regression models on KDE balanced data by remark," yet the table columns are sensitivity and specificity; the threshold used to derive these classification metrics from regression outputs should be stated in the caption or text.
  3. [Figures 8 and 9] The captions refer to "validation (blue) and test (red)" sets, while the text says binary classification is on validation and multiclass is on test; the figure labels and text should be harmonized, and the split ratio and whether KDE was applied before or after the split should be stated.
  4. [II.F] The hardware naming is inconsistent: Section II.A and Section IV.2 refer to the Jetson Nano, while Section II.F refers to the Jetson Orin Nano; the authors should use one accurate device name throughout.
  5. [General] The manuscript would benefit from a data/code availability statement, since the screening model is trained on a public dataset and reproducibility would be considerably improved by releasing the preprocessing and evaluation scripts.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: empirical benchmarks on external data, with evaluation-protocol caveats.

full rationale

This is an empirical systems paper, not a formal derivation. The central performance claims (RMSE 1.969 g/dL, 79.2% sensitivity, YOLO latency reduction) are measurements obtained by training standard models on a public dataset [25] and benchmarking on Jetson hardware; none of the reported quantities is identical by construction to a fitted parameter or to a self-citation. The KDE rebalancing reported in Section III.A.2 and Tables II-III is an evaluation-protocol concern: if the test set was rebalanced rather than reflecting the original 27% prevalence, the metrics may not transfer to real screening populations, and Section IV.2 concedes the system 'has yet to undergo clinical validation in real-world field deployments.' That is an evidentiary limitation, not a circular step, because the model outputs could have been poor even on the rebalanced data. References are to external prior work (e.g., the Yakimov dataset and standard quantization surveys); no load-bearing self-citation or imported uniqueness theorem is invoked. The Discussion's wording 'confirms acceptable clinical screening performance' overstates the evidence, but overstatement is a correctness risk rather than a circularity.

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

The model's performance depends on several unstated choices: Random Forest hyperparameters, KDE rebalancing parameters, the feature list, and the classification threshold. The paper also relies on the public dataset's ground-truth labels and the clinical premise that nail color correlates with hemoglobin. No new entities are introduced.

free parameters (3)
  • Random Forest hyperparameters = 100 trees, max depth 10
    Chosen without a systematic search; directly affect the reported RMSE and sensitivity.
  • KDE rebalancing parameters = not specified
    The degree of undersampling or reweighting is not quantified; it shapes the test distribution and the reported sensitivity and RMSE.
  • Severity classification threshold = not specified
    The 79.2% sensitivity and 65.7% specificity depend on an unspecified decision threshold for the severity model.
assumptions (3)
  • domain assumption The Yakimov et al. dataset is representative of real-world fingernail images and populations.
    The study uses 250 images from a public dataset; generalizability to other populations, skin tones, and lighting conditions is assumed without external validation.
  • domain assumption Hemoglobin concentration is recoverable from RGB/L*a*b color statistics of fingernail images.
    This is the clinical premise of the screening module; the paper does not establish it independently beyond the dataset's original publication.
  • domain assumption The laboratory hemoglobin values in the dataset are accurate ground truth.
    The paper trusts the dataset's FDA-certified lab measurements; any systematic error in those labels would propagate to the model evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Design of an Edge-based Portable EHR System for Anemia Screening in Remote Health Applications." pith.science (2026). https://pith.science/paper/V2KQL2MM

@misc{pith2026250715146,
  author       = {Pith},
  title        = {Pith review of: Design of an Edge-based Portable EHR System for Anemia Screening in Remote Health Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V2KQL2MM}},
  note         = {Machine review of arXiv:2507.15146}
}
read the original abstract

The design of medical systems for remote, resource-limited environments faces persistent challenges due to poor interoperability, lack of offline support, and dependency on costly infrastructure. Many existing digital health solutions neglect these constraints, limiting their effectiveness for frontline health workers in underserved regions. This paper presents a portable, edge-enabled Electronic Health Record platform optimized for offline-first operation, secure patient data management, and modular diagnostic integration. Running on small-form factor embedded devices, it provides AES-256 encrypted local storage with optional cloud synchronization for interoperability. As a use case, we integrated a non-invasive anemia screening module leveraging fingernail pallor analysis. Trained on 250 patient cases (27\% anemia prevalence) with KDE-balanced data, the Random Forest model achieved a test RMSE of 1.969 g/dL and MAE of 1.490 g/dL. A severity-based model reached 79.2\% sensitivity. To optimize performance, a YOLOv8n-based nail bed detector was quantized to INT8, reducing inference latency from 46.96 ms to 21.50 ms while maintaining mAP@0.5 at 0.995. The system emphasizes low-cost deployment, modularity, and data privacy compliance (HIPAA/GDPR), addressing critical barriers to digital health adoption in disconnected settings. Our work demonstrates a scalable approach to enhance portable health information systems and support frontline healthcare in underserved regions.

Figures

Figures reproduced from arXiv: 2507.15146 by the authors.

Figure 1
Figure 1. System architecture for our patient monitoring and clinical records platform, integrating edge and smart devices with [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Histogram showing the empirical distribution of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Kernel Density Estimation (KDE) was used to rebal [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Image samples from Yakimov et. al. [25] including a [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 6
Figure 6. Figure 6: Single-use app view for quick intervention [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Key interfaces of the Point-of-Care system, including login, patient records, chart overview, and sample submission [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 9
Figure 9. Figure 9: The severity-balanced model shows improved linearity [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 28 canonical work pages

  1. [1]

    Contextual barriers to im- plementing open-source electronic health record systems in low-resource settings,

    V . Makinen, E. Osei, R. Smith, and N. Patel, “Contextual barriers to im- plementing open-source electronic health record systems in low-resource settings,” Journal of Medical Internet Research , vol. 26, p. e11327637, 2024

    work page 2024

  2. [2]

    Improved healthcare access in low-resource regions: A review of technological solutions,

    A. Ravichandran, M. Li, A. Patel, and S. Kim, “Improved healthcare access in low-resource regions: A review of technological solutions,” IEEE Access, vol. 10, pp. 112345–112362, 2022

    work page 2022

  3. [3]

    Iot-enabled low-cost fog computing system with online machine learning for accurate and low- latency heart monitoring in rural healthcare settings,

    F. Alshahrani, S. Kumar, and M. Gonzalez, “Iot-enabled low-cost fog computing system with online machine learning for accurate and low- latency heart monitoring in rural healthcare settings,” in Proceedings of the IEEE Global Humanitarian Technology Conference (GHTC) , 2023

    work page 2023

  4. [4]

    Harnessing the digital revolution: A comprehensive review of mhealth applications for remote monitor- ing in transforming healthcare delivery,

    L. Chen, A. Rahman, and H. Singh, “Harnessing the digital revolution: A comprehensive review of mhealth applications for remote monitor- ing in transforming healthcare delivery,” IEEE Reviews in Biomedical Engineering, vol. 17, pp. 120–138, 2024

    work page 2024

  5. [5]

    Integration of remote patient monitoring systems into physicians’ work in underserved communities: Survey of healthcare provider perspectives,

    A. T. Nguyen, D. Morales, and P. Chandra, “Integration of remote patient monitoring systems into physicians’ work in underserved communities: Survey of healthcare provider perspectives,” in Proceedings of the 2022 IEEE International Conference on Health Informatics (ICHI) , 2022

    work page 2022

  6. [6]

    Edge Computing For Smart Health: Context-aware Approaches, Opportunities, and Challenges

    A. A. Abdellatif, A. Mohamed, C. F. Chiasserini, M. Tlili, and A. Erbad, “Edge computing for smart health: Context-aware approaches, opportu- nities, and challenges,” arXiv preprint arXiv:2004.07311 , 2020

    work page Pith review arXiv 2004

  7. [7]

    I-Health: Leveraging Edge Computing and Blockchain for Epidemic Management

    A. A. Abdellatif, L. Samara, A. Mohamed, A. Erbad, C. F. Chiasserini, M. Guizani, M. D. O’Connor, and J. Laughton, “I-health: Leveraging edge computing and blockchain for epidemic management,” arXiv preprint arXiv:2012.14294, 2020

    work page Pith review arXiv 2012

  8. [8]

    BEdgeHealth: A Decentralized Architecture for Edge-based IoMT Networks Using Blockchain

    D. C. Nguyen, P. N. Pathirana, M. Ding, and A. Seneviratne, “Bedge- health: A decentralized architecture for edge-based iomt networks using blockchain,” arXiv preprint arXiv:2109.14295 , 2021

    work page Pith review arXiv 2021

Show all 29 references
  1. [9]

    Transforming healthcare in low- resource settings with artificial intelligence,

    J. Smith, K. Osei, and N. Patel, “Transforming healthcare in low- resource settings with artificial intelligence,” Public Health Nursing , vol. 40, no. 1, pp. 123–130, 2025

    2025

  2. [10]

    Saving lives at the edge: How edge computing drives healthcare in disadvantaged networks,

    B. Cushing, “Saving lives at the edge: How edge computing drives healthcare in disadvantaged networks,” HealthTech Magazine, 2024

    2024

  3. [11]

    A comprehensive review on ai- driven mhealth systems: Barriers and opportunities in remote health- care,

    L. Chen, A. Rahman, and H. Singh, “A comprehensive review on ai- driven mhealth systems: Barriers and opportunities in remote health- care,” IEEE Access, vol. 11, pp. 120345–120369, 2023

    2023

  4. [12]

    Medaide: Leveraging large language models for on-premise medical assistance on edge devices,

    A. Basit, K. Hussain, M. A. Hanif, and M. Shafique, “Medaide: Leveraging large language models for on-premise medical assistance on edge devices,” arXiv preprint arXiv:2403.00830 , 2024

    2024 arXiv

  5. [13]

    Iron deficiency anemia: A global public health concern,

    A. Manish, “Iron deficiency anemia: A global public health concern,” Int. J. Clin. Biochem. Res. , vol. 11, pp. 229–236, Jan. 2025

    2025

  6. [14]

    Anemia in low-income and middle-income countries,

    Y . Balarajan, U. Ramakrishnan, E. ¨Ozaltin, A. H. Shankar, and S. V . Subramanian, “Anemia in low-income and middle-income countries,” The Lancet, vol. 378, no. 9809, pp. 2123–2135, 2011

    2011

  7. [15]

    The global burden of anemia,

    N. J. Kassebaum et al. , “The global burden of anemia,” The Lancet Global Health, 2014

    2014

  8. [16]

    Prevalence and factors associated with anemia in women of reproductive age across low- and middle-income countries based on national data,

    A. Z. Alem, F. Efendi, L. McKenna, E. B. Felipe-Dimog, D. Chilot, S. I. Tonapa, I. A. Susanti, and A. Zainuri, “Prevalence and factors associated with anemia in women of reproductive age across low- and middle-income countries based on national data,” Sci. Rep. , vol. 13, p. 2...

    2023

  9. [17]

    An approach to hemequity: Identifying the barriers and facilitators of iron deficiency reduction strategies in low- to middle-income countries,

    S. Ge, S. Ali, V . Haldane, C. Bekdache, G. H. Tang, and M. Sholzberg, “An approach to hemequity: Identifying the barriers and facilitators of iron deficiency reduction strategies in low- to middle-income countries,” Br. J. Haematol., vol. 206, pp. 428–442, Feb. 2025

    2025

  10. [18]

    Edge ai: A taxonomy, systematic review and future directions,

    S. S. Gill, M. Golec, J. Hu, M. Xu, J. Du, H. Wu, G. K. Walia, S. S. Murugesan, B. Ali, M. Kumar, K. Ye, P. Verma, S. Kumar, F. Cuadrado, and S. Uhlig, “Edge ai: A taxonomy, systematic review and future directions,” 2024

    2024

  11. [19]

    Edge ai: A survey,

    R. Singh and S. S. Gill, “Edge ai: A survey,” Internet of Things and Cyber-Physical Systems, vol. 3, pp. 71–92, 2023

    2023

  12. [20]

    Engineering a sustainable future for point-of- care diagnostics and single-use microfluidic devices,

    A. E. Ongaro, Z. Ndlovu, E. Sollier, C. Otieno, P. Ondoa, A. Street, and M. Kersaudy-Kerhoas, “Engineering a sustainable future for point-of- care diagnostics and single-use microfluidic devices,” Lab Chip, vol. 22, pp. 3122–3137, Aug. 2022

    2022

  13. [21]

    Health-care waste,

    World Health Organization, “Health-care waste,” 2023. [Online]. Available: https://www.who.int/news-room/fact-sheets/detail/ health-care-waste

    2023

  14. [22]

    Sustainable medical technology: ad- dressing the environmental impact of single-use devices,

    J. Bressler and S. B. Bhatia, “Sustainable medical technology: ad- dressing the environmental impact of single-use devices,” Journal of Environmental Health, vol. 86, no. 4, pp. 23–29, 2023

    2023

  15. [23]

    NVIDIA Corporation, Santa Clara, CA, March 2023

    NVIDIA Corporation, Jetson Orin Nano Series Datasheet . NVIDIA Corporation, Santa Clara, CA, March 2023. Document Number: DS- 11105-001 v1.1

    2023

  16. [24]

    Advanced encryption standard (aes),

    N. I. of Standards and Technology, “Advanced encryption standard (aes),” Tech. Rep. FIPS 197-upd1, U.S. Department of Commerce, Gaithersburg, MD, 2023

    2023

  17. [25]

    Dataset of human skin and fingernails images for non- invasive haemoglobin level assessment,

    B. Yakimov, K. Buiankin, G. Denisenko, I. Bardadin, O. Pavlov, Y . Shitova, A. Yuriev, L. Pankratieva, A. Pukhov, A. Shkoda, and E. Shirshin, “Dataset of human skin and fingernails images for non- invasive haemoglobin level assessment,” Scientific Data, vol. 11, no. 1, p. 1070, 2024

    2024

  18. [26]

    Health insurance portability and accountability act (hipaa),

    U.S. Department of Health & Human Services, “Health insurance portability and accountability act (hipaa),” 2023. Office for Civil Rights (OCR)

    2023

  19. [27]

    General data protection regulation (gdpr),

    European Union, “General data protection regulation (gdpr),” 2016. Regulation (EU) 2016/679, Apr. 27, 2016

    2016

  20. [28]

    P3 International Corporation, New York, NY , August 2015

    P3 International Corporation, Kill A Watt™ Electricity Usage Monitor - Model P4400 . P3 International Corporation, New York, NY , August 2015

    2015

  21. [29]

    A survey of quantization methods for efficient neural network inference,

    A. Gholami, S. Kim, Z. Dong, et al., “A survey of quantization methods for efficient neural network inference,” 2021. arXiv preprint: https:// arxiv.org/abs/2103.13630

    2021 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.