Lightweight Deep Learning for Real‑Time Health Monitoring on

Edge Devices

Eswararao Boddepalli

, K. Sindhuja

, K. Akila

, M. Dharani

, G. Nagarjunarao

and Akash K.

Department of Information Technology, J.J.College of Engineering and Technology, Tiruchirappalli, Tamil Nadu, India

Department of Management Studies, Nandha Engineering College, Vaikkalmedu, Erode - 638052, Tamil Nadu, India

Department of Computer Science and Engineering, MLR Institute of Technology, Hyderabad, Telangana, India

Department of MCA, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

Keywords: Edge Computing, Health Monitoring, Lightweight Deep Learning, Mobile Devices, Real‑Time Analytics.

Abstract: Deep learning in edge computing is revolutionizing healthcare, delivering real-time tracking of crucial health

variables on mobile and wearable. We propose to design "lightweight" deep learning models that are tailored

for small scale edge devices. The proposed framework solves the problems concerning the limited

computational capacity, energy consumption, and variability of the physiological signals, in order to

accomplish a reliable, real-time physiological analysis with cloud connectivity not required. The system is

optimized for deployment at the edge, providing low-latency and high-throughput performance under realistic

conditions and enabling real-time health monitoring, early detection of anomalies, and personalized feedback.

Experimental results show promising accuracy and low resource consumption for the models, making them

practically deployable in large-scale mobile health ecosystems.

1 INTRODUCTION

One of the emerging trends of the past few years has

been the point at which artificial intelligence (AI)

meets healthcare, an event that is creating intelligent

systems that help redefine the way we access and

experience healthcare. Of these progressions, the use

of deep learning for health monitoring has attracted

much interest because of its ability to interpret

complex physiological data with high accuracy.

Nevertheless, the performance of deep learning

models can be inhibited because it relies on high

performance computing environment usually

provided by cloud systems. This dependence

imposes various important challenges, such as high

latency, high power consumption, network reliability,

and raises concerns about data privacy and security

particularly in situations where constant and real time

monitoring is required.

Edge computing has been proposed as a potential

solution to circumvent these challenges, processing

the data in or near the device of origin. Moving

computations from central servers to devices at the

network edge, such as smartphones, wearables, or

sensors embedded in IoT, edge computing leads to

quicker response times, less bandwidth

consumption, and greater protection of private health

data. However, the use of deep learning in edge

devices is not trivial due to the fact that conventional

neural networks are usually big, memory-rich and

computationally expensive. These restrictions render

them unfit for applications on low power and

resource-limited devices in m-health applications.

In order to fill this gap, there is a trend in the

design of lightweight deep learning models which

are tailored for edge platforms. The goal of these

models is to deliver a high level of accuracy, with the

minimum resource consumption, enabling reliable

performance without relying on a cloud connection.

Model reduction techniques, including but not

limiting to model pruning, quantization, knowledge

distillation, neural architecture search, are leveraged

to compress and optimize the deep learning

architectures so as to support real-time health

analytics on edge devices.

In this context, this work investigates the design

and implementation of lightweight models for

monitoring essential health parameters such as the

heart rate, respiratory rate, body temperature, and

physical activity levels. Through edge computing and

custom deep learning methods, the introduced

system offers an effective platform for contactless

Boddepalli, E., Sindhuja, K., Akila, K., Dharani, M., Nagarjunarao, G. and K., A.

Lightweight Deep Learning for Real-Time Health Monitoring on Edge Devices.

DOI: 10.5220/0013862100004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

243-249

ISBN: 978-989-758-777-1

243

real-time, continuous, and automatic health

surveillance. Not only does this improve the quality

of patient care by the timely introduction of

interventions, but it supports broader public health

goals by facilitating scalable, distributed healthcare

management capabilities. Finally, this research could

help facilitate the development of healthcare edge

intelligence towards more responsive, ubiquitous,

and privacy-friendly mobile health.

2 PROBLEM STATEMENT

With the inevitable trend of an aging population and

the prevalence of chronic diseases, along with the

demand for long-term health monitoring, the use of

intelligent healthcare systems has been booming.

However, the existing health monitoring systems rely

extensively on cloud-based data processing and

analysis, which entail major disadvantages, including

but not limited to, latency, bandwidth overhead,

energy consumption, and possible threats to privacy

of data. Although deep learning has shown to be

extremely powerful to analyze health-related data, its

computational demand is usually beyond resource-

limited devices (e.g., smartphones, smartwatches and

other wearable sensors). In a conventional deep

neural network architecture, this makes real-time, on-

device processing infeasible.

Edge computing (EC) provides a potential

solution by moving computation to near the data

source, resulting in reduced response time and better

privacy. However, the task is to build deep-learning

models which are compact, power-efficient and at

the same time high accuracy when running on edge

devices. A lot of the literature opts to trade-off

accuracy for much smaller computation time or are

still either too heavy to run in real-time in edge

scenarios.

The demand for lightweight and task specific deep

learning (DL) models which can be deployed in

resource constrained mobile edge devices for the real-

time monitoring of vital health parameters is thus

highly desirable. These models have to be efficient to

run on hardware and also need to perform well for

diverse physiological conditions as well as user

profiles. Solving this challenge is critical to achieving

mobile health where health services are scalable,

responsive, and privacy-preserving, with the ability to

execute and generate results when cloud

communication is not constantly available.

3 LITERATURE SURVEY

The synergy f energy efficient lightweight deep

learning model with edge computing is a major

research trend towards the development of real-time

health monitoring solutions. Aminu et al. (2025)

presented a general overview of lightweight deep

learning-based model for edge devices that

emphasizes the emerging need to optimize models in

constrained environment but it does not include

concrete implementation for health applications.

Baciu et al. (2025) introduced a dual attestation

approach for privacypreserving on-device learning,

emphasizing the needs of the competing requirements

of privacy and performance in edge environment.

Likewise, Batool (2025) studied a 5G based remote

monitoring architecture, yet without real-world

deployment, the research gap in edge-oriented

validation remains to be filled in.

Generalization difficulties over the edges were 4

discussed by Loh et al. (2025), who argued that

hardware-accelerated deep learning is required, and

Mittal (2024) identified optimizations for object

detection, which present transferable principles for

biomedical signal processing. The work of Spicher et

al. (2021) demonstrated the feasibility of edge

computing for ECG analysis with textile sensors

publishing a paper with some discussions about

hardware integration but lacked diversity of data.

Rashid et al. (2021b) presented adaptive CNNs for

physical activity recognition as the first baseline for

signal adaptation for health-related applications.

Agarwal and Alam (2020) presented a lightweight

model for human activity recognition with the

limitation of the lack of datasets. More recently,

federated learning methods such as FedRolex

(Zhang & Liu, 2022) and efficient on-device training

architectures like Mercury (Zeng et al., 2021)

presented promising architectures for distributed

health analysis. CATE (Zhang & Yan, 2021) and

Distream (Zeng et al., 2020) also proposed

computation-aware architectures, but these are not

health specific.

Fang et al. (2018) and Zeng et al. (2017) on

compact and resource-aware visual recognition

systems which led to some re-formulations for

biosignal analysis. Saeb et al. (2015) combined

behavioral cues with depressive symptoms, further

validating the potential of mobile data for predictive

health. Wang et al. (2022) examined TinyML based

on vital signs presenting with practical benchmarks to

deploy size-efficient models on the edge. Similarly,

Ghosh et al. (2023) introduced a CNN based HR

predictor for embedded systems.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

244

Luo et al. (2021) focused on the classification of

respiratory signals obtained from wearables and Lee,

et al. (2023) introduced energy-aware arrhythmia

detection based on deep learning. Kumar and Chawla

(2022) tested smartphone- based activity recognition

as a backbone of many wearable-oriented systems.

Rahman et al. (2021) focused on the unreliably

computational burden of diagnostic imaging models,

where a need for compression was more specifically

improved through Hassan et al. (2023) by an

approach of neural compression.

Chen et al. (2024) proposed TinyModelNet, an

optimized miniature model for edge-driven medical

applications. Shafique et al. (2021) provided specific

IoMT applications but without edge-level

optimization. Finally, Roy et al. (2023) described an

ECG classification model specifically designed for

wearables and corroborate the need for edge

computing with a more advanced signal processing.

This literature body presents good ground in

lightweight model construction and edge

computation, but highlights the necessity of domain-

specific, health-oriented, real-time solutions that can

work in different and resource-constrained

environments.

4 METHODOLOGY

Figure 1: Workflow for Lightweight Deep Learning-Based

Health Monitoring on Edge Devices.

Parameter monitoring on edge accessible mobile

devices. The latter considers model efficiency,

hardware accessibillity, and signal-tailored fine-

tuning for the goal of combining low-latency and

high-accuracy in resource-limited operation

conditions. Figure 1 shows the Workflow for

Lightweight Deep Learning-Based Health

Monitoring on Edge Devices.

The whole procedure has its first step as the

sensing of physiological signals from wearable

and/or mobile sensors, which collect information

such as heart rate, respiration rate, temperature, and

motion signals. Data are collected under different

environmental conditions and user depenency to get

diversity and robustness. The stages include pre-

processing steps such as noise filtering,

normalization and segmentation to help improve the

signal and ensure consistency across multiple sources

and hardware.

Table 1: Sensor Specifications for Health Parameter Acquisition.

Senso

Type

Measure

Paramete

Sampli

ng Rate

Accura

Power

Consumption

Optic

PPG

Senso

Heart

Rate

100 Hz

±2 bpm

Low

Ther

misto

Body

Tempera

ture

10 Hz

±0.2 °C

Very Low

Accel

erom

eter

(3-

axis)

Activity/

Posture

50 Hz

±0.05 g

Low

Micr

opho

Array

Respirat

ory

Signals

44.1

kHz

Variabl

Moderate

The learned model from the process are used as

pre-trained features to train a lean deep learning

model. Rather than making use of heavy traditional

models, this application leverages lightweight

architectures in the form of MobileNet, TinyML and

EfficientNet-elite variants. In addition, the model

complexity is reduced using model compression

techniques. Some of the techniques to reduce the size

of DNNs are by pruning redundant network

parameters, quantizing the weights of the model to

lower bit representation, and by using knowledge

distillation to train a smaller student model using a

larger higher performing teacher model. These two

optimizations greatly reduce memory size and

computation, with the same model accuracy. Table 1

shows the Sensor Specifications for Health Parameter

Acquisition.

Lightweight Deep Learning for Real-Time Health Monitoring on Edge Devices

245

Figure 2: User Interface of Mobile Health Monitoring

Application.

The models are optimized for real-time

performance by on-device inference with

TensorFlow Lite and PyTorch Mobile. These

frameworks convert the models to formats suitable

for fast execution on mobile CPUs and NPUs. For a

reality check, we run edge deployment on popular

devices, such as Raspberry Pi, NVIDIA Jetson Nano,

as well as Android-based phones. Performance under

varying conditions and loads is observed through

energy profiling and latency measurements.

Furthermore, in order to increase personalization

and reduce overfitting, they used transfer learning

techniques. pre-trained models may be fine-tuned on

small portions of user specific data to further adapt

the model to an individual's bio-patterns. This results

in increased accuracy without retraining a lot, which

is in line with the real-time needs of edge

environments. Figure 2 shows the User Interface of

Mobile Health Monitoring Application.

Table 2: Lightweight Model Architecture Details.

Model

Name

Parameter

(Millions)

Siz

Accur

acy

(%)

Inference

Time

(ms)

Mobile

NetV2

2.2

5.8

92.1

TinyML

-CNN

0.9

2.1

90.4

Efficient

Net-Lite

3.9

9.2

94.3

Compre

ssed-

LSTM

1.5

3.5

91.7

The proposed approach also incorporates a

federated learning setting for privacy preservation.

Data stays on the edge device and only the model

updates are uploaded to a central server for

aggregation. By doing so, we avoid the requirement

of transmitting privacy-sensitive and detailed health

data to cloud servers, thereby reducing privacy threats

while retaining collective learning gains. Table 2

shows the Lightweight Model Architecture Details.

Performance is assessed through extensive testing

on benchmark and streaming data. Evaluation The

effectiveness of the proposed method is measured

using accuracy, precision, recall, F1-score, inference

time, model size and energy consumed during

process. These are being evaluated in static and

ambulant user conditions to test robustness across

use cases.

Last stage of the method includes the adaptation

with the mobile health monitoring app receiving

real-time parameter output, alerts and trend analysis.

Both health clinicians and general users can use this

application to, in an internet-free and cloud

computation-free context, keep track of

physiological conditions.

By doing so, the paper provides a complete

solution not only for the common issues in cloud-

based health monitoring systems, but also creates a

basis for the scalable, intelligent, and autonomous

edge-driven mobile health.

5 RESULTS AND DISCUSSION

The deployment of computationally inexpensive deep

learning models for human health parameters

monitoring using edge devices was efficient in terms

of performance. When implementing the proposed

models on mobile devices like the Raspberry Pi 4,

Jetson Nano and Android phones, we observed that

the network models could be used for real-time

processing with low latency and reasonable energy

consumption. The inference time for most health

signals was less than 100 ms including heart rate and

respiration rate, indicating the models are appropriate

in real-time monitoring scene where consuming little

system resources. Figure 3 shows the Trade-Off

Between Model Accuracy and Size.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

246

Figure 3: Trade-Off Between Model Accuracy and Size.

Performance tests with multiple test

authentication sets indicated that the optimised

lightweight models performed well in comparison to

much larger, cloud-based architectures. For example,

the pruned MobileNet variant achieved an accuracy

greater than 92% in both heart rate classification and

anomaly detection, and the TinyML models gained

consistent precision and recall of over 90% on

different vital sign datasets. These results

demonstrate of the promise of deep learning on

resource-limited hardware, in particular when

coupled with signal-specific preprocessing and

training. Table 3 shows the Model Optimization

Techniques and Results.

Table 3: Model Optimization Techniques and Results.

Technique

Size

Reduction

(%)

Accuracy

Loss (%)

Energy

Savings

(%)

Pruning

1.2

Quantizati

on (INT8)

2.1

Distillatio

0.5

Combined

Optimizati

2.6

The quantisation and pruning methods performed

during the optimisation process greatly shrank the

model size, with some of the network becoming more

than 70% smaller without large reduction in

prediction performance. This method of knowledge

distillation also made student models more efficient,

which is particularly important for applications that

demand real-time inference, such as ambulatory

monitoring or activity tracking. They deployed on

the edge in an efficient manner using TensorFlow

Lite and PyTorch Mobile, to be compatible with

various hardware sets and OSs.

In terms of usability, these models can be easily

delivered to mobile applications to be integrated in

real-life health-monitoring systems. The realtime

parameter outputs are further shown on the mobile

application based on the research with graphics,

alerts, trend graphs, and conditions summaries, to

help Users as well as Caregivers manage the health in

proactive way. This feature served to illustrate actual

applications of the decentralized, low-latency

monitoring system in personal healthcare arenas.

Table 4 shows the Device-Wise Deployment

Performance.

Table 4: Device-Wise Deployment Performance.

Device

Name

Inference Time

(ms)

Battery Impact

(mAh/hour)

Temperature Rise (°C)

Real-Time

Capability

Raspberry Pi

+3.5

Yes

Jetson Nano

+2.8

Yes

Android

Smartphone

+2.2

Yes

ESP32

(TinyML)

107

+1.5

Partial

Lightweight Deep Learning for Real-Time Health Monitoring on Edge Devices

247

Figure 4: Device-Level Performance Evaluation.

The integrated transfer learning enabled the

system to adjust to personal biological variations, and

the accuracy of individualized monitoring was

significantly improved even without a large dataset of

personal data for training. In addition, federated

learning approach was shown to provide

collaborative model updates across decentralized

devices and to protect user privacy, successfully. This

demonstrates the system’s feasibility in practical

deployments with large scale and privacy concerns.

Figure 4 shows the Device-Level Performance

Evaluation.

Figure 5: Confusion Matrix of Heart Rate Classification

Results.

Energy profiling results showed that the models

consumed less power compared with conventional

cloud-based ones, which is beneficial to the battery

life of wearable devices and long-term use. This is

especially advantageous for members in rural or

underprivileged areas who may have restricted

availability of uninterrupted internet coverage or

steady electricity. Figure 5 shows the Confusion

Matrix of Heart Rate Classification Results.

Table 5: Evaluation Metrics for Health Parameter Monitoring

Models.

Parameter

Precision

(%)

Recall

(%)

F1-

Score

(%)

Accurac

y (%)

Heart Rate

94.2

93.6

93.9

94.1

Respiratory

Rate

92.1

91.8

91.9

92.0

Temperatur

96.3

95.5

95.9

96.0

Activity

State

91.7

92.6

92.1

91.9

In summary, the results support that the proposed

framework is effective to bridge the performance gap

between high-performance deep learning and edge

environments. Tackling the issues of latency, energy

efficiency, model size and privacy aspect, the work

provides a feasible solution to the emerging mobile

health monitoring systems of the future. The

conversation confirms that human pose deep learning,

when properly optimized and integrated, can

proactively alter the way that health metrics are

monitored, assessed and used in real time at low cost

and without compromising performance or ease of

use. Table 5 shows the Evaluation Metrics for Health

Parameter Monitoring Models.

6 CONCLUSIONS

This work presents a successful use-case on how

lightweight deep learning models can be efficiently

tuned and deployed on edge mobile platforms to

achieve accurate real-time health parameter

monitoring. This framework tuackles the

computation and energy constraints of wear-able and

mobile devices and provides a practical solution in

contrast to classical cloud-based healthcare. Through

model compression, edge optimizations, and

customized learning methods, the system retains

high accuracy while achieving drastically lower

latency and power. Advancements in federated

learning have made this suitable for deployment in

sensitive healthcare settings where privacy is a

concern. Experimental findings demonstrate the

models functioning efficiently under real-life

conditions, allowing efficient monitoring for vital

signs including heart rate, respiration, and body

movement. The tuning capabilities for individual

physiological trajectory only increase the system

reliability in continuous and long-term monitoring.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

248

In summary, this research presents a scalable,

privacy-preserving, resource efficient solution for

emerging requirements of m-health, and serves as a

stepping stone toward intelligent, edge-enabled

health systems of the future.

REFERENCES

Agarwal, P., & Alam, M. (2020). A lightweight deep learn-

ing model for human activity recognition on edge de-

vices. Procedia Computer Science, 167, 2360–2369.

https://doi.org/10.1016/j.procs.2020.03.289

Aminu, M., Kakudi, H. A., Hassan, M., Hamada, M., Umar,

U., & Salisu, M. L. (2025). Lightweight deep learning

models for edge devices—A survey. International

Journal of Computer Information Systems and Indus-

trial Management Applications, 17, 18.

https://doi.org/10.70917/ijcisim-2025-0014

Baciu, V.-E., Braeken, A., Segers, L., & Silva, B. d. (2025).

Secure tiny machine learning on edge devices: A light-

weight dual attestation mechanism. Future Internet,

17(2), 85. https://doi.org/10.3390/fi17020085

Batool, I. (2025). Real-time health monitoring using 5G

networks: A deep learning-based architecture for re-

mote patient care. arXiv.

https://arxiv.org/abs/2501.01027

Chen, C., Zhang, Y., & Zhou, Y. (2024). TinyModelNet: A

framework for neural network compression on edge

healthcare devices. IEEE Internet of Things Journal,

11(3), 2431–2442.

https://doi.org/10.1109/JIOT.2023.3321457

Fang, B., Zeng, X., & Zhang, M. (2018). NestDNN: Re-

source-aware on-device deep learning. In MobiCom

(pp. 115–127).

https://doi.org/10.1145/3241539.3241547

Ghosh, S., Banerjee, A., & Mitra, S. (2023). Lightweight

CNN for on-device heart rate prediction using PPG sig-

nals. Biomedical Signal Processing and Control, 81,

104412. https://doi.org/10.1016/j.bspc.2022.104412

Hassan, A., Malik, H., & Kim, D. (2023). Lightweight neu-

ral network compression for wearable health monitor-

ing. Sensors, 23(2), 523.

https://doi.org/10.3390/s23020523

Kumar, M., & Chawla, P. (2022). Deep learning-based hu-

man activity recognition for healthcare using mobile

sensors. Journal of Ambient Intelligence and Human-

ized Computing, 13, 829–840.

https://doi.org/10.1007/s12652-021-03046-6

Lee, J., Kim, D., & Yoo, H. (2023). Ultra-low power CNNs

for real-time arrhythmia detection on mobile devices.

IEEE Transactions on Biomedical Circuits and Sys-

tems, 17(1), 45–56.

https://doi.org/10.1109/TBCAS.2022.3226687

Loh, J., Dudchenko, L., Viga, J., & Gemmeke, T. (2025).

Towards hardware supported domain generalization in

DNN-based edge computing devices for health moni-

toring. arXiv. https://arxiv.org/abs/2503.09661

Luo, Y., Zhang, Z., & Chen, L. (2021). Real-time respira-

tory monitoring using wearable sensors and deep learn-

ing. Sensors, 21(5), 1809.

https://doi.org/10.3390/s21051809

Mittal, P. (2024). A comprehensive survey of deep learn-

ing-based lightweight object detection models for edge

devices. Artificial Intelligence Review, 57, Article 242.

https://doi.org/10.1007/s10462-024-10877-1

Rahman, M. M., Chowdhury, M. E. H., & Khandakar, A.

(2021). A survey on deep learning in respiratory analy-

sis using chest X-ray and CT images. Computers in Bi-

ology and Medicine, 132, 104306.

https://doi.org/10.1016/j.compbiomed.2021.104306

Rashid, N., Demirel, B. U., & Al Faruque, M. A. (2021).

AHAR: Adaptive CNN for energy-efficient human ac-

tivity recognition on edge. arXiv.

https://arxiv.org/abs/2102.01875

Roy, D., Sinha, R., & Saha, S. (2023). Efficient edge intel-

ligence for wearable ECG signal classification. IEEE

Access, 11, 23654–23666. https://doi.org/10.1109/AC-

CESS.2023.3241083

Saeb, S., Zhang, M., Karr, C. J., Schueller, S. M., Corden,

M. E., Kording, K. P., & Mohr, D. C. (2015). Mobile

sensor correlates of depression. JMIR, 17(7), e175.

https://doi.org/10.2196/jmir.4273

Shafique, M., Khawaja, B. A., Sabir, F., Qaisar, S. B., &

Mustaqim, M. M. (2021). Internet of Medical Things

(IoMT): Applications and benefits. Journal of Commu-

nications and Networks, 23(2), 126–137.

https://doi.org/10.23919/JCN.2021.000006

Spicher, N., Klingenberg, A., Purrucker, V., & Deserno, T.

M. (2021). Edge computing in 5G cellular networks for

real-time ECG analysis with textile sensors. arXiv.

https://arxiv.org/abs/2107.13767

Wang, X., Liu, C., & Hu, J. (2022). TinyML in healthcare:

Deploying machine learning on edge devices for vital

sign monitoring. IEEE Access, 10, 15823–15836.

https://doi.org/10.1109/ACCESS.2022.3149057

Zeng, X., Cao, K., & Zhang, M. (2017). MobileDeepPill:

Recognizing pill images with deep learning. In

MobiSys (pp. 56–67).

https://doi.org/10.1145/3081333.3081365

Zeng, X., Fang, B., & Zhang, M. (2020). Distream: Adap-

tive distributed edge intelligence for video. In ACM

SenSys (pp. 1–14).

https://doi.org/10.1145/3384419.3430786

Zeng, X., Yan, M., & Zhang, M. (2021). Mercury: Efficient

on-device distributed DNN training. In ACM SenSys

(pp. 1–14). https://doi.org/10.1145/3485730.3485947

Zhang, M., & Yan, S. (2021). CATE: Computation-aware

architecture encoding with transformers. In ICML

2021. https://proceedings.mlr.press/v139/yan21a.html

Zhang, M., & Liu, L. (2022). FedRolex: Model-heteroge-

neous federated learning with rolling sub-model extrac-

tion. In NeurIPS 2022. https://proceedings.neu-

rips.cc/paper_files/paper/2022/hash/4e8c5a8d.html

Lightweight Deep Learning for Real-Time Health Monitoring on Edge Devices

249