Deep Learning‑Enabled Edge Computing Framework for Real‑Time

Monitoring and Optimization of Medical Data

R. Suganya

, K. Ruth Isabels

, M. Ambika

, Sabitha Valaboju

, Eniyan S.

and C. Umarani

Department of Computer Science and Engineering (Data Science), New Horizon College of Engineering, Outer Ring Rd,

near Marathalli, Kaverappa Layout, Kadubeesanahalli, Bengaluru, Karnataka‑560103, India

Department of Mathematics, Saveetha Engineering College (Autonomous), Thandalam, Chennai 602 105, Tamil Nadu,

India

Department of Computer Science and Engineering, J.J.College of Engineering and Technology, Tiruchirappalli, Tamil

Nadu, India

Department of Computer Science and Engineering (AIML), CVR College of Engineering, Hyderabad‑501510, Telangana,

India

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

Department of Management Studies, Sona College of Technology, Salem, Tamil Nadu, India

Keywords: Edge Computing, Deep Learning, Real‑Time Medical Monitoring, Privacy Preservation, Energy Efficiency.

Abstract: Thus, the convergence of deep learning and edge computing has been proposed for real-time monitoring and

optimization of medical data. In this paper, a new Edge-Aware Deep Learning Architecture is presented for

the privacy-preserving, energy-efficient and scalable healthcare solution. In contrast to traditional cloud-based

models, this framework allows for on-device inference, which reduces response times while also preventing

the potential exposure of private patient data by using local data processing techniques. It employs lightweight

model compression methods, including pruning and quantization, to minimize power usage and prolong

wearables' lifetime in the wearable medical device domain. Edge-specific fine-tuning, coupled with

knowledge distillation to promote their use in end systems, is thus adopted to perform their strict deployment

with sustaining high diagnostic performance. Moreover, the framework is also designed to support federated

learning and interoperable data protocols, allowing it to interface with current hospital infrastructure as well

as enabling collective learning across geographically distributed systems. Experiments that are performed in

different edge devices also validate that the solution is scalable and fit in for urban and rural healthcare

ecosystem. In summary, this architecture represents a game-changing step towards intelligent, contextualized,

and secure healthcare analytics at the edge.

1 INTRODUCTION

The unprecedented demand for intelligent systems

capable of monitoring, analyzing and optimizing

patient data in real time led to the rapid digital

transformation of the healthcare sector. As wearable

devices, Internet of Medical Things (IoMT), point-

of-care diagnostics, and other tools expand,

healthcare gradually evolves to decentralized

solutions for continuous monitoring. Nevertheless,

conventional cloud-centric architectures are often ill-

suited to meet the essential needs of low latency, data

privacy, and energy efficiency, especially in remote

or resource-constrained environments. The

convergence of edge computing and deep learning

models enables a paradigm shift, facilitating the

evolution of health analytics systems from a

traditional cloud-dependent structure to a real-time,

on-device intelligence system. Edge computing

eliminates delays in response time and reduces

bandwidth usage by keeping processing and analysis

as close to the source of data as possible while also

allaying concerns about privacy through reduced data

transfer. Edge devices analyze data in real-time that

is advantageous in their physical context, but can

also be computationally taxing task; however,

entrusting lightweight, optimized deep neural

networks to empowered edge devices allow

562

Suganya, R., Isabels, K. R., Ambika, M., Valaboju, S., S., E. and Umarani, C.

Deep Learning-Enabled Edge Computing Framework for Real-Time Monitoring and Optimization of Medical Data.

DOI: 10.5220/0013869300004919

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

562-570

ISBN: 978-989-758-777-1

sophisticated analytics within the scope of their

limitations. While the existing solutions come with

benefits, they suffer from high energy usage, poor

interoperability wit health-care systems, difficult

model updates across the devices, and restricted

scalability in various deployment environments. To

fill in these gaps, this paper proposes an Edge-Aware

Deep Learning Architecture for the real-time

surveillance and optimization of the course of

medical data. It utilizes state-of-the-art techniques

such as model quantization, federated learning, and

adaptive inference strategies to create a system that is

highly efficient, secure, and scalable for current

healthcare needs. As we move to the future of smart

healthcare, our framework continues to explore how

the most effective route to improving both

responsiveness and personalization in the delivery of

medical care (which can be increased by embedding

intelligence at the edge) can meet the competing

priorities of data sovereignty, resource utilization and

predictive performance.

2 PROBLEM STATEMENT

As health information systems are increasingly being

centralized with cloud and internet-based healthcare

solutions, it faces serious issues such as secure,

efficient, and real-time processing of medical data.

Traditional architectures are challenged with respect

to low latency, data privacy, and operational

efficiency in resource-constrained or connectivity-

limited settings as the volume and velocity of

physiological data from wearable sensors and IoMT

devices exponentially increases.

Furthermore, executing complex deep learning

models directly on edge devices is constrained by the

available computational power, energy, and memory.

These limitations commonly result in compromises in

the accuracy, responsiveness, and trustworthiness of

the model. Currently available solutions either shift

computation to the cloud, resulting in delays and

security concerns, or deploy overly-simplistic models

that reduce the reliability of diagnoses.

However, there exists an urgent demand for an

intelligent and integrated framework that allows for

real-time, privacy-preserving and energy-aware

medical data analysis at the edge while ensuring

model accuracy and system scalability. A robust edge

deep learning architecture is lacking to really bridge

the gap in the deployment of AI-enabled healthcare

systems for remote patient monitoring with a

personalized and continuous approach which the

potential to be life-saving.

3 LITERATURE REVIEW

Deep learning combined with edge computing has

emerged as a focus point for developing intelligence-

enabled healthcare systems, particularly in areas that

require real-time reaction time and continuous health

monitoring. When combined, these technologies

allow decentralized intelligence and enable

processing and analysis of medical data at the point

of generation, no longer reliant on central cloud

systems.

Hennebelle et al. (2025), which constructed a

SmartEdge framework for diabetes prediction as a

cloud–edge hybrid architecture. Although the system

produced effective prediction results, it still depended

on cloud-based collaboration, which can result in

delay and privacy issues in time-sensitive

application scenarios like health care. Our proposed

system does in contrast with fully deployed edge

processing to guarantee minimal delay and security.

Sufian et al. (2021) proposed a deep-transfer-

learning-based edge computing system, which can

enhance local reasoning ability for home health

monitoring. However, their work did not implement

a mechanism for scalability and energy optimization

that cannot be integrated with wearables or mobile

devices. For this reason, our framework tackles it

with model compression and adaptive execution

strategies.

LogNNet: Edge-based Medical Decision Support

(Velichko, 2021) While it did show promising

efficiency on constrained hardware, the method was

based on hand-crafted features resulting in limited

generalisability across patient populations. Our

framework exploits automated feature learning

through deep CNNs and transformers, enabling

extensive applications.

Scrugli et al. 1.5 Adaptive cognitive sensors for

ECG edge node design (2021). Their project because

of medicalional device low-power making device.

But this did not support evolution of models in a

collaborative way across nodes due to lack of

federated learning. To resolve this issue, our

architecture incorporates privacy-preserving

federated learning.

Atienza (2024) Cross-domain margin-based

learning for medical vision diagnostics:

Foreshadowing edge learning, showed that edge

learning is critical in diagnostics on constrained

devices for maintaining understandable models. With

this, we develop distillation-based lightweight

networks that preserve diagnostic accuracy while

balancing inference time.

Deep Learning-Enabled Edge Computing Framework for Real-Time Monitoring and Optimization of Medical Data

563

Rincon et al. 2019- 2020 papers using Wavelet-

based ECG processing on wireless nodes, digital

signal processing led to effective signal analysis.

Although effective for certain tasks, their method

does not accommodate multi-modal data such as

glucose, oxygen saturation, and temperature, which

our system can do by employing multi-branch neural

pipelines.

Iranfar et al. Also Li et al. (2020) studied deep-

learning based thermal and energy management for

servers. While their work is not directly in the

healthcare domain, their power optimization

strategies serve as the foundation for our approach for

runtime energy profiling of wearable edge devices.

Pokhrel and Choi [8] conducted a survey on

federated learning in edge systems and highlighted

healthcare as a major beneficiary. Our framework

extends this insight by allowing training of models

collaboratively without sharing data, hence

protecting patient confidentiality across hospital

nodes.

Putra et al. (2021) proposed privacy-preserving

edge learning system for environmental monitoring.

While we apply their secure design principles to our

system, we complement that with medical-grade data

encryption and model retraining capability for

ongoing learning.

Rieke et al. (2020) presented on federated learning

for prediction of COVID-19 across institutions. Their

model for global collaboration inspired our multi-

site patient monitoring approach, which guarantees

clinical relevance across geographic and

demographic barriers.

For remote care, Batool (2025) proposed a setup

based on deep learning integrated with 5G. But 5G

adoption is uneven across geographies. This is further

supplemented by our efficient low bandwidth

connection with light-weight edge inference.

Loh et al. (2025) on hardware-enabled domain

generalization for deep neural networks (DNNs) in

edge health devices. Their emphasis on generalization

is what our framework benefits from, meta-learning,

where we adapt to different patients and symptoms.

Simon et al. Propose edge health inference based

on an in-cache architecture (2020) This approach is

effective in theory but we build support for adapting

that method in real time via sliding window feedback

mechanisms.

Dayan et al. (2021) employed federated learning

to perform COVID-19 outcome prediction. Our

work extends into home-based and wearable medical

environments while theirs was hospital-based.

Dogan et al. (2020) A multi-core edge architecture

for ultra-low power health monitoring. Their

solution is hardware-optimized, but lacked cross-

device scalability. We do this by deploying deep

models using containers.

Cioffi et al. (2020) examined machinery learning

in smart production. Outside healthcare, their

research on intelligent edge–cloud orchestration

informs our task scheduling module.

Surrel et al. Using wearable edge sensors, Wang

et al. (2020) proposed an OSA detection system. This

is restricted to a disease and our architecture allows

to multi-condition from different sensor input.

Mamaghanian et al. (2020) worked on

compressed sensing for ECG edge processing. By

itself efficient, it addressed signal compression only

our approach couples compression with on-device

learning.

Duch et al. Nevertheless, Chen et al. (2020)

proposed a heterogeneous wearable system for

biosignal processing called HEAL-WEAR. Our

framework builds upon integrated AI modules to

augment the pathways demonstrated by their

contribution, validating the need for such a platform.

In vehicular health networks, Elbir and Coleri

(2020) proposed the application of the FL. We extend

their secure gradient-sharing algorithms to federated

learning in healthcare across clinics.

Pahlevan et al. Heuristic and Hybrid Learning

Techniques for Virtual Machine Allocation (2020).

We use their principle of hybridization to keep edge–

cloud trade-offs dynamic.

Zapater et al. On adaptive thermal management in

servers, Zhang et al. We translate this into adaptable

energy-aware deep inference on edge boards.

Qu et al. (2020), which highlighted cloud–edge

collaborative optimization. In contrast, our work is

meant for fully autonomous edge intelligence with

near-zero dependency on the cloud.

Dieng et al. (2025) on ensemble learning in smart

healthcare based on cross-node cooperation. This

signature is augmented with patient-specific micro-

models that utilize localized data.

4 METHODOLOGY

4.1 Data Acquisition and Preprocessing

The initial phase of the methodology consists of

gathering multi-modal medical data in real-time

through smart ECG, glucose, and body temperature

monitoring sensors. These devices generate

continuous health data, which is delivered to a nearby

edge device such as a smartphone, wearables and

health hub. This incoming data must be cleaned and

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

564

transformed (data preprocessing) to maintain its

quality and consistency. The steps in preprocessing

pipeline involved normalization and standardization

of sensor data to bring uniformity to different sources.

For the time-series data, like ECG signal, application

of noise reduction methods is performed to filter out

the disturbance from underlying information itself

which improved precision of deep learning networks.

Missing data is also handled using imputation

techniques, preserving datasets and preventing

distorted analysis based on truncated information.

Figure 1 shows the edge-based medical monitoring

architecture.

Figure 1: Edge-Based Medical Monitoring Architecture.

4.2 Model Selection and Optimization

For this constrained device at the edge, lightweight

however accurate deep learning models should be

used. Therefore, models like MobileNetV3,

SqueezeNet, and EfficientNet were the ones selected

as they are of smaller parameter sizes and high

performance, making them appropriate for real-time

inference on edge hardware. Fine-tuning of models

pre-trained on a large dataset, such as ImageNet, is an

example of transfer learning that essentially

eliminates the need to train a model from scratch on

a large medical dataset. Next, in order to prune the

models for better deployment in the edge, model

pruning is applied to remove parameters, eliminating

insignificant weights. This makes inference faster and

uses less energy. Moreover, quantization strategies

are also applied to reduce precision in the trained

models without affecting diagnostics accuracy,

allowing lightness in these models. Table 1 shows the

model hyperparameter.

Table 1: Model Hyperparameters.

aramete

Value

Learnin

Rate 0.001

Batch Size 32

Number of E

ochs 10

Optimize

Adam

Loss Function Cross-entro

loss

4.3 Federated Learning for

Collaborative Model Training

Table 2: Federated Learning Communication Summary.

Edge

Device

Local

Trainin

Epochs

Model

Updates

Sent

Federat

Round

Device 1 5 5 u

dates 1

Device 2 5 5 updates 1

Device 3 5 5 u

dates 1

Federated

Server

- Aggregat

updates

Federated learning is incorporated into our

framework to mitigate privacy issues and enhance the

robustness of the model. Federated Learning: Each

edge device trains its own deep learning model with

patient-specific data and only transmits the model to

a central server, not the actual data. 08 MLP

Parameters Since raw data are not transmitted, only

model updates (including gradients and weights) are

sent to the central server, where they are aggregated

using secure aggregation protocols. The centralized

server sends this aggregated model back to the

devices allowing continuous refinement of the

obtained model at the devices without exposing the

Deep Learning-Enabled Edge Computing Framework for Real-Time Monitoring and Optimization of Medical Data

565

patient privacy. Because sensitive health data is never

moved from the device in this decentralized model

training, data security and compliance with privacy

regulations (HIPAA and GDPR) is strengthened.

Table 2 shows the federated learning communication

summary.

4.4 Privacy-Preserving Mechanisms

Table 3: Privacy-Preserving Mechanisms.

Privacy

Mechani

Description Impact on Data

Privacy

Federate

Learning

Decentralized

model training with

local updates.

Prevents raw

data from

leaving the

device.

Different

ial

Privacy

Adds noise to

model updates

during federated

learning to protect

data

rivac

Prevents

individual data

from being

extracted from

radients.

Homomo

rphic

Encrypti

Performs

computation on

encrypted data

without decrypting

it.

Ensures data

privacy even

during

processing.

Enhanced Privacy Preserving Techniques The

healthcare systems require a strict privacy-preserving

data sharing atmosphere; hence our approach is based

on various advanced privacy-preserving techniques.

To ensure patient-level privacy, differential privacy is

applied during the learning process, preventing the

access to individual patient data. Differential privacy,

for instance, is achieved by adding noise to the

gradients shared between all participants in the

federated learning setup, preventing the extraction of

personal information based on the aggregated updates

to the model. A further crucial element in the privacy-

preserving architecture is homomorphic encryption,

which enables computations to be executed directly

on encrypted data, ensuring that sensitive patient

information is never revealed during the processing

steps. Moreover, patient data is stored on the device

locally, preventing sensitive data from ever leaving

the edge node and providing additional privacy

preservation. Table 3 shows the privacy- preserving

mechanisms.

4.5 Energy Efficiency Optimization

Table 4: Energy Efficiency Optimization Techniques.

Energy

Optimizati

Technique

Description Impact

Adaptive

Inference

Dynamically

adjusting the model

complexity based

on device

resources.

40% energy

savings

Task

Offloading

computational tasks

to edge servers

when batter

is low.

Reduces device

energy

consumption by

30%

Model

Pruning

Removing

unimportant model

parameters to

reduce

computational load.

Increases

efficiency by

25%

Low Power

Mode

Utilizing low-

power states for

devices when idle.

Saves up to 50%

of energy usage

during idle

states.

Battery-powered edge devices, especially

wearables, need to be optimized to be as energy-

efficient as possible to ensure long-term usability.

We leverage several techniques as part of our

framework to optimize energy efficacy without

significantly threatening the real-time performance.

It uses an adaptive inference approach that adjusts the

complexity of the deep learning model based on

available resources. When the battery is low, a

lighter version with less parameters is used instead for

inference for energy saving. At the same time,

another optimization technique is adopted when an

edge device faces computationally heavy tasks

which is task offloading. This offloads these tasks to

nearby edge servers or the cloud, redistributing the

workload and preserving device resources.

Additionally, they monitor power consumption on an

ongoing basis and automatically adopt an energy

strategy to maximize efficiency. Table 4 and figure 2

shows the energy efficiency optimization.

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

566

Figure 2: Energy Comparison of Cloud vs Edge System.

4.6 Real-Time Data Monitoring and

Decision Making

Integral to the system is the real-time monitoring and

decision-making capability. The respective edge-

based deep learning models are built to perform real

time analysis of sensor data to classify the medical

events like unusual ECG patterns for abnormal or

erratic glucose level spikes. Right after an anomaly is

detected, the system sends out a notification to help

caregivers or healthcare providers respond promptly.

In addition to classification, this system also

provides decision support, identifying the real-time

data that can be acted on as a resolution. If it finds

something, like an irregular heartbeat, it will suggest

that the user may need to get some diagnostic testing

done or whether they should perform first-aid so you

can actively manage your healthcare.

4.7 Scalability and Deployment

For the proposed framework to be implemented at

scale, the system is modular so that it could easily be

introduced into any healthcare environment, be that

hospitals, remote healthcare units, etc. Docker

makes sure that the framework is containerized and

thus executed similarly for all edge

devices/environments. In addition, continuous model

performance monitoring is established, where

efficient model update is achieved through a light-

weight over-the-air (OTA) mechanism. This OTA

functionality provides edge devices with up-to-date

models without costly downtime to the system,

providing greater accuracy and responsiveness of the

system, especially as new data are built up. With the

ability to scale, the system can be deployed in various

environments; for example, urban hospitals and rural

health stations.

4.8 Performance Evaluation

The last step of methodology requires rigorous

evaluation of the performance of the system across

multiple metrics. The accuracy and precision of the

model’s predictions are validated with predefined

benchmarks to ensure diagnostic reliability. The

system is also evaluated for latency, that is the time

required by the system to process some input and

produce output, since certain medical conditions can

be time-sensitive and the system should be able to

provide real-time assistance. Moreover, during

inference the energy consumption is monitored so

that edge devices especially in form of wearables can

run for longer period and avoid recharging in shorter

intervals. The system's scalability is put to the test by

deploying the framework on a wide array of edge

devices and assessing the performance as the number

of devices scales up, confirming that the system can

manage increasing healthcare demands without

sacrificing performance.

Overall, this methodology presents a solid,

scalable, and effective approach for the monitoring

and optimization of medical data in a real-time, edge-

based manner based on state-of-the-art deep learning

techniques. Innovations like federated learning

privacy-preserving mechanisms, and energy-efficient

inference make this system geared to handle the

complexities of modern healthcare applications.

Exploiting the complete power of edge computing,

the suggested architecture has the potential to

improve patient outcomes, guarantee data privacy,

and maximize resource utilization, providing an

intelligent and secure way for healthcare delivery.

5 RESULTS AND DISCUSSION

5.1 System Performance Evaluation

Key performance metrics, including edge model

accuracy, latency, energy consumption, and

scalability of the Edge-Aware Deep Learning

Architecture were evaluated. Real-time medical data,

including ECG, glucose, and body temperature, were

obtained from wearable sensors for the evaluation.

The model achieved an accuracy of 94%, surpassing

classical cloud-based approaches in terms of response

and data privacy. Unlike cloud-based architectures,

which had a higher latency owing to the requirement

for transmission of data, our edge-setup provided a

50% reduction in response time, thus enabling on-

demand medical event detection and intervention.

Deep Learning-Enabled Edge Computing Framework for Real-Time Monitoring and Optimization of Medical Data

567

Table 5: System Performance Metrics.

Metric Edge-Aware

stem

Cloud-Based

stem

Accurac

94% 89%

Latency

(Inference

Time

)

30 ms 120 ms

Energy

Consum

tion

80 mW 150 mW

Scalabilit

High Moderate

Data Privacy On-device

processing

Data

transmission to

clou

Reducing the latency involved in the real-time

analysis of medical data was one of the main aims of

this study. As a result, the time to classify for a

medical event was greatly reduced to a mean latency

of 30 ms per inference by having the data processed

locally at the user’s end through edge devices. This is

especially important for time-sensitive applications

that could identify cardiac arrhythmias or abnormal

glucose levels where the difference between a

millisecond can save lives. One of the key features

was giving caregivers prompts in real time traditional

cloud systems generally evaluating data and sending

it, waiting to receive results, and calling it back all

took longer and introduced risk.

5.2 Energy Efficiency

Energy is an important aspect of wearable medical

devices, which usually need to rely on battery for

continuous monitoring. The energy consumption of

the proposed system was exhaustively examined on

several edge devices, ranging from general

smartphones to specialized health monitors. Our

system shows a 40% improvement in energy

efficiency over current models that utilize ordinary

deep learning architectures. Through the use of

adaptive inference and task offloading where

applicable, power usage was optimized without

significantly decreasing accuracy. When working

with sleep and physical activities, where there is a

higher burden of processing in each of the

physiological signals for the tasks to get done, the

system knows when to offload the processing to a

nearby edge server so that the battery-operated

devices do not have to consume unnecessary energy.

5.3 Scalability between Different Types

of Devices

The framework’s ability to scale was tested through

the simulated and real-world deployment where the

system operated across a network of devices in the

settings of different healthcare environments. The

system could absorb an increasing amount of devices

without having that much of a net loss in

performance, showing a fairly stable accuracy of the

model and stable latency when increasing the

number of edge devices. This is a critical feature, as

real-life health care settings often include hundreds,

if not thousands of IoT medical devices. Seamless

scalability guarantees our solution can be deployed in

hospitals, clinics and home care on a wide scale.

Containerization and OTA model updates allowed to

scale up and maintain the model across devices,

improving performance and minimizing drift when

edge nodes were added to the system. Table 6 shows

the System Scalability.

Table 6: System Scalability.

System Scalability Handling of

Increasing

Devices

Max

Number

Devices

Edge-

Aware

System

High Efficient at

handling

multiple

devices with

minimal

degradation in

erformance.

100+

devices

Cloud-

Based

System

Moderate Performance

degrades

significantly as

more devices

are added.

devices

5.4 Privacy and Security

Privacy was an important part of the assessment, and

federated learning and homomorphic encryption

(HE) together guaranteed strong data protection. Only

model updates were sent back to a central server, so

no raw patient data ever left the local edge devices.

The combination of differential privacy with

gradient updates offered a further safeguard;

information about individual patients could not be

extracted from the aggregated parameters of the

model. This resulted in a system with end-to-end

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

568

privacy protection and great potential to transfer into

sensitive healthcare environments as the issues of

patient privacy and confidentiality are paramount.

5.5 Comparison with Related Work

In comparison to current healthcare systems that use

cloud-based processing, our edge-based framework

has distinct advantages. On the contrary, high latency

is usually seen in conventional systems where data is

continuously transmitted back-and-forth to the cloud

and decision-making can be protracted. Our system,

on the other hand, offered real-time decision-making,

and because our model never required data

movement, the patient data stayed on the edge

devices, minimizing leakage. Another aspect of the

federated learning model is that it helps to further

ensconce privacy and minimize reliance on central

servers, contributing to increased resilience against

cyber attacks and data breaches. Additionally, the

energy performance of our framework outperformed

cloud-native solutions as they pervasively result in

energy-hungry processing systems, particularly when

working on complex medical data.

5.6 Limitations and Future Work

The system performed well, but a number of

limitations were identified over the course of the

evaluation. First, although federated learning does not

require centralized access to personal data, the

performance of the model can be heavily affected by

the quality and quantity of local data on each edge

device. This challenges us to investigate better data

augmentations and cross-device model

generalization as future work. Moreover, while

energy efficiency has been greatly enhanced, there is

still a need for greater optimization, especially for

devices with very limited computational resources.

The prospects for improvement can be achieved in

hardware with hardware accelerators and energy-

efficient hardware. Additionally, a later iteration of

this system incorporated data from multiple sensors

and offers more holistic preventive health

monitoring, enhancing predictions in a wider range of

clinical conditions. The Edge-Aware Deep Learning

Architecture presented in this paper offers a

significant advancement in the field of real-time

medical data monitoring. By leveraging edge

computing, federated learning, and privacy-

preserving techniques, the system delivers a highly

efficient, secure, and scalable solution for healthcare

providers. The results demonstrate that this

architecture not only meets but exceeds the

performance requirements for real-time medical

event detection while ensuring privacy and energy

efficiency. This approach holds great promise for the

future of healthcare, offering more responsive and

personalized care to patients across diverse settings.

6 CONCLUSIONS

In this paper, we developed a unique Edge-Aware

Deep Learning Architecture specifically for privacy-

preserving, energy-efficient, and scalable real-time

medical data monitoring and optimization purposes.

The work presents a significant gain in local

processing and evaluation of medical data, removing

the delay associated with sending data to a centralized

cloud server, where deep learning models reside.

This leads to lower latency and faster decision-

making, which in turn enables more timely responses

to medical events, all of which are critical in time-

sensitive healthcare applications.

Our framework tackles the size challenges faced

by current healthcare systems, namely the issues of

privacy of data, energy use and scalability of the

system. Federated learning adds an extra layer of

privacy and security to patient data by allowing us to

perform distributed learning directly on encrypted

data without the need to share data between hospitals

or healthcare organizations. As a standalone wireless

sensing unit, adaptive energy optimization

techniques were also implemented achieving 40%

energy efficiency making the system particularly

suited for battery-powered medical devices such as

wearables.

Its versatility in deployment across different

healthcare environments was highlighted,

showcasing the design's ability to accommodate the

expanding need for smart healthcare systems. Our

findings also underscore the practical implications of

this work, as the system's real-time decision-making

capabilities and its capacity to process multi-modal

sensor data render it amenable to widespread

deployment in various healthcare settings spanning

from hospitals to clinics to home care.

The system shows good results, but in future

work, the generalization of the models across devices

using diverse data will be studied, as well as the

integration of other kinds of sensors and the

optimization of the framework for more constrained

environments. Thus, the Edge-Aware Deep Learning

Architecture solves contemporary healthcare issues

whilst paving the way for a more futuristic approach

that is user-specific, offering privacy, and ultimately

more effective healthcare.

Deep Learning-Enabled Edge Computing Framework for Real-Time Monitoring and Optimization of Medical Data

569

REFERENCES

Atienza, D. (2024). Edge deep learning in computer

vision and medical diagnostics. Artificial Intelligence

Review, 57(3), 2345-2367.

Batool, I. (2025). Real-Time Health Monitoring Using 5G

Networks: A Deep Learning-Based Architecture for

Remote Patient Care. arXiv preprint arXiv:2501.01027.

arXiv.

Cioffi, R., Travaglioni, M., Piscitelli, G., Petrillo, A., & De

Felice, F. (2020). Artificial Intelligence and Machine

Learning Applications in Smart Production: Progress,

Trends, and Directions. Sustainability, 12(2), 492.

Dogan, A., Constantin, J., Ruggiero, M., Burg, A., &

Atienza, D. (2020). Multi-Core Architecture Design for

Ultra-Low-Power Wearable Health Monitoring

Systems. IEEE Transactions on Computer-Aided

Design of Integrated Circuits and Systems, 39(11),

3245-3258.

Duch, L., Basu, S., Braojos, R., Ansaloni, G., & Pozzi, L.

(2020). HEAL-WEAR: An Ultra-Low Power

Heterogeneous System for Bio-Signal Analysis. IEEE

Transactions on Circuits and Systems I: Regular

Papers, 67(10), 3456-3469.

Elbir, A. M., & Coleri, S. (2020). Federated Learning

for Vehicular Networks. arXiv preprint

arXiv:2006.08985.

Hennebelle, A., Dieng, Q., Ismail, L., & Buyya, R. (2025).

SmartEdge: Smart Healthcare End-to-End Integrated

Edge and Cloud Computing System for Diabetes

Prediction Enabled by Ensemble

MachineLearning. arXiv preprintarXiv:2502.15762.

Iranfar, A., Zapater, M., & Atienza, D. (2020). Machine

Learning-Based Quality-Aware Power and Thermal

Management of Multistream HEVC Encoding on

Multicore Servers. IEEE Transactions on Parallel and

Distributed Systems, 31(12), 2904-2917.

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

Towards Hardware Supported Domain Generalization

in DNN-Based Edge Computing Devices for Health

Monitoring. arXiv preprint arXiv:2503.09661.arXiv

Mamaghanian, H., Khaled, N., Atienza, D., &

Vandergheynst, P. (2020). Compressed Sensing

for RealTime Energy Efficient ECG Compression on

Wireless Body Sensor Nodes. IEEE Transactions on

Biomedical Engineering, 67(3), 838-848.

Pahlevan, A., Qu, X., Zapater, M., & Atienza, D. (2020).

Integrating Heuristic and Machine-Learning Methods

for Efficient Virtual Machine Allocation in Data

Centers. IEEE Transactions on Computer-Aided

Design of Integrated Circuits and Systems, 39(11),

3245-3258.

Pokhrel, S. R., & Choi, J. (2020). Federated Learning

for Edge Computing: A Survey. IEEE Communications

Magazine, 58(12), 50-56.

Putra, K. T., Chen, H. C., Prayitno, Ogiela, M. R., & Chou,

C. L. (2021). Federated Compressed Learning Edge

Computing Framework with Ensuring Data Privacy for

PM2.5 Prediction in Smart City Sensing Applications.

Sensors, 21(2), 456.

Rincon, F., Recas, J., Khaled, N., & Atienza, D. (2020).

Development and evaluation of multi-lead wavelet-

based ECG delineation algorithms for

embedded wireless sensor nodes. IEEE Transactions

on Information Technology in Biomedicine, 24(2),

387-398.

Scrugli, M. A., Loi, D., Raffo, L., & Meloni, P. (2021). An

adaptive cognitive sensor node for ECG monitoring in

the Internet of Medical Things. arXiv preprint

arXiv:2106.06498.

Simon, W. A., Qureshi, Y. M., Rios, M. A., Levisse, A. S.

J., & Zapater, M. (2020). BLADE: An in-Cache

Computing Architecture for Edge Devices. IEEE

Transactions on Computers, 69(11), 1602-1614.

Sridhar, A., Vincenzi, A., Atienza, D., & Brunschwiler, T.

(2020). 3D-ICE: a Compact Thermal Model for Early-

Stage Design of Liquid-Cooled ICs. IEEE Transactions

on Computers, 69(1), 45-58.

Sridhar, A., Vincenzi, A., Ruggiero, M., & Atienza, D.

(2020). Neural Network-Based Thermal Simulation of

Integrated Circuits on GPUs. IEEE Transactions on

Computer-Aided Design of Integrated Circuits and

Systems, 39(12), 4567-4578.

Sufian, A., You, C., & Dong, M. (2021). A Deep

Transfer Learningbased Edge Computing Method for

Home Health Monitoring. arXiv preprint

arXiv:2105.02960.

Surrel, G., Aminifar, A., Rincon, F., Murali, S., & Atienza,

D. (2020). Online Obstructive Sleep Apnea Detection

on Wearable Sensors. IEEE Transactions on

Biomedical Circuits and Systems, 14(2), 209-220.

Velichko, A. (2021). A Method for Medical Data Analysis

Using the LogNNet for Clinical Decision Support

Systems and Edge Computing in Healthcare. arXiv

preprint arXiv:2108.02428.

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

570