Intelligent Healthcare with Federated Learning: A Brief Investigation

Hengjie Ma

Data Science and Big Data Technology with Contemporary Entrepreneurialism, Xi’an Jiaotong-Liverpool University,

Suzhou, China

Keywords: Federated Learning, Machine Learning, Healthcare, Medical.

Abstract: Intelligent healthcare is an emerging field that leverages technologies such as wearable Internet of Things

(IoT) devices and deep learning to analyze various types of medical data, including traditional records,

medical images, and sensor data from wearables. These innovations facilitate more accurate diagnosis and

personalized treatment. However, they also raise significant privacy concerns, as sensitive data collected from

devices like smart speakers and IoT wearables may be vulnerable to breaches. Federated Learning (FL) offers

a promising solution by allowing data to remain on local devices while sharing only model updates with a

central server. This method enhances privacy and reduces the risks associated with transferring personal

medical data. This paper summarizes some of the recent research outcomes in this field, including a brief

introduction to the federated learning algorithm and its variants, a privacy-preserved medical data processing

model SCALT, different FL-IoMT architectures according to data partition, a clustered federated learning

based multimodal COVID-19 diagnosis model and voice recognition-based Alzheimer’s disease detection

ADDetector. However, challenges such as data heterogeneity and hardware limitations remain, requiring

further algorithmic improvements and specialized hardware development. As FL holds the potential to

revolutionize healthcare, enabling safer, more efficient processing of medical data while protecting patient

privacy, this paper gives this brief review to investigate the current outcomes of this field and gives out.

1 INTRODUCTION

Intelligent healthcare is one of these days’ emerging

fields. This technology usually combines the usage of

wearable Internet of Things (IoT) devices and deep

learning methods to utilize the healthcare data

analysis and treatment from a variety of types of data

including regular medical records, e.g.,

Electrocardiogram (ECG), medical image data, gene

data and data collected from wearable IoT devices

(Sun & Wu, 2022; Kumar & Singla, 2021), other

types of data that is not from medical inspections

while could be processed through deep learning

methods, e.g., voice based neural diseases diagnostic

(Li et al., 2021).

However, with large numbers of healthcare data

collected for intelligent healthcare model training,

this technology soon faces some limitations.

According to Li et al. (Li et al., 2021), voice-based

Alzheimer's Disease detection implemented on smart

speakers would involve recording voices from users’

https://orcid.org/0009-0003-4218-7707

home environment, indicating a serious privacy

problem. Medical data collected from wearable IoT

devices used for training these models have the same

privacy issues as illustrated by Thilakarathne et al., as

those data typically have a strong connection with a

specific personnel, which centralized cloud

computing could lead to potential data leakage.

Furthermore, traditional centralized cloud computing

requires other infrastructure to provide smart service,

consuming more resources and less efficient

compared to if this intelligent could be implement on

device-side, which would enable real-time processing

(Guo et al., 2022). In this case, many emerging

studies combine the Federated Learning (FL)

algorithm to give these issues a potential solution.

Instead of centralized cloud computing, FL processes

data on the local client and then send a new iteration

of model back to the server for aggregation,

preventing the potential of sensitive data leakage

caused by data transferring. Based on the

implementation of FL, edge-computing is more

important as the model would be trained natively, as

516

Ma and H.

Intelligent Healthcare with Federated Learning: A Brief Investigation.

DOI: 10.5220/0013527500004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 516-520

ISBN: 978-989-758-754-2

a result edge intelligent could be achieved (Akter et

al., 2022). This could enable some IoT health data

sensors to provide intelligent services to the users

directly, rather than relying on a specified platform,

which contributes to energy efficiency according to

Akter et al (Akter et al., 2022). Considering that the

development of intelligent healthcare would be

greatly rely on large amounts of data as deep learning

algorithms seems gradually plays a more important

role in smart healthcare, involving FL would

contribute to balance data collection and privacy

protection. Besides, since no sensitive data

transferred in this process, this enables facilities

including hospitals and medical institutions to share

the training results, so that the final aggregated model

would generally perform better compared to the

models from only one institution (Kumar & Singla,

2021). Hence, FL provided an essential opportunity

for training outcomes to be shared, while keeping all

the data stored and processed natively.

Currently, the study of implementing FL on

intelligent healthcare has a range of outcomes, with

many of these already being able to be introduced to

practice. Sun and Wu (Sun et al., 2022) introduced

SCALT, an efficient privacy-protective FL-based

medical sensor data classification system. Li et al. (Li

et al., 2021) developed a voice-based Alzheimer’s

disease (AD) detection system deployed on smart

speakers with FL for privacy protection. In the

meantime, with the continuous development and

growth of FL algorithms and edge-computing

hardware, intelligent healthcare now has a relatively

firm platform and infrastructure to be further

developed. This study will give a review of recent

research outcomes in the field of Federated Learning

and intelligent healthcare. The organization of this

article is as follows. Section 2 introduces current

methods of implementing FL algorithms on

intelligent healthcare, Section 3 discusses current

hurdles and possible future developments, and

Section 4 concludes the whole paper.

2 METHOD

2.1 Introduction to Federated Learning

FL shown in Figure 1 is a distributive deep learning

algorithm where only locally trained models will be

sent back to the central server for aggregation, as the

training data will be kept on the local devices. This

algorithm was developed to reduce the privacy risks

since it prevents data from being transmitted between

devices and the cloud (McMahan et al., 2017). With

the further development of FL, there comes with a

range of variants including FedAvg, which performs

stochastic gradient descent locally with a server for

model aggregation, and FedProx, which modifies

FedAvg and adds a proximal term 𝜇, providing that

FedProx performs more robustly in heterogeneous

networks (Li, Sahu et al., 2020). FL is currently

widely deployed especially on edge-computing

devices including mobile phones and IoT hardware,

providing features containing input predictions, photo

processing and Optical Character Recognition (OCR)

without violating privacy.

Figure 1: The workflow of Federated Learning (Ullah et al.,

2023).

2.2 Federated Learning Based

Healthcare Data Analysis

2.2.1 SCALT

Federated Learning concept could be implemented

onto IoMT devices, provide a better privacy

protection while processing the patients’ health

information. A system named SCALT (Sun & Wu,

2022) is proposed to deploy on edge devices

including wearables and cell phones. Those devices

will collect health data including ECG, heart rate,

body temperature and so on. Then those data will be

processed by SCALT, generate the result to assist the

doctor’s decision. During this process, the local

model will be trained on edge devices using locally

collected dataset. The training process of SCALT first

denoise the collected dataset by wavelet transform,

then it will be reconstruct and segment into short

slices. Each slice will then be normalized and extract

the feature. Since SCALT is mainly used on edge

Intelligent Healthcare with Federated Learning: A Brief Investigation

517

devices with limited computing power, it uses a

lightweight 1-D CNN network. The extracted slice

features will then be classified using per-class

classifier (PCC). The use of PCC allows SCALT to

adapt different tasks as a PCC could be simply added

for a new class. An original SCALT model will be

initially trained on the cloud, then distribute to edge

devices for local training. By implementing FL, only

new parameters will be required to send back to the

cloud server, avoiding the exposure of original health

sensor data.

2.2.2 FL-IoMT

Another study done by Thilakarathne et al.

(Thilakarathne et al., 2022) also introduced FL-IoMT

architecture. This study categorized FL into three

different types according to their data partition

architecture: vertical FL (VFL), horizontal FL (HFL)

and federated transfer learning (FTL). In a HFL case

the local databases are with the same feature space but

different sample spaces, while a VFL case is an

inversion of HFL. In a FTL case the local databases

will have different feature spaces and sample spaces.

The study also divided the architecture into two

categories with respect to whether the network is

centralized or decentralized. A centralized FL

network, according to Thilakarathne et al., is usually

for protect the privacy and security of the training

dataset, as the decentralized network will not

mandatorily require a central server to aggregate each

local model. In this case peer-to-peer (P2P)

communications will be introduced between each

client for model exchanges. By applying FL

architecture, researchers could access to datasets

which may contain sensitive medical information

safely, which could also accelerate the improvement

of algorithms as mentioned by Thilakarathne et al. FL

could also be implemented in medical imaging

processing and analysis. Thilakarathne et al. provides

that computer vision algorithms with FL could be

used for tumor segmentation and computed

tomography (CT) and magnetic resonance imaging

(MRI) image diagnosis.

2.3 Federated Learning Based Edge

Diagnosis

2.3.1 Clustered Federated Learning

A study conducted by Qayyum et al. (Qayyum et al.,

2022) focused on multi-model edge diagnosis of

COVID-19 using FL. As proposed by Qayyum et al.,

traditional methods of diagnosing COVID-19 usually

only use one modality and without FL, some FL-

based methods still use single modality to diagnose.

Considering deploying ML at the edge could easily

face privacy issues, along with data heterogeneity,

communication costs and other challenges.

Considering those challenges, Qayyum et al.

(Qayyum et al., 2022) proposes a chest image

classification model using multiple sources based on

Clustered Federated Learning (CFL). The two

clusters are two medical facilities with one of them

have clients with X-ray images, another have clients

with Ultrasound images. Local training will be

performed at those two facilities, then send the

weights updates to central server to aggregate the

multimodal model. In addition, as noted by this study,

conventional FL could not train a single multi-modal.

For the dataset, they first convert the X-ray and

Ultrasound images into gray-scale image, then resize

to 256 × 256. Those resized images will then be

normalized for training. Adam optimizer was used

during their training process. The study compared the

performance of CFL multimodal model and

conventional FL multimodal model with specialized

conventional FL model. As shown by the study, the

CFL model performs better compares to conventional

FL methods, and it could also confirm that a

collaboratively trained model is able to recognize the

test images without having explicit knowledge about

all these modalities.

2.3.2 ADDetector

Another study completed by Li et al. proposed a FL

based privacy-preserving smart healthcare system

named ADDetector that could be deployed onto smart

speakers and use users’ voice input to detect whether

the user has AD. They designed this system to be easy

to deploy, high efficiency and privacy preserving, and

FL is relatively suitable for those requirements (Li et

al., 2021).

The ADDetector, as mentioned by Li et al. (Li et

al., 2021), is constructed by three layers: user layer,

detection layer and the cloud layer. The data

collection model was used in the user layer, prompt

the user to provide voice samples for AD detection.

The detection process will extract features from both

acoustic and linguistic aspects. Then the data will be

processed by the FL-based decision module and

assign the features to detection clients, and then

optimize the AD classification by interacting between

clients and the cloud. By implementing FL network,

those raw data could be processed at user level,

avoiding transferring voice records containing the

users’ home environment and personal voice to the

DAML 2024 - International Conference on Data Analysis and Machine Learning

518

cloud. Finally, they use the Asynchronous Privacy-

Preserving Aggregation Module for model

aggregation, as well as ensuring the integrity of the

interaction between clients and the cloud.

For the performance, they utilized the ADReSS

Challenge dataset for testing. When using 3 clients in

the FL network, ADDetector have an accuracy of

81.9% under Laplace-based DP protection and

cryptography-based scheme, and it takes 711.55ms

per user detection on a desktop system, which

considered acceptable in the smart home scenario.

Further study by them shows that the most time-

consuming stage is feature generation (97.9% of total

time), proving that the privacy-preserving schemes by

this study are time efficient. For real world scenarios,

the model proposed by them maintains an accuracy of

78%, demonstrating the effectiveness and robustness

of ADDetector.

3 DISCUSSION

As more studies concentrate on FL implementation

on intelligent healthcare, there do have more

approaches for medical data processing and disease

diagnosis based on different modal of data. However,

there still some hurdles on the way to overcome. As

the FL models will mainly be trained locally, which

means the dataset would mainly come from the edge

device, usually varies a lot compared to centralized

training as centralized training would usually use pre-

made high quality datasets, while edge devices may

not be able to perform this selection process on the

dataset. As a result, this heterogeneity would cause

the aggregated model’s performance to vary from

device to device, so do the training process. This

would require improving the FL algorithm to adapt

such varying environments or design the original

model to be client-specific to minimize the issue

(Qayyum et al., 2022).

Another issue would be potential impacts of data

heterogeneity in the system. This would become more

significant when involving model sharing between

medical authorities, as a well-trained model from

hospital A might have a decreased performance when

using data inputs from hospital B. Although as

mentioned above, FedProx model would help to

minimize the impact of data heterogeneity, however

considering the application is in the field of

healthcare, this issue might still require the

development of FL algorithm to further improve the

overall performance of the models under data

heterogeneous conditions, which is considered

common in realistic implementation.

Besides the algorithm itself, the performance

bottlenecks of the hardware are another potential

hurdle. Considering this technique has a strong

connection with IoT devices, it is a necessity that FL

models should be designed to be efficient, drawing a

little amount of computing power while still maintain

an acceptable accuracy. This would require the

optimization and possibly specialized models to focus

on a certain objective to reduce the overall

performance consumption, or the improvement of IoT

hardware to allow running higher performance

models with no significant higher consumption of

power.

While besides those difficulties, the future of FL

based network on healthcare is still bright. For the FL

algorithm itself, the original FedAvg algorithm would

simply drop those clients who could not reach the

required local steps, causing the system could easily

be disrupted by the heterogeneity inside the network.

While improved FL algorithms including FedProx

and Scaffold (Karimireddy et al., 2020) avoids to do

so and use certain computations to add correction

onto the local updates from each client, in this way

optimized the speed and robustness of the training

process.

In the meantime, those improved FL algorithms

also perform better on heterogeneous data compared

to the original FedAvg. According to Li et al.,

FedProx would have a significantly improved

training loss in normal data heterogeneous conditions

and could converge in extreme data heterogeneous

conditions compared to FedAvg, which would not

converge in this case. This made FL based medical

models more applicable, since real world data would

have some heterogeneity, it is considered essential for

a medical model to maintain a certain accuracy in

heterogeneous conditions.

The development of IoT devices also supported to

solve the performance bottleneck. With more

efficient chips developed and the appearance of deep

learning specified processors including Neural

network Processing Unit (NPU) and Tensor

Processing Unit (TPU), low power IoT devices were

able to process high performance models efficiently,

eventually improving the overall accuracy of the

model while only requiring a little amount of power,

making the actual large-scale application to be

possible.

4 CONCLUSIONS

This paper summarized some of the recent methods

and implementations of Federated Learning in

Intelligent Healthcare with Federated Learning: A Brief Investigation

519

intelligent healthcare. The paper briefly introduced

the mechanism of FL and one improved FL method:

FedProx. The paper reviewed two methods: SCALT

and FL-IoMT architecture, and two edge diagnosis

implementations: edge COVID-19 diagnosis and

ADDetector. Currently the implementation of FL

models helped to provide a privacy-preserved way for

better data processing and earlier diagnosis, however

those methods would require further validation to

build trust among users, or be verified by medical

authorities to be used for realistic applications. Those

methods would need to be utilized for better

performance with reduced hardware cost, which also

requires to combine with the development of FL

algorithm itself.

REFERENCES

Akter, M., Moustafa, N., Lynar, T., & Razzak, I. 2022.

Edge intelligence: Federated learning-based privacy

protection framework for smart healthcare systems.

IEEE Journal of Biomedical and Health Informatics,

26(12), 5805-5816.

Guo, K., Chen, T., Ren, S., Li, N., Hu, M., & Kang, J. 2022.

Federated learning empowered real-time medical data

processing method for smart healthcare. IEEE/ACM

Transactions on Computational Biology and

Bioinformatics.

Karimireddy, S. P., Kale, S., Mohri, M., Reddi, S., Stich,

S., & Suresh, A. T. 2020. Scaffold: Stochastic

controlled averaging for federated learning. In

International Conference on Machine Learning (pp.

5132-5143). PMLR.

Kumar, Y., & Singla, R. 2021. Federated learning systems

for healthcare: perspective and recent progress. In

Federated Learning Systems: Towards Next-

Generation AI (pp. 141-156).

Li, J., Meng, Y., Ma, L., Du, S., Zhu, H., Pei, Q., & Shen,

X. 2021. A federated learning based privacy-

preserving smart healthcare system. IEEE

Transactions on Industrial Informatics, 18(3).

Li, T., Sahu, A. K., Zaheer, M., Sanjabi, M., Talwalkar, A.,

& Smith, V. 2020. Federated optimization in

heterogeneous networks. Proceedings of Machine

Learning and Systems, 2, 429-450.

McMahan, B., Moore, E., Ramage, D., Hampson, S., & y

Arcas, B. A. 2017. Communication-efficient learning

of deep networks from decentralized data. In

Artificial Intelligence and Statistics (pp. 1273-1282).

PMLR.

Qayyum, A., Ahmad, K., Ahsan, M. A., Al-Fuqaha, A., &

Qadir, J. 2022. Collaborative federated learning for

healthcare: Multi-modal COVID-19 diagnosis at the

edge. IEEE Open Journal of the Computer Society, 3,

172-184.

Sun, L., & Wu, J. 2022. A scalable and transferable

federated learning system for classifying healthcare

sensor data. IEEE Journal of Biomedical and Health

Informatics, 27(2), 866-877.

Thilakarathne, N. N., Muneeswari, G., Parthasarathy, V.,

Alassery, F., Hamam, H., Mahendran, R. K., &

Shafiq, M. 2022. Federated learning for privacy-

preserved medical internet of things. Intelligent

Automation & Soft Computing, 33(1), 157-172.

Ullah, F., Srivastava, G., Xiao, H., Ullah, S., Lin, J. C. W.,

& Zhao, Y. 2023. A scalable federated learning

approach for collaborative smart healthcare systems

with intermittent clients using medical imaging. IEEE

Journal of Biomedical and Health Informatics.

DAML 2024 - International Conference on Data Analysis and Machine Learning

520