Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr626

Liver Disease Prediction using Federated Learning

P. Uma Shankar1; E. Surya Rahul2; K. Durga Rao3; K. Satish4;
U. Dayanand Kumar5; D. Ravindra6; G. Subbarao7
1 Asst.Professor,
2,3,4,5,6,7 B.Tech Student

1,2,3,4,5,6,7
Sri Vasavi Engineering College(A), Pedatadepalli, Tadepalligudem

Publication Date: 2025/04/22

ABSTRACT: Developing a model through a centralized approach, where data is shared among all stakeholders, enhances
its reliability. However, privacy concerns—especially regarding medical datasets—often impede data sharing. Numerous
machine learning models have been created using isolated datasets, leading to challenges with overfitting and poor
performance on new datasets. Consequently, there is an urgent need to create a model that achieves accuracy comparable to
centralized models while upholding security standards.

Efficient diagnosis of liver disease typically depends on analyzing imaging techniques such as CT and MRI scans.
Traditional machine learning methods face difficulties due to the decentralized nature of medical data across institutions,
which is further complicated by stringent privacy regulations. Federated learning offers a solution by enabling local model
training, allowing institutions to collaborate without exchanging raw data; instead, they share only model updates. This
approach safeguards data privacy while enhancing model reliability.

Keywords: Federated Learning, Decentralized Model, Model Aggregation, privacy preservation, Medical Imaging, Collaborative
Training.

How to Cite: P. Uma Shankar; E. Surya Rahul; K. Durga Rao; K. Satish; U. Dayanand Kumar; D. Ravindra; G. Subbarao (2025),
Liver Disease Prediction using Federated Learning. International Journal of Innovative Science and Research Technology, 10(4),
880-887. https://doi.org/10.38124/ijisrt/25apr626

I. INTRODUCTION diseases such as cirrhosis and hepatitis. The model achieved

an accuracy of 82.7%, demonstrating the potential of ML in
Liver disease is a critical health issue requiring early liver disease prediction. However, the study faced limitations
and accurate diagnosis. Traditional machine learning models related to data imbalance, where a significantly larger number
struggle due to privacy concerns and decentralized medical of healthy cases compared to diseased cases resulted in biased
data. Our project uses Federated Learning (FL) to enable predictions. The absence of imaging data in the analysis also
multiple institutions to train a shared model without restricted the model’s ability to detect structural liver
exchanging raw patient data. A Convolutional Neural abnormalities.
Network (CNN) is used to analyze CT and MRI scans for
liver disease prediction. FL ensures data privacy, improved In 2017, Zhang et al. introduced a hybrid
model generalization, and compliance with regulations classification approach that integrated Support Vector
like HIPAA and GDPR. Hospitals train models locally and Machines (SVM) with Random Forest (RF) to enhance liver
share only updates, enhancing security. This approach disease detection. Their dataset included both clinical
advances medical AI while maintaining patient attributes (such as bilirubin levels and albumin count) and
confidentiality. ultrasound imaging features. The hybrid approach was able to
capture complex relationships between different attributes,
II. LITERATURE SURVEY leading to an improved accuracy of 85.4%. Despite these
improvements, the model struggled with overfitting when
In 2016, Gupta et al. developed a machine learning- applied to new patient data. The reliance on manually
based model for early-stage liver disease classification using extracted features from ultrasound images also reduced its
a combination of Decision Trees, Naïve Bayes, and K- scalability for real-world applications.
Nearest Neighbors (KNN). Their study utilized a dataset
containing patient symptoms, liver function test results, and In 2018, Smith et al. proposed an end-to-end deep
medical history to identify patterns indicative of liver learning framework that utilized Convolutional Neural

IJISRT25APR626 www.ijisrt.com 880

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr626
Networks (CNNs) to classify liver ultrasound images. Unlike which limited the feasibility of deployment in low-resource
previous studies, this approach automated feature extraction, healthcare facilities.
removing the need for manual selection of relevant
parameters. The CNN-based model was trained on a large In 2020, Lee et al. explored a hybrid deep learning
dataset consisting of thousands of liver ultrasound images architecture that combined CNNs with Long Short-Term
collected from multiple hospitals. The model achieved an Memory (LSTM) networks to analyze both static and
accuracy of 87.5%, outperforming traditional ML classifiers. sequential medical imaging data. The inclusion of LSTMs
However, the study relied on a centralized dataset, which allowed the model to process temporal changes in liver
raised concerns about data privacy and security. Additionally, structure, which is crucial for monitoring disease progression
the requirement for large amounts of annotated training data over time. The study achieved 91.2% accuracy, making it
made the approach resource-intensive. one of the most effective approaches at the time. However,
the centralized data training approach raised security risks, as
In 2019, Kumar et al. implemented an advanced deep patient data had to be aggregated in a single location, making
learning model using a combination of CNNs and it vulnerable to cyberattacks.
autoencoders to extract meaningful features from liver CT
scan images. Their model utilized unsupervised learning to III. METHODOLOGY
pre-train the feature extraction layers, reducing dependency
on labeled datasets. The proposed approach demonstrated an  System Structure
accuracy of 89.6% and showed improved generalization The system structure consists of multiple AI
across different liver disease types, including fatty liver and components working together to predict liver disease
fibrosis. However, the study highlighted a major drawback— using Federated Learning (FL). The key components are:
the high computational cost of training deep learning models,

Fig 1 Federated Learning System Architecture

 Data Collection & Preprocessing Module  Local AI Training Agents (Client Nodes)

 Gathers medical imaging data (CT scans, MRIs,  Each hospital or medical institution acts as a local AI
ultrasound) and clinical test results (bilirubin, ALT, AST, agent, training a deep learning model on its own patient
albumin, etc.). data.
 Applies data cleaning, normalization, augmentation, and  Uses CNN-based models for image classification and
feature selection to improve training efficiency. MLP/ANN for structured clinical data.

IJISRT25APR626 www.ijisrt.com 881

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr626
 Ensures raw patient data never leaves the hospital, V. PROPOSED SYSTEM
preserving privacy.
The proposed Federated Learning (FL)-based system
 Federated Learning Coordinator (Global Server) enables multiple hospitals to collaboratively train a CNN-
based model without sharing raw medical data. Instead of
 Manages communication between all hospitals transferring sensitive patient information, each hospital trains
participating in Federated Learning. its local model on its own dataset, ensuring data privacy and
 Aggregates model updates from different hospitals using compliance with regulatory standards. The trained model
Federated Averaging (FedAvg). updates are then sent to a central server, where they are
 Sends the updated global model back to clients for the next aggregated using the Federated Averaging (FedAvg)
training round. algorithm to create a global model. This decentralized
approach preserves patient privacy by keeping medical data
 Secure Model Update Transmission Module within hospital premises while also reducing the reliance on
centralized storage. Additionally, by learning from diverse
 Encrypts and transmits model weights and gradients from datasets across multiple hospitals, the model achieves better
client hospitals to the global server. generalization, leading to improved predictive accuracy for
 Uses end-to-end encryption and differential privacy liver disease diagnosis across different patient demographics.
techniques to secure communications.
VI. IMPLEMENTATION
 Model Evaluation & Performance Analysis Module
The implementation of this project follows a structured
 Tests the aggregated global model using independent approach, ensuring privacy-preserving federated learning,
validation datasets. deep learning model training, and secure communication
 Measures accuracy, precision, recall, F1-score, and AUC- between clients and the central server.
ROC to assess model effectiveness.
 Fine-tunes the model to improve prediction reliability.  Step 1: Configuring the Development Environment
The development environment is set up with necessary
 Liver Disease Prediction & Deployment Module frameworks and libraries, ensuring seamless execution of the
federated learning model. A suitable code editor such as VS
 Deploys the final trained model in hospitals, cloud-based Code or PyCharm is used for development. Two key scripts,
AI platforms, and mobile applications. one for the central server and another for the client nodes, are
 Allows doctors to input new patient data (medical created to manage the federated learning workflow.
images, test results) for real-time liver disease risk
assessment. Install the dependencies required for project such as
 Assists in early diagnosis and treatment recommendations. tensorflow numpy socket pickle open cv- python matplotlib

 Implementation Details  Step 2: System Architecture Setup

The system follows a federated learning approach where
 Programming Language: Python 3.11 multiple clients, representing hospitals, train a deep learning
 Frameworks & Libraries: Streamlit (UI), model locally and send only the model updates to a central
Tensorflow(Training), Keras (Model evaluation) server. The key components of the system include:
 Pickle: Serializes model architecture and weights before
transmission.  Client Nodes (Hospitals):
Each hospital trains a Convolutional Neural Network
IV. EXISTING SYSTEM (CNN) model on locally stored liver disease image data
without sharing the raw images.
The conventional liver disease diagnosis system
involves collecting liver ultrasound and CT scan images and  Central Server:
storing them in a centralized database. A deep learning model The server receives the trained model updates from
is then trained on the combined dataset to classify liver different clients and aggregates them using the Federated
diseases, and the trained model is deployed for real-time Averaging (FedAvg) algorithm to update the global model.
diagnosis. However, this system has several limitations.
Firstly, storing patient data in a centralized server increases  Communication Mechanism:
privacy risks, making it vulnerable to potential data breaches. A secure connection is established between the server
Additionally, sharing medical data across institutions may and the clients, enabling the exchange of model updates while
conflict with regulatory compliance frameworks, restricting preserving privacy.
data accessibility and collaboration. Furthermore, a
centralized approach often suffers from data imbalance  Step 3: Federated Learning Model Training
issues, where the model may not generalize well across Each client follows a systematic training process:
diverse populations, leading to biased predictions and
reduced accuracy in real-world applications.

IJISRT25APR626 www.ijisrt.com 882

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr626

 The local dataset is preprocessed to ensure uniformity in

training.  Step 5: Model Evaluation & Performance Analysis
 A CNN model is initialized and trained on the hospital’s Once the global model is trained, its performance is
local dataset. evaluated using standard metrics:
 Once training is complete, the model weights are extracted
 Accuracy and loss graphs are generated to analyze training
and prepared for transmission to the central server.
efficiency.
 The weights are securely sent to the central server without  Confusion matrix and precision-recall analysis are
sharing raw medical data. On the server side, the conducted to assess the classification performance.
following steps are executed:  Comparative analysis is performed to measure
improvements over a centrally trained model.
 Model updates from multiple clients are received.
 The FedAvg algorithm is applied to aggregate the  Step 6: Deployment & Testing
weights and update the global model.
 The updated global model is redistributed to clients for The final trained model is deployed for liver disease
the next round of training. prediction. The system is tested by running the central server
and multiple client nodes, allowing them to collaborate in
 Step 4: Security & Privacy Mechanisms training the model. The global model can then be used for real-
time liver disease diagnosis, potentially integrated into a web-
 Privacy Preservation: based application for practical deployment.
Raw medical images are not shared, ensuring
compliance with data protection regulations. VII. RESULT AND ANALYSIS

 Secure Model Transmission: The evaluation of the Liver Disease Prediction using
Encrypted communication is used to prevent Federated Learning system is based on key performance
unauthorized access during model updates. metrics, model accuracy, and comparative analysis. The
results demonstrate the efficiency, accuracy, and advantages
 Homomorphic Encryption (Future Scope): of the proposed approach.
Additional security measures, such as encryption Home page.
techniques, could be implemented to enhance privacy
further.

Fig 2 Liver Disease Classification

IJISRT25APR626 www.ijisrt.com 883

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr626

Fig 3 Running of clients

 Clients Training  The asynchronous updates indicate clients training at

different speeds, reflecting real-world hospital
 Each window represents a different client training a local collaboration
model before sharing updates for global aggregation.  This setup helps build a robust liver disease prediction
 The logs show training progress, loss, and accuracy, model without compromising patient data security.
highlighting variations in local datasets.

Fig 4 Global Server Training

IJISRT25APR626 www.ijisrt.com 884

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr626
 Global Model Training  Hospitals return only the updated model (not the data) to
the server.
 Global server builds an initial model.  Server aggregates all the updates to better the model.
 The model is shipped to all the hospitals (clients).  This is repeated for a number of rounds.
 Every hospital trains the model on their local liver images.  Final model is trained using all hospitals’ knowledge,
without sharing any patient data.

Fig 5 Liver Disease Prediction (Fibrosis)

IJISRT25APR626 www.ijisrt.com 885

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr626
 Liver Disease Prediction (nodes). Federated learning minimizes having to move huge
image datasets to a central server, reducing communication
 Users upload liver scan images (e.g., ultrasound, CT, overhead and allowing for real-time or near-real-time
MRI). collaborative training.
 The system predicts liver disease using a deep learning
model.  User Accessibility
 Federated Learning ensures training happens across Our solution gives a straightforward and easy
hospitals without sharing raw data. mechanism for federated training. Each client updates the
 Only model updates are shared, keeping patient data global model with their local image data alone, and
private. predictions are sent back immediately, making the procedure
 The system shows if the liver is Healthy or Diseased, with easy for use in clinical or research settings.
a confidence score.
 Designed for secure and privacy-preserving liver disease  Robustness Across Institutions
detection. Multi-dataset training improves the stability of the
model. Through experience with diverse medical images, the
 Result Analysis: model gains ability to handle novel and unseen data, and the
The proposed liver disease prediction system with model is therefore an appropriate resource for support in the
strong performance and practical applicability will be diagnosis of liver disease across institutions.
illustrated using Federated Learning on medical image data.
The system addresses the key challenges in Medical AI by VIII. CONCLUSION
enabling collaborative model training without the need for
centralized data collection. Our Project is a liver disease prediction system based
on Federated Learning for medical image data. It allows
 Accuracy of Prediction several hospitals to jointly train a model without exchanging
The federated learning model is extremely accurate at sensitive patient data, thus maintaining data confidentiality
predicting liver disease from image-based data. By utilizing while achieving high prediction accuracy. The decentralized
disparate datasets across different sources, the model takes approach facilitates early detection of liver disease,
advantage of rich feature learning, which maximizes its supporting timely diagnosis and intervention. The system is
capacity to generalize across diverse patient cases. highly scalable and adaptable with different medical datasets,
which corresponds to real-world healthcare privacy
 Data Privacy (Decentralization) requirements. By not using centralized storage of data, it
One of the key benefits of our method is that institutions minimizes legal and ethical issues related to the handling of
can train on their individual local data without revealing raw sensitive medical data. The model can generalize well across
images. This local training is privacy-preserving and adheres institutions, improving its diagnostic accuracy. As a whole,
to real-world restrictions regarding sensitive medical our solution is effective, secure, and pragmatic for medical
information, although encryption was not utilized. AI use. It facilitates collaboration in healthcare without
breaching patient privacy and provides a platform for future
 Efficiency and Performance development. This project showcases how federated learning
The system shows effective training performance can be harnessed in creating smart and moral healthcare
through parallelization of computation between clients products.

disease even in resource-constrained hospitals, which makes

FUTURE ENHANCEMENT the solution more accessible and effective at scale.

Improved features can comprise the deployment of REFERENCES

personalized federated learning to personalize the global
model to fit each hospital's individual features of the liver [1] A. Padthe, R. Ashtagi, S. Mohite, P. Gaikwad, R.
disease data, enhancing the accuracy of diagnostics. Bidwe, and H. M. Naveen, “Harnessing Federated
Federated hyperparameter optimization can be utilized to Learning for Efficient Analysis of Large-Scale
optimize model parameters among clients automatically to Healthcare Image Datasets in IoT- Enabled Healthcare
deliver the best performance without human adjustment. Systems,” International Journal of Intelligent Systems
Synthetic data generation methods applied at the client level and Applications in Engineering (IJISAE), vol. 12, no.
will ensure class distribution balancing and the treatment of 10s, pp. 253–263, 2024.
rare liver disease instances to enhance model robustness. The [2] R. Ashtagi, V. Musale, V. S. Rajput, S.
system can be extended to process multi-modal inputs, like Chinchmalatpure, S. Mohite, and R. V. Bidwe,
integrating liver ultrasound images with blood tests reports, “Revolutionizing Early Liver Disease Detection:
to increase diagnostic accuracy. Adding explainable AI in the Exploring Machine Learning and Ensemble Models,”
federated structure will enable clinicians to comprehend and IJISAE, vol. 12, no. 13s, pp. 528–534, 2024.
believe model predictions, which is essential in healthcare
[3] M. A. Kuzhippallil, C. Joseph, and A. Kannan,
environments. Lastly, using light-weight, low-energy models
“Comparative Analysis of Machine Learning
on edge devices can facilitate real- time prediction of liver

IJISRT25APR626 www.ijisrt.com 886

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr626
Techniques for Indian Liver Disease Patients,” in
Proc. 2020 6th Int. Conf. Advanced Computing and
Communication Systems (ICACCS), IEEE, pp. 778–
782, 2020.
[4] J. Xu, C. Gentry, S. Weller, and C. Wu, “Secure
Federated Learning for Liver Cancer Diagnosis Using
CT Images,” IEEE Access, vol. 9, pp. 162519–162530,
2021.
[5] M. J. Sheller, B. Edwards, G. A. Reina, J. Martin, and
S. Bakas, “Federated Learning in Medicine:
Facilitating Multi-Institutional Collaborations Without
Sharing Patient Data,” Scientific Reports, vol. 10, no.
1, p. 12598, 2020.
[6] P. Vepakomma, O. Gupta, T. Swedish, and R. Raskar,
“Split Learning for Health: Distributed Deep Learning
Without Sharing Raw Patient Data,” in NeurIPS
Workshop on Machine Learning for Health (ML4H),
2018.

IJISRT25APR626 www.ijisrt.com 887