← All theses

Department of Computer Science and Engineering

Privacy-Preserving Federated Learning for Clinical Text Classification

Shithy, Muktadiru Mannan

Abstract

Automatic classification of symptom descriptions into disease categories can support clinical decision-making. However, centralized machine learning requires training data to be collected in one location, which can create privacy concerns when dealing with sensitive clinical text. This thesis investigates Federated Learning (FL) as a privacy-preserving approach for symptom-to-disease text classification. A client-server FL pipeline is implemented using the Flower framework with a regularized Text Convolutional Neural Network (TextCNN). The Symptom2Disease dataset contains 1,200 samples from 24 disease classes and is divided among three simulated clients using a non-IID Dirichlet distribution (α = 0.5). The federated model is trained using the FedAvg aggregation strategy for 30 communication rounds and is compared with centralized machine learning models, including Logistic Regression, Naive Bayes, Support Vector Machine, and Random Forest. The proposed federated TextCNN achieves 93.75% test accuracy, 93.50% weighted F1-score, and a ROC-AUC of 0.9976. The results show that Federated Learning can maintain strong classification performance while keeping the raw training data distributed across clients. This study demonstrates the potential of a lightweight federated TextCNN approach for privacy-sensitive clinical text classification and provides a basis for future improvements using stronger privacy mechanisms and advanced language models.