Implementasi Federated Learning untuk Klasifikasi Penyakit dengan Dataset Tidak Seimbang

Abstract

Imbalanced data is one of the main challenges in applying machine learning to stroke disease classification, where the number of stroke patient data is significantly smaller than non-stroke patient data. This condition causes classification models to become biased toward the majority class, resulting in poor detection of stroke cases. In addition, centralized learning approaches raise privacy and security concerns due to the sensitive nature of patient data. Therefore, this study proposes a Federated Learning Random Forest (FL-RF) approach to improve classification performance on imbalanced data while preserving data privacy. The study uses the Cerebral Stroke Prediction – Imbalanced Dataset obtained from Kaggle. The research stages include data preprocessing, semi non-IID data distribution across multiple clients, local training using Balanced Random Forest, and model aggregation within a federated learning environment. Model evaluation was conducted using accuracy, precision, recall, F1-score, confusion matrix, and ROC curve metrics. The results show that the conventional Random Forest achieved an accuracy of 0.9817 but failed to detect the minority class. Meanwhile, the proposed FL-RF model obtained an accuracy of 0.7976 with an improved recall of 0.56 and an F1-score of 0.0910 for the stroke class. These findings indicate that the FL-RF approach is more effective in improving sensitivity toward minority classes compared to conventional Random Forest.