Advancing elderly social care dropout prediction with federated learning: client selection and imbalanced data management

Accurate prediction of user dropout is crucial for enhancing the effectiveness of social care applications developed for the elderly. Given the sensitive nature of healthcare data, this study introduces a Federated Learning (FL) approach to predict user dropout while preserving data privacy by keeping personal data on local devices and only sharing model updates. We propose three client selection methods-Balanced label distribution, Data-Rich client prioritization, and a Combined strategy-that enhance predictive performance, computational efficiency,reduce carbon emissions, and decrease training duration. Additionally, we evaluate sampling strategies, including oversampling, undersampling, and combined over-under sampling, to address class imbalance in real-world dataset. Our experimental results demonstrate that our FL approach with the Balanced client selection method achieves an F1 score of 76.4%, surpassing the traditional centralized model by 13.5% and reducing training time and carbon emissions by 88.7%. This study underscores the potential of FL combined with strategic client selection and sampling methods to improve dropout prediction in healthcare applications while preserving privacy and promoting environmental sustainability.