Abstract
Federated Learning (FL) offers a privacy-preserving and decentralized paradigm for machine learning, making it particularly suitable for analyzing sensitive psychological and physiological data. This study aims to develop and evaluate a federated learning framework for assessing the psycho-emotional well-being of students in Kazakhstani educational institutions, where data privacy and infrastructural constraints pose significant challenges. We benchmark three FL algorithms, such as FedAvg, FedOpt, and FedProx, on heterogeneous, institution-level datasets that combine sleep, dietary, activity, and self-reported emotional measures. Experiments simulate cross-device, non-IID deployments and evaluate convergence, accuracy, and stability across ten communication rounds. Results show that FedProx attains the best trade-off between accuracy and stability under non-IID conditions (peak accuracy is 99.9%), while FedOpt provides faster early convergence, and FedAvg performs well for more homogeneous partitions. The methodological contribution comprises optimized aggregation and adaptive client weighting to mitigate non-IID effects in resource-constrained educational settings. These findings validate FL as a scalable, privacy-preserving approach for mental health monitoring in education and support its use for early intervention and resilience tracking. The proposed framework contributes to data-driven mental health policy design in educational systems, addressing both ethical and infrastructural considerations. The study discusses limitations of the simulated setup and outlines directions for broader deployment and cross-silo validation.