Privacy-preserving multimodal student well-being analytics via federated temporal transformers

Student well-being analytics may support timely educational assistance, but behavioural, psychological, and academic records are sensitive and difficult to pool. We present a federated temporal Transformer that combines smartphone sensing, ecological momentary assessment (EMA), academic records, and digital behavioural traces while retaining raw records on student clients. Because psychometric variables can overlap with a composite well-being target, evaluation is organized around a leakage-free multimodal protocol, a temporally separated forecasting protocol, and a full-feature diagnostic upper bound.

On 48 StudentLife participants, student-level leave-one-student-out predictions yielded an area under the receiver operating characteristic curve (AUROC) of 0.834 for centralized training and 0.824 for FedProx (\(p=0.21\)). Record-level differentially private training at \(\varepsilon =4\) achieved an AUROC of 0.791; the corresponding membership-inference AUROC decreased from 0.92 without privacy noise to 0.64. Including target-overlapping survey variables inflated AUROC by 0.094–0.107, demonstrating why the leakage-free protocol is used for the main claims.

Directional findings were reproduced on 497 independent GLOBEM participants, where FedProx achieved an AUROC of 0.783. Cross-modal ablation, calibration, privacy attacks, external validation, and explanation-stability analyses jointly characterize predictive utility and deployment limits. The framework is intended for research and educational risk stratification, not clinical diagnosis or autonomous intervention.

This research received no external funding. School of International Education, Shandong Vocational and Technical University of International Studies, No. 99 Shanhai Road, Rizhao, 276826, Shandong, China Student Affairs Office, Shandong Water Conservancy Vocational College, No. 677 Xueyuan Road, Rizhao, 276826, Shandong, China The authors declare no competing interests. This work is a secondary analysis of de-identified public research datasets and involved no new recruitment, intervention, or contact with participants.

Data were used in accordance with their repository access conditions and the ethical procedures reported by the original dataset investigators. The proposed outputs are intended for research and educational support, not clinical diagnosis. Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

A post hoc capacity stress test was conducted under L4 only to identify gross overfitting patterns; it did not supply any headline result or change the L2 outer-test predictions. The 4-layer, 256-dimensional candidate achieved L4 diagnostic AUROC 0.941. Increasing the temporal depth to six layers reduced AUROC to 0.930 and increased the training–validation accuracy gap from 0.017 to 0.033; increasing \(d_}\) to 512 reduced AUROC to 0.929 and increased the gap to 0.041.