Privacy-Preserving Prediction of Psychological Distress Across Schools Using Academic Workload, Bullying, Sleep Quality, and Social Support: A Federated XGBoost Study

Abstract

Objective: This study aimed to develop and evaluate a privacy-preserving federated XGBoost model for predicting psychological distress among Norwegian upper secondary school students using academic workload, bullying victimization, sleep quality, and perceived social support.

Methods and Materials: A multicenter cross-sectional predictive study was conducted among 2,428 students aged 16–19 years from 12 public upper secondary schools in Norway. Psychological distress was assessed using the Kessler Psychological Distress Scale, while academic workload, bullying victimization, sleep quality, and perceived social support were measured using standardized self-report instruments. Each school retained its raw data locally and participated as an independent node in a federated learning network. Federated XGBoost was trained using locally computed gradient and Hessian statistics transmitted through secure aggregation. Data were divided into training, validation, and independent test subsets. Model performance was evaluated using the area under the receiver operating characteristic curve, area under the precision–recall curve, sensitivity, specificity, balanced accuracy, F1 score, and Brier score. Performance was compared with centralized XGBoost, locally trained XGBoost models, and federated logistic regression.

Findings: Federated XGBoost achieved an area under the receiver operating characteristic curve of .906 and an area under the precision–recall curve of .756 in the independent test sample. The model produced balanced accuracy of .848, sensitivity of .844, specificity of .851, an F1 score of .734, and a Brier score of .112. Its performance did not differ significantly from centralized XGBoost, whereas it significantly outperformed locally trained XGBoost models and federated logistic regression. Academic workload, poor sleep quality, recurrent bullying, low family support, short sleep duration, cyberbullying, and low friend support were the strongest predictors. Leave-one-school-out validation showed stable performance across participating schools.

Conclusion: Federated XGBoost provided accurate and well-calibrated prediction of psychological distress while preserving institutional control over sensitive student data, supporting its potential use as a privacy-conscious screening aid in school mental health systems.