Federated Learning Model for Mental Health Prediction and Personalized Intervention

Abstract

Mental health screening is dependent on self-reported behavioural information, but the privacy issue is a major constraint on the collection of data. To overcome this problem, we introduce a privacy-preserving federated learning (FL) model for mental health prediction and severity classification, where the raw psychological information is not transmitted from the user’s device. The system handles 27 local input variables, including normalized demographic variables and DASS-21 questionnaire results, and uses the Flower framework with PyTorch to train a Multilayer Perceptron (MLP) model using the Federated Averaging (FedAvg) algorithm. Regression models were separately trained for stress, anxiety, and depression, and the continuous scores ranging from 0 to 21 were mapped to standardized DASS-21 severity levels. The performance of the model was tested on a stratified test dataset, showing stable federated learning convergence with a Mean Absolute Error (MAE) of 0.9-1.2, R² score of 0.85+, and severity classification accuracy of up to 90.36%. The trained models were integrated into a full-stack system with a React frontend and FastAPI backend, allowing real-time prediction, tracking, and downloadable clinical-style reports. These findings confirm that federated learning is a safe and efficient method for digital mental health screening with high predictive accuracy and strict data privacy.