Enhancing Privacy in Mental Health Prediction Using Centralized and Federated Learning Models

A large percentage of the population suffers from mental health disorders, which imposes serious challenges to healthcare systems around the globe. Early detection of conditions such as depression and anxiety disorders is vital; however, diagnosing methods incorporate wrong predictions and expose patients to data breaches. To bridge this gap of predictions with secured data, this paper presents an approach for analyzing mental health disorders using centralized models and federated learning considering a dataset of workplaces and demographic data with various conditions gathered from surveys. The results indicate that centralized models achieved an accuracy of 82.01 % as compared to federated learning models with 80.68%, federated learning models provided comparable performance along with a decrease in losses shared among multiple clients. The calculated privacy loss <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\epsilon= 0.5321)$</tex> and delta <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\delta=1\!\!\mathrm{x}10^{-5})$</tex> for federated learning setup using differential privacy which indicates enhanced privacy and security among multiple clients along with accurate results, highlighting their potential to safeguard sensitive mental health data.