Privacy-First Mental Health Solutions: Federated Learning for Depression Detection in Marathi Speech and Text

Abstract

Federated Learning (FL) is a cutting-edge approach that allows machines to learn from data without compromising privacy, making it especially valuable in sensitive areas like mental health. This research focuses on using FL to detect depression through speech and text data from a Marathi-speaking population. Depression, a widespread mental health issue, often leaves subtle clues in the way people speak and write, making speech and text analysis a powerful tool for early identification. However, mental health data is highly personal, and protecting it is crucial. FL addresses this by enabling training across multiple devices without ever sharing the raw data. In this study, we introduce a federated learning framework designed specifically to detect depression in Marathi speakers. The framework combines Natural Language Processing (NLP) for analyzing text and audio processing techniques for studying speech patterns. Using a Marathi dataset that includes both speech and text samples from individuals with and without depression, we train local models on individual devices. These models are then combined into a global model, which is continuously improved through a process called federated averaging. Our findings show that this FL-based approach performs well in detecting depression while keeping the data private and secure. This highlights the potential of FL in mental health applications, especially for languages like Marathi, where gathering and processing data centrally can be difficult. By prioritizing privacy, this work opens the door for future research into using federated learning for other regional languages and mental health challenges. The Federated Learning model outperforms the non-FL model, achieving around 97.9% across accuracy, precision, recall, and F1-score, compared to 97.4% without FL. with speech dataset the model demonstrates high parameter values of above 96.0%., This demonstrates FL’s effectiveness in improving performance on the text and speech based depression detection task.