Multilingual MLP system for early detection of mental health indicators

Abstract

This thesis explores the development of a multilingual Natural Language Processing (NLP) system designed to detect mental health indicators, specifically anxiety and depression, in textual data. Leveraging datasets in both English and Spanish, we employed a comprehensive feature set that includes lexical, phonetic, emotional, and transformer-based features to enhance detection performance. Our study reveals that the “All” approach, which combines these diverse features, consistently outperforms other models across various evaluation metrics, demonstrating superior accuracy, precision, recall, and F1-score. The inclusion of phonestemes, particularly in the Spanish dataset, provided additional insights and improved performance in specific contexts. The Random Forest classifier was identified as the most robust and effective, consistently delivering high performance across all feature combinations. The findings underscore the importance of using comprehensive and diverse feature sets in developing accurate and reliable NLP systems for mental health detection, with significant implications for future research and real-world applications.