Mental Health Detection Using AI : From Social Media Texts Using NLP

Abstract

A comparative study on Artificial intelligence (AI) was conducted, focused on an adaptable and discrete detection of severe mental health related issues, especially given digital health priorities in contexts like Finland. The feasibility of using Natural Language Processing (NLP) to classify suicidal thoughts from 232,074 social media posts was investigated to identify the best model architecture for this important task. The methodology involved comparing a traditional TF-IDF/Logistic Regression baseline which achieved 93.0% accuracy against a specialized Bidirectional Encoder Representations from Transformers (BERT) model fine-tuned on a GPU. The BERT model achieved an overall test accuracy of 97.9% demonstrating a significant five-percentage-point improvement and a highly balanced F1-score of 0.98 for suicide detection. The results validated the study's main hypothesis, confirming that the advanced contextual understanding of the Transformer architecture was significantly more effective than linear models. It was concluded that AI models can reliably and ethically serve as high-accuracy triage systems, offering robust support to professionals, provided ethical safeguards concerning bias and non-diagnostic function are strictly maintained.