Explorando Métodos de Aprendizado Semi-supervisionado e Inteligência Artificial Generativa na Identificação de Ideação Suicida em Textos Não-Clínicos

Abstract

Suicide remains a serious public health issue worldwide, and the early identification of suicidal ideation is essential for preventing fatal outcomes. In this context, the present work aimed to enhance the Boamente system, an artificial intelligence–based tool developed to support mental health professionals in detecting and monitoring suicidal ideation in non-clinical texts written in Brazilian Portuguese. Boamente operates by collecting user texts via a virtual keyboard on mobile devices, which are then sent to a web platform for automatic analysis. To improve the system’s accuracy and reliability, two complementary studies were conducted. The first study aimed to improve Boamente’s classification model performance through the application of semi-supervised learning techniques. New data were collected, and different approaches were tested, resulting in an accuracy improvement ranging from 2.39% to 4.30% compared to the baseline model. Notably, the self-learning method achieved the best overall performance. These results highlight the potential of combining labeled and unlabeled data when training models for sensitive tasks such as suicidal ideation detection. The second study proposed expanding Boamente’s architecture by incorporating an explanation generation module, using prompt engineering with Large Language Models (LLMs). This addition aimed to increase the system’s interpretability, enabling mental health professionals to understand the reasons why a text was classified as containing or not containing suicidal ideation. For this purpose, a quantitative evaluation was conducted using various LLMs, in which the Qwen 2.5 model (14B) achieved the best AUC score (0.9898), while the 3B and 7B versions of Qwen 2.5 achieved the highest recall values a critical metric for minimizing false negatives in clinical contexts. In parallel, a qualitative assessment was conducted with experts from different fields (computer science, linguistics, and psychology), who rated the explanations generated by the LLaMA 3.1 (8B) model as the most coherent and clear. The results demonstrate that integrating semi-supervised learning methods with explanatory LLMs can significantly improve the quality of Boamente, both in terms of technical performance and perceived trustworthiness among professionals. Furthermore, explanation generation has the potential to strengthen user trust in the AI system, preventing blind acceptance of predictions and promoting more critical and informed decision-making. This work, therefore, contributes to the development of more transparent and effective artificial intelligence systems to meet the demands of mental health care, particularly in the context of suicide prevention in Brazil.