Assessing text input from telegram for early suicide risk detection

Author

Mutete, Elton.

Title

Assessing text input from telegram for early suicide risk detection

Abstract

This work addresses the central world problem of suicide prevention by leveraging innovative artificial intelligence methods, emphasizing the crossing cultural and language gaps in low-resource settings. The research constructed and evaluated a multilingual suicide risk evaluation system grounded on AfriBERTa architecture, optimized for a specially prepared dataset of 15,000 labeled English posts across mental health forums and professionally certified Shona translations. The system was accurate to 88.11% in classification and stable across risk categories and performed particularly well in detecting culturally-specific distress expressions through specialized adaptation strategies. Practical deployment achieved real-time processing by integrating Telegram bot and ONNX optimization, with deploy ability towards deployment in bandwidth-constrained environments. While the results confirm AI as scalable mental health surveillance, the study also identifies primary areas of improvement, namely sensitivity for identifying high-risk cases. The paper delivers methodological contributions in natural language processing and operational models for ethically deploying AI mental health technologies cross-culturally. Scaling up high-risk training sets and latency minimization to response times in clinical settings are crucial to be addressed by future research.

Date

August 2025

Publisher

BUSE

Keywords

Suicide

Telegram

Text Input

Risk Detection

Supervisor

Mr. C. Zano,