# Explainable AI for suicide risk detection: gender- and age-specific patterns from real-time crisis chats

## TL;DR

This study uses AI and chat data to identify patterns in suicide risk across different ages and genders, helping improve early detection and personalized prevention.

## Contribution

The novel use of explainable AI and a theory-driven lexicon to detect real-time suicide risk patterns in crisis chats, stratified by gender and age.

## Key findings

- Hopelessness and prior suicide attempts were strong predictors of suicide risk across all groups.
- Gender and age-specific patterns emerged, such as loneliness being a consistent predictor for women and thwarted belongingness for men.
- Age-specific factors like bullying and LGBTQ identity showed inverse associations with suicide risk in certain subgroups.

## Abstract

Suicide remains a leading cause of death worldwide, yet conventional risk models based on static demographic or diagnostic factors show limited predictive value. Advances in explainable artificial intelligence (AI) and natural language processing (NLP) offer new opportunities for real-time, personalized risk detection.

We analyzed 17,564 chat sessions (2017–2021) from Sahar, a digital crisis helpline. Suicide risk (SR) was defined by explicit suicidal ideation. A theory-driven lexicon of 20 psychological constructs (e.g., hopelessness, loneliness, self-harm), derived from leading SR frameworks, was applied using NLP. Logistic regression models estimated associations between constructs and SR, stratified by gender and age (10–17, 18–20, 21–40, and 41+). Temporal trajectories of predictors were examined across five conversation stages.

Previous suicide attempts and hopelessness were the strongest predictors across all groups. Gender differences emerged: among women, loneliness was a consistent predictor, whereas in men, thwarted belongingness and late-session depression were more salient. Age analyses showed developmental specificity: prior attempts were strongest in adolescents, hopelessness and self-harm peaked in young adults, thwarted belongingness strengthened with age, and loneliness predicted risk only among adults aged 41+. Several factors, including bullying/cyberbullying, LGBTQ identity, and perfectionism, were inversely associated with SR in specific subgroups.

This study demonstrates how explainable, theory-informed NLP can capture dynamic SR factors in real-world crisis interactions. Findings reveal distinct gender- and age-specific pathways, underscoring the need for personalized prevention strategies. Beyond theoretical insights, the approach highlights the potential of AI-driven, interpretable monitoring tools to support crisis counselors in detecting escalating risk earlier and tailoring interventions. Such methods can enhance the accuracy, timeliness, and equity of suicide prevention, and illustrate how explainable AI can translate psychological theory into clinically meaningful tools for mental health screening and early intervention.