Machine-Learning-Based Prediction of Suicide Risk Using Preliminary Questionnaire and Consultation Text

Area of Science:

• Artificial Intelligence
• Mental Health Technology
• Computational Psychiatry

Background:

• Chat-based mental health services in Japan face challenges with low response rates due to understaffing.
• There is a critical need for efficient methods to assess suicide risk in real-time.
• Leveraging preliminary information alongside consultation text is crucial for accurate risk assessment.

Purpose of the Study:

• To propose and evaluate machine learning (ML) based methods for suicide risk classification.
• To determine the optimal combination of preliminary information (medical questionnaire - MQ) and consultation text (CT) for risk assessment.
• To enhance the prioritization of high-risk users in mental health services.

Main Methods:

• Development of five ML-based suicide risk classification methods.
• Construction of a dataset including MQ, CT, chat logs, and six-level risk assessments.
• Evaluation of methods using ROC-AUC, with a focus on the M3 approach outputting intermediate predictions for MQ and CT separately.

Main Results:

• The M3 method achieved the highest ROC-AUC of 0.879.
• Combining both MQ and CT data significantly outperformed using either data source alone.
• Suicidal ideation within the MQ was identified as a key predictive item.

Conclusions:

• ML methods, particularly the M3 approach, show significant promise in classifying suicide risk.
• Integrating diverse data sources (MQ and CT) is essential for robust suicide risk assessment.
• The proposed methods can effectively assist in prioritizing high-risk users, despite classification challenges in some cases.