Developing and validating a parser-based suicidality detection model in text-based mental health services.

Area of Science:

• Computational linguistics
• Mental health technology
• Natural Language Processing (NLP)

Background:

• Text-mining holds potential for enhancing online mental health services, particularly in identifying suicidality.
• False positives in automated suicidality detection pose a significant challenge for these services.

Purpose of the Study:

• To develop a novel parser-based algorithm (PBSD) for detecting suicidal ideation in text-based mental health services.
• To minimize false alarms while accurately identifying potential suicidality.

Main Methods:

• Utilized data from a 24/7 online text-based counseling service in Hong Kong (N=1267 sessions).
• Developed PBSD using sentence parsing to analyze grammatical structure and apply syntax rules for true/false positive classification.
• Compared PBSD against a standard keyword matching model using accuracy and recall metrics.

Main Results:

• PBSD significantly outperformed the baseline keyword model in accuracy (0.68 vs 0.53), a 28.3% improvement.
• The algorithm corrected 36.8% of false alarms generated by lexicon matching.
• A marginal reduction in recall was observed (1 vs 0.96).

Conclusions:

• The parser-based model substantially enhances lexicon-based methods by reducing false alarms and improving suicidality detection accuracy.
• This advancement can decrease unnecessary distress and disruption for frontline mental health service providers.