A Reproducible Model Based on Clinical Text for Predicting Suicidal Behavior.

Area of Science:

• Computational psychiatry
• Clinical informatics
• Natural language processing

Background:

• Accurate identification of suicide risk is crucial for timely clinical intervention.
• Existing methods for suicide risk assessment have limitations in scalability and real-time application.

Purpose of the Study:

• To develop and validate a reproducible approach for training text classifiers to identify patients at risk of suicide.
• To evaluate the effectiveness of these models in phenotyping suicidal behavior and predicting future suicidal events.

Main Methods:

• Utilized a reproducible machine learning pipeline to train text classifiers on clinical notes.
• Employed natural language processing (NLP) techniques to extract relevant features from patient data.
• Evaluated model performance using standard metrics, including F1-score.

Main Results:

• The developed text classifiers demonstrated high effectiveness in phenotyping suicidal behavior, achieving an F1-score of 0.94.
• The models showed moderate effectiveness in predicting future suicidal events, with an F1-score of 0.63.

Conclusions:

• Reproducible text classification models can effectively identify patients with suicidal behavior.
• These NLP-driven approaches offer a promising tool for enhancing suicide risk assessment and patient prioritization in clinical settings.