ScAN: Suicide Attempt and Ideation Events Dataset.

Area of Science:

• Public Health
• Medical Informatics
• Computational Linguistics

Background:

• Suicide is a major global health issue, with suicidal behaviors like attempts (SA) and ideations (SI) being significant risk factors.
• Electronic Health Records (EHR) contain valuable information on patient SA and SI, crucial for surveillance and prevention.

Purpose of the Study:

• To develop a dataset (ScAN) and a model (ScANER) for accurately extracting and classifying suicidal behavior events from EHR notes.
• To improve the identification of patients at risk for suicide and support timely medical intervention.

Main Methods:

• Construction of the ScAN dataset using MIMIC III EHR notes, annotated for SA and SI events (19k+ events).
• Development of ScANER, a multi-task RoBERTa-based model with retrieval and prediction modules for analyzing EHR data.
• Evaluation of ScANER's performance in identifying suicidal behavioral evidence and classifying SA/SI events.

Main Results:

• ScAN dataset comprises over 19,000 annotated SA and SI events from 12,000+ EHR notes.
• ScANER achieved a macro-weighted F1-score of 0.83 for evidence retrieval and F1-scores of 0.78 (SA) and 0.60 (SI) for classification.
• Both the ScAN dataset and ScANER model are publicly available for research use.

Conclusions:

• Automated detection of suicidal behaviors from EHRs is feasible and can enhance suicide prevention strategies.
• The ScAN dataset and ScANER model provide valuable resources for advancing research in this critical public health area.