Identifying suicide documentation in clinical notes through zero-shot learning.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Clinical Informatics

Background:

• Identifying suicidality in clinical notes is challenging due to limited training data.
• A scarcity of positive instances hinders the development of deep learning models for suicide risk detection.

Purpose of the Study:

• To develop a novel methodology using zero-shot learning to identify suicidality in clinical notes.
• To address the data sparsity issue in deep learning for clinical note analysis.

Main Methods:

• Utilized US Veterans Affairs clinical notes, with labels derived from suicide attempt and self-harm diagnostic codes.
• Employed a deep neural network trained with a base string for auxiliary information, mapping document content to a semantic space.
• Compared performance against a baseline deep neural network using bag-of-words features.

Main Results:

• The zero-shot learning model significantly outperformed the baseline across various metrics, including area under the curve, sensitivity, and specificity.
• Achieved 94% accuracy in identifying suicidality documentation not linked to ICD-10-CM codes at a 0.90 probability threshold.

Conclusions:

• Zero-shot learning offers an effective approach to identify suicidality in clinical notes.
• This methodology enables accurate detection of suicide risk without the need for manual annotation, addressing data limitations.