Enhancing suicidal behavior detection in EHRs: A multi-label NLP framework with transformer models and semantic...

Area of Science:

• Computational linguistics
• Clinical informatics
• Public health

Background:

• Suicide is a significant global health concern, necessitating early identification of at-risk individuals.
• Existing Natural Language Processing (NLP) methods for suicide risk assessment in Electronic Health Records (EHRs) have limitations in capturing the full spectrum of suicidal behaviors.

Purpose of the Study:

• To develop a novel NLP approach for fine-grained, multi-label classification of suicidal behaviors in EHRs.
• To create a multi-class labeled dataset with comprehensive annotation guidelines to improve NLP accuracy and annotation efficiency.

Main Methods:

• A multi-class labeling system was developed, categorizing six types of suicidal behaviors with multi-label capability.
• An MPNet-based semantic retrieval framework was employed to extract relevant EHR sentences, followed by expert annotation.
• Transformer-based models were fine-tuned on the curated dataset for multi-label classification of suicidal behaviors.

Main Results:

• Lexical analysis identified key themes in suicide risk assessment, including personal history, mental health, and substance use.
• Fine-tuned models, particularly Bio_ClinicalBERT, BioBERT, and XLNet, achieved an F1 score of 0.81 in identifying suicidal behaviors.
• The multi-label classification system effectively captured the complexity of suicidal behaviors, outperforming traditional binary classification.

Conclusions:

• A robust NLP framework was established for detecting suicidal behaviors in EHRs using fine-tuned transformer models and a semi-automated pipeline.
• The study highlights the potential of advanced NLP in improving the identification of suicidal behaviors.
• Future research should focus on expanding and integrating these models to enhance patient care and clinical decision-making.