Selection of Clinical Text Features for Classifying Suicide Attempts

Area of Science:

• Medical Informatics
• Computational Linguistics
• Psychiatry

Background:

• Traditional methods using ICD-9/10-CM diagnosis codes lead to misclassification in studies of suicidal thoughts and behaviors (STBs).
• Electronic Health Record (EHR) phenotyping offers a promising alternative for patient identification.
• Incorporating unstructured clinical text into EHR phenotyping is an emerging area of research.

Purpose of the Study:

• To evaluate the effectiveness of using natural language processing (NLP) and machine learning to identify STBs from unstructured clinical text.
• To develop and validate predictive models for STB phenotyping using features extracted from discharge summaries.

Main Methods:

• Utilized a publicly accessible NLP program (MetaMap) to process clinical text from 810 inpatient admissions.
• Employed iterative elastic net regression to extract and select predictive text features.
• Developed two phenotyping models based on the selected textual features.

Main Results:

• Reduced initial feature sets of 5,866 and 2,709 text features to 18 and 11, respectively.
• Achieved high performance metrics: Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.866-0.895.
• Attained strong performance: Area Under the Precision-Recall Curve (AUPRC) of 0.800-0.838.

Conclusions:

• The NLP and regression approach demonstrates significant potential for accurate STB phenotyping.
• Extracting features from unstructured clinical text can substantially improve patient identification in EHR data.
• This methodology offers a valuable tool for research on suicidal thoughts and behaviors.