Identification of suicidal behavior among psychiatrically hospitalized adolescents using natural language processing and machine learning of electronic health records

• DOI: 10.1371/journal.pone.0211116
• Publication Date: 2019-02-19T18:25:47Z
• Authors: Nicholas J. Carson; Brian Mullin; Maria Jose Sanchez; Frederick Lu; Kelly Yang; Michelle Menezes; Benjamin Lê Cook
• Keywords: Medical record

Objective The rapid proliferation of machine learning research using electronic health records to classify healthcare outcomes offers an opportunity address the pressing public problem adolescent suicidal behavior. We describe development and evaluation a algorithm natural language processing identify behavior among psychiatrically hospitalized adolescents. Methods Adolescents on psychiatric inpatient unit in community system northeastern United States were surveyed for history suicide attempt past 12 months. A total 73 respondents had available prior index admission. Unstructured clinical notes downloaded from year preceding Natural identified phrases associated with outcome. enriched this group clinically focused list terms representing known risk protective factors then applied random forest develop classification model. model performance was evaluated sensitivity, specificity, positive predictive value (PPV), negative (NPV), accuracy. Results final sensitivity 0.83, specificity 0.22, AUC 0.68, PPV 0.42, NPV 0.67, accuracy 0.47. mostly highly clustered around related suicide, family members, disorders, psychotropic medications. Conclusion This analysis demonstrates modest success approach identifying small sample adolescents setting.