347A LATE BREAKING ABSTRACT: Homicidal and suicidal ideation as a chief complaint: considering the influence of news events using a large, de-identified electronic medical record dataset

Background: In the wake of nationally reported tragedies such as school shootings, pediatric mental health providers anecdotally report an increase in patient reports of homicidal or suicidal ideation. However, it is challenging to quickly assess if these trends are present in the general patient population and thus worthy of additional prevention research. As such, we explore toolkits and techniques that provide rapid, deidentified pilot data using an electronic medical record (EMR) database.

Methods: To assess if there was an increase in reported suicidal or homicidal ideation in our pediatric patients following the Florida Parkland tragedy (2-14-18), our EMR was probed using an integrated tool which automatically deidentifies and provides filters to “slice” the dataset for research purposes (e.g. queries, inclusion/exclusion criteria). We began with the initial database of 5,328,007 individual patients (all records in our health system, past year to date). This number lowered to 747,486 by applying the age filter of 0-19 years of age, reflecting the range of school-age youth. A “chief compliant” filter was then applied, using ‘suicidal or suicide attempt or suicidal ideation’ (the three options available in our EMR), resulting in 802 cases. Graphing data by week showed variability with a range of 10-40 case per week. To allow direct comparison by time range, filters were applied to parse from February 14 (the date of the event) through the date of the query (March 6), for both 2017 and 2018.

Results: The mean number of pediatric patients with the chief complaint of ‘suicidal or suicide attempt or suicidal ideation’ per week during the time range (2/14 to 3/6) was 23.67 (SD 4.51) and in 2018 was 33.00 (SD 3.46); the difference was significant: p = 0.046, t = 2.8430, df = 4 (95% CI -18.45 to -.22). We then propose an automated pipeline to identify major suicide risk factors (e.g. traumas) in clinical notes using Natural Language Processing (NLP) and Machine Learning (ML) within large EMR datasets. The suggested pipeline includes: 1) querying for target keywords; 2) manually annotating n number of sentences containing those keywords to identify if patients are expressing risk factors; 3) automating this process using NLP and ML; and 4) applying the automated models to a large set of data within EMR.

Conclusions: Using an integrated toolkit, we identified a signal in our EMR demonstrating increased suicidal/homicidal ideation after a national tragedy. EMR toolkits, NLP and ML can be applied to deidentified EMR datasets, providing signals that are otherwise time-consuming to identify in traditional pilot studies for prevention research.