In a NIMH-funded, five-year study, researchers at Vanderbilt University Medical Center will use artificial intelligence (AI)-based, natural language analysis of EHRs to shed light on suicidal ideation and its relationship to attempted suicide, and to explore the genetic underpinnings of suicide risk.
Principal investigators for the study are Colin Walsh, M.D., assistant professor of biomedical informatics and psychiatry and behavioral sciences, and Douglas Ruderfer, Ph.D., assistant professor of genetic medicine, psychiatry and behavioral sciences and biomedical informatics.
“Many cases of suicidal thoughts and behaviors are missed if we rely on structured data alone.”
“Many cases of suicidal thoughts and behaviors are missed if we rely on structured data alone,” Walsh said. “In one study, suicidal thoughts were only coded three percent of the time even when documented in text in primary care. Text features extracted through natural language processing of physician notes, patient messages and more should allow us to improve our predictive algorithms by capturing more nuanced risk factors.”
Beyond Structured Data
Rates of suicidal behavior continue to increase in the U.S. (up approximately 30 percent from 1999 to 2017), which only elevates the importance of developing better tools for predictive analyses.
Walsh, Ruderfer and colleagues previously used EHR data and machine learning techniques to develop a predictive model of suicide attempt risk. They employed this algorithm across two biobanks – Vanderbilt’s BioVU, the world’s largest collection of human DNA stored at a single site, and UK Biobank – to assign suicide risk scores to thousands of genotyped patients.
While this previous study relied strictly on structured data such as health care billing codes to find high-risk patients, this time the researchers will work with Cosmin Bejan, Ph.D., assistant professor of biomedical informatics at Vanderbilt, to include EHR text in their analyses. Free text could help identify and predict suicidal ideation and suicide attempt, Ruderfer said.
“Expanding our capture of suicidal phenotypes will enable improved understanding of the genetic architecture of suicidal thoughts and behaviors and the contributing genetic and clinical risk factors.”
“Expanding our capture of suicidal phenotypes will enable improved understanding of the genetic architecture of suicidal thoughts and behaviors and the contributing genetic and clinical risk factors. While most individuals who consider suicide do not attempt it, we know ideation is an important risk factor that we should flag for follow-up.”
Measuring Genetic Risk
The group’s earlier research was notable in part for establishing that significant heritability for suicide attempt risk can be measured using genotype data. By combining genetic correlation, genomewide association and polygenic risk scoring, they found a high genetic correlation between attempted suicide and suicide attempt probability, in the 70 to 100 percent range.
The study presented a new strategy for measuring genetic risk when patient outcomes are less common or more difficult to pin down: instead of relying on large numbers of patients with the outcome of interest, use clinical data and a predictive model to assign a probability to every patient for that outcome.
“If you can show that the model’s performance and calibration are good, you can convert that to a quantitative measure for that outcome. Then, you don’t need to ascertain more cases to increase statistical power, you just need to increase your sample size,” Ruderfer said.
Related Research and What It Means
If successful, the newly funded work may clarify biologic risk factors for suicide and related psychiatric disorders. The team hopes that improvements in the identification of high-risk individuals considering suicide will contribute to the ultimate goal of preventing patients from acting on those thoughts. “The text-based approaches used here may identify those who might otherwise fall through the cracks,” Walsh said.
Ruderfer and Walsh have also received NIMH funding to apply a similar approach to identify patients that do not respond to standard treatments for depression – a major risk factor for suicide attempt.
