Researchers mine EHRs to predict suicide

Researchers combined data from electronic health records (EHRs) with results from self-report questionnaires to create a model that predicts the risk of suicide substantially better than existing prediction tools, according to the results of a study published in the American Journal of Psychiatry.

Prediction models cannot replace clinical judgment, the researchers acknowledged, but risk scores can help practitioners make better-informed clinical decisions.

“We believe these risk prediction tools are now accurate enough to help clinicians identify people at high risk and to help health systems reach out to people at risk who miss or cancel appointments,” said lead author Gregory Simon, MD, MPH, psychiatrist and senior investigator, Kaiser Permanente Washington Health Research Institute, Seattle, WA.

Statistics show that half of people who die by suicide and two-thirds of those who survive suicide attempts received a mental health diagnosis or treatment during the previous year. This suggests an opportunity for doctors to identify people at risk for suicide and to intervene before they act.

For this study, Dr. Simon and colleagues used anonymized EHR data from nearly 3 million patients who received a mental health diagnosis at either a primary care or a mental health clinic visit between January 2009 and June 2015. Patients were at least age 13 years or older (mean age: 46 years; 62% female).

The researchers sought to identify potential predictive characteristics present in EHRs up to 5 years before each visit, including prior suicide attempts, mental health and substance use diagnoses, medical diagnoses, prescriptions for psychiatric medications, inpatient or emergency department care, and recorded scores on depression questionnaires. Health system records identified 24,133 probable suicide attempts and state mortality records identified 1,240 suicide deaths within 90 days of a visit.

After the researchers analyzed the data, they found that the strongest predictors of suicide or suicide attempt included mental health diagnoses, substance use diagnoses, use of mental health emergency and inpatient care, and history of self-harm.

Dr. Simon and colleagues developed prediction models for both mental health clinic visits and primary care visits. Mental health specialty visits with risk scores in the top 5% accounted for 43% of subsequent suicide attempts and 48% of suicide deaths. Primary care visits with scores in the top 5% accounted for 48% of subsequent suicide attempts and 43% of suicide deaths.

“This prediction model was more accurate than previous models using health records,” Dr. Simon said. “For example, people with risk scores in the highest 5% accounted for almost half of suicide attempts, compared to about one-third with previous models.”

The overall accuracy of the new model also exceeded that of models used to predict other health conditions, such as rehospitalization for heart failure, in-hospital sepsis mortality, and high emergency department use. Further, the model’s predictive value was either equal to or greater than many widely accepted prediction tools for major medical outcomes, the researchers noted.

Accurate risk stratification can better inform providers’ and health systems’ decisions about frequency of follow-up, referral for intensive treatment, or outreach following missed or canceled appointments, concluded Dr. Simon and coauthors.

This research was supported by the National Institute of Mental Health.