EHR-based Model Predicts Suicide Risk in Mental-health Outpatients
By Anne Harding
NEW YORK—Electronic health records can be used to predict suicidal behavior after an outpatient mental-health visit, new findings show.
"We can identify people at risk probably more accurately than we can identify risk for a lot of other adverse health events," such as the risk of stroke in patients with atrial fibrillation or the risk of hospital readmission with congestive heart failure, Dr. Gregory E. Simon of Kaiser Permanente Washington Health Research Institute, in Seattle, told Reuters Health by phone.
While suicide-risk detection "across health care" has been recommended, Dr. Simon and his team note in the American Journal of Psychiatry, online May 24, "traditional clinical detection of suicide risk is hardly better than chance."
To develop their prediction model, the researchers analyzed data from seven health systems on more than 2.9 million patients 13 years and older who had 10.3 million specialty mental-health visits and 9.7 million primary-care visits with a mental-health diagnosis in 2009-2015.
During the 90 days following each visit, there were 24,133 suicide attempts and 1,240 suicide deaths. The authors used 65% of the visits for model training and 35% for validation.
For both the specialty mental-health and the primary-care visits, the authors found, prior suicide attempt, mental-health and substance-abuse diagnoses, thoughts of death or self-harm "nearly every day" on item 9 of the Patient Health Questionnaire (PHQ-9), and previous inpatient or emergency mental health treatment were strongly associated with suicide attempts and death by suicide.
C statistics for predicting suicide attempts were 0.851 for mental-health specialty visits and 0.853 for primary care visits, with no difference between training and validation samples. For predicting suicide death, c-statistics were 0.861 and 0.833, respectively.
The risk of suicidal behavior was more than 200 times as high for those in the top 1% for risk based on the model compared to those in the bottom 50% based on predicted risk.
The factors associated with suicidal behavior are exactly what clinicians would use in determining whether a patient seeking outpatient mental health care is at risk, Dr. Simon noted.
"What these machine learning models do better than a human being can do, is they can think about 150 things simultaneously, so they can come up with a more accurate prediction," he added.
"We are not saying that this calculation would stand alone. The idea is you would deliver the results of this to a human being, who would then be able to act on it. We don't believe these calculations will replace human beings, we believe that they would be helpful to human beings."
Am J Psychiatry 2018.
(c) Copyright Thomson Reuters 2018. Click For Restrictions - https://agency.reuters.com/en/copyright.html