Suicide Risk Assessment of Psychiatric Inpatients by Natural Language Processing

Abstract

Suicidal act of patients is one of the most serious issues in psychiatry. Every medical staff should ideally always be alert for patients’ mental condition and share it among the colleagues, but in fact that is not an easy task. We contrived a system which automatically assesses the suicidal risk of patients and help staffs to share it. We selected approximately 170 patients who showed suicidal acts while hospitalized. Then we examined their medical records and extracted about 500 texts which were thought to suggest high suicidal risk and about 800 texts suggesting low suicidal risk. After that, we applied naïve bayes method to those data and inspected the power of classification. It demonstrated the accuracy of better than 80% in both 10-fold cross validation and hold-out. By setting this classification algorithm into electronic medical record systems, sensitivity of medical staffs to patients’ suicidal risk might be improved.