Login / Signup

Toward Reduced Burden in Evidence-Based Assessment of PTSD: A Machine Learning Study.

Tammy JiangSunny DutraDaniel J LeeAnthony J RoselliniGabrielle M GauthierTerence M KeaneJaimie L GradusBrian P Marx
Published in: Assessment (2020)
Structured diagnostic interviews involve significant respondent burden and clinician administration time. This study examined whether we can maintain diagnostic accuracy using fewer posttraumatic stress disorder (PTSD) assessment questions. Our study included 1,265 U.S. veterans of the Afghanistan and Iraq conflicts who were assessed for PTSD using the Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders, fifth edition (SCID-5). We used random forests to assess the importance of each diagnostic item in predicting a SCID-5 PTSD diagnosis. We used variable importance to rank each item and removed the lowest ranking items while maintaining ≥90% accuracy (i.e., efficiency), sensitivity, and other metrics. We eliminated six diagnostic items among the overall sample, four items among male veterans, and six items among female veterans. Our findings demonstrate that we may shorten the SCID-5 PTSD module while maintaining excellent diagnostic performance. These findings have implications for potentially reducing patient and provider burden of PTSD diagnostic assessment.
Keyphrases
  • posttraumatic stress disorder
  • machine learning
  • social support
  • risk factors
  • climate change
  • deep learning
  • big data