Text mining to improve screening for trauma-related symptoms in a global sample

Previous studies showed that textual information could be used to screen respondents for posttraumatic stress disorder (PTSD). In this study, we explored the feasibility of using language features extracted from short text descriptions respondents provided of stressful events to predict trauma-related symptoms assessed using the Global Psychotrauma Screen. Texts were analyzed with both closed- and open-vocabulary methods to extract language features representing the occurrence of words, phrases, or specific topics in the description of stressful events. We also evaluated whether combining language features with self-report information, including respondents’ demographics, event characteristics, and risk factors for trauma-related disorders, would improve the prediction performance. Data were collected using an online survey on a cross-national sample of 5048 respondents. Results showed that language data achieved the highest predictive power when both closed- and open-vocabulary features were included as predictors. Combining language data and self-report information resulted in a significant increase in performance and in a model which achieved good accuracy as a screener for probable PTSD diagnosis (.7 < AUC ≤ .8), with similar results regardless of the length of the text description of the event. Overall, results indicated that short texts add to the detection of trauma-related symptoms and probable PTSD diagnosis.

Highlights

•  Text-mining was applied to short textual descriptions of stressful events.

•  Extracted language features were examined as predictors of trauma-related symptoms.

•  Combining language and self-report data on the event resulted in the best predictions.

•  The tested model showed promising results in screening for probable PTSD diagnosis.