Identifying ICU survivors and relatives with post-traumatic stress disorder using text mining: An explorative study

Purpose: A quarter of ICU-patients develop post-traumatic stress disorder (PTSD) after discharge. These patients could benefit from early detection of PTSD. Therefore, we explored the accuracy of text mining with self-narratives to identify intensive care unit (ICU) patients and surviving relatives at risk of PTSD in a pilot study. Methods: In this prospective cohort study with self-administered questionnaires, discharged ICU-patients and surviving relatives participated. In a single centre study at a 32-bed ICU of a large teaching hospital, we used an online screening tool with self-narratives, to identify ICU-patients and surviving relatives at risk of PTSD using text mining. Study variables were Trauma Screening Questionnaire (TSQ) and self-narratives, administered 3 to 6 months after ICU discharge. Results: Of the participants 15% had an indication of PTSD based on TSQ. The median length of the self-narratives was 101 words. Using self-narratives, PTSD was predictable with a reasonable performance (AUROC of 0.67), compared to TSQ as gold standard. The most important words of the prediction model were ‘happen’ ‘again’ and ‘done’. These words are difficult to interpret without context. Conclusions: It is possible to predict risk of PTSD for ICU-patients and surviving relatives using text mining applied on self-narratives, 3 to 6 months after ICU discharge. The model performance is reasonable and helps to identify patients and surviving relatives at risk. Implications for Clinical Practice: Based on the large proportion of participants with an indication for PTSD, it remains important to persuade patients and surviving relatives to seek help when experiencing mental health problems after discharge.