Dynamic suicide topic modelling: Deriving population-specific, psychosocial and time-sensitive suicide risk variables from Electronic Health Record psychotherapy notes

Abstract

In the machine learning subfield of natural language processing, a topic model is a type of unsupervised method that is used to uncover abstract topics within a corpus of text. Dynamic topic modelling (DTM) is used for capturing change in these topics over time. The study deploys DTM on corpus of electronic health record psychotherapy notes. This retrospective study examines whether DTM helps distinguish closely matched patients that did and did not die by suicide. Cohort consists of United States Department of Veterans Affairs (VA) patients diagnosed with Posttraumatic Stress Disorder (PTSD) between 2004 and 2013. Each case (those who died by suicide during the year following diagnosis) was matched with five controls (those who remained alive) that shared psychotherapists and had similar suicide risk based on VA's suicide prediction algorithm. Cohort was restricted to patients who received psychotherapy for 9+ months after initial PTSD diagnoses (cases = 77; controls = 362). For cases, psychotherapy notes from diagnosis until death were examined. For controls, psychotherapy notes from diagnosis until matched case's death date were examined. A Python-based DTM algorithm was utilized. Derived topics identified population-specific themes, including PTSD, psychotherapy, medication, communication and relationships. Control topics changed significantly more over time than case topics. Topic differences highlighted engagement, expressivity and therapeutic alliance. This study strengthens groundwork for deriving population-specific, psychosocial and time-sensitive suicide risk variables.

Keywords: dynamic topic models; electronic medical records; natural language processing; suicide prediction.

FIGURE 1

Quantitative method of topic mapping between derived case and control topics. We first identified the top 20 words (see Table 2 for words) in each topic from the first month of treatment and then used Word2Vec to map these words’ similarities across topics. Respective topics are numerated as follows: Topic 0 = treatment; topic 1 = medication; topic 3 = engagement; topic 4 = expressivity; topic 4 = symptomology. Thicker blue lines are indicative of increased topic similarity.

FIGURE 2

Word changes over time within cases (top row) and controls’ (bottom row) identified topics, respectively. Each topics’ selected words (the same words presented in Table 2) were derived using dynamic topic modelling. X axis = time (0 is index, 11 is suicide for cases/end of observation for controls); Y axis = prominence (lower values are more prominent).

FIGURE 3

Spearman correlations over time for cases (top row) and controls’ (bottom row) identified topics at index and end of treatment year (X axis: index month; Y axis: month of suicide for cases/end of observation for controls). Each word is presented at its associated coordinate; 95% confidence interval (CI) was calculated using Bonett and Wright’s recommended method for non-parametric data (2000). We determined that one Spearman coefficient was statistically significantly greater based on lack of CI overlap. We determined that one Spearman coefficient was statistically significantly greater based on lack of CI overlap (see Table 3) for CI.