Research Aims. The loss of a spouse is among life's most painful experiences, associated with significant distress and elevated risk for suicide. Concomitant persistent stress responses to bereavement may manifest as Prolonged Grief Disorder (PGD), which is further associated with increased risk for suicidal ideation and behaviors. However, the mechanisms by which the loss of a loved one leads to suicidal ideation remains unclear. Individual differences in emotional distress in response to loss-related memories may help identify those with increased suicide risk, but more research is needed.
To improve our assessment of potential suicide risk factors following loss, we will use machine learning methods to conduct a preliminary examination of stress markers in voice and language. We intend to apply advances in affective computing to a large archival dataset of recently-bereaved individuals, consisting of clinical assessments and audio-recording of semi-structured interviews. We will extract digital phenotypes of speech acoustic characteristics and content from the recordings to identify markers. We will then develop a predictive model and assess its accuracy in predicting future suicidal ideation as well as long-term adjustment. Further, we will validate these findings using an independent bereaved sample dataset to ensure their generalizability.
Methods. We will analyze archival data from 305 individuals who had recently lost a spouse and were followed longitudinally at 3-, 14-, and 25-months post-loss. At each time point, they were administered structured clinical interviews to assess PGD, MDD, and PTSD symptoms as well as C-SSRS suicidal ideation items. At 3- and 14-months, participants engaged in interviews about their loss, and other questions about their relationship with the deceased and non-loss topics.
Interview audio recordings will be analyzed to extract digital phenotypes. For linguistic content, we will apply a pre-trained Natural Language Processing model to the interview transcripts. For acoustic features, we will extract speech characteristics using phonetics analysis. We will then combine these markers in a model to assess prediction of suicidal ideation, as measured by the C-SSRS at 14- and 25-months. We will also assess the accuracy of the computational markers extracted at baseline (3-months post-loss) in predicting long-term adjustment trajectories (3-25 months). Lastly, we will use an auxiliary dataset of loss-narratives from 25 patients (6+ months post-loss) undergoing manualized complicated grief treatment, to independently evaluate model accuracy.
Hypotheses. We hypothesize that the algorithm will identify digital phenotypes from the semi-structured interviews, which will be able to predict suicidal ideation in bereaved individuals at 14- and 25-months post-loss, while controlling for other symptoms. We further hypothesize that the model will predict differences in long-term adjustment (as measured by symptom trajectories) using baseline data.
Impact. The proposed project will pioneer a new approach for the collection of information associated with suicidal thoughts from trauma narratives, and will provide a substantial proof-of-concept toward using computational methods with voice and language features for detecting risk factors that predict suicidal thoughts and adjustment difficulties after loss. This will provide the groundwork for more empirical measurement and personalized medicine approaches to the assessment of suicide risk.
