(PS14-61) Using Machine Learning to Examine the Effects of Worry, Secondary Stressors, and Other Known Risk Factors on Grief Severity

Worry is a natural concomitant of bereavement that is triggered by the uncertainty of the future and is associated with post-loss adjustment problems. After the loss of a loved one, worry may further increase in response to secondary stressors (SS), or newly surfaced situations resulting from the loss that cause additional burdens (e.g., financial problems, deterioration of relationships, or increased responsibilities). Perceived stressfulness of SS have been positively related to worry and shown to predict prolonged grief and depressive symptoms. Despite existing research suggesting the importance of these variables, neither worry nor SS have been adequately examined within the field of grief. 

This study aimed to use a machine learning-based random forest regression model to determine if increases in worry and SS would be significant predictors of the intensity of grief responses when examined amongst other known predictors of grief severity. To our best knowledge, this study is the second to examine the role of both worry and SS on grief, and the first to use machine learning to do this. Study participants included 428 adults who lost a loved one within 1 to 12 months of completing the survey. Results showed that 56% of participants were female, the median age was 37, and 77% lost their parents. Participants completed questionnaires on demographics, worry, SS, grief intensity, and other known predictors of grief. 

The random forest regression model randomly selected a subsample of the participants in the total sample which was iteratively further split repeatedly into a specified number of subgroups to assess the contribution of predictors in reducing the Root Mean Square Error and maximize the fit during a training phase. The resultant model was then validated by assessing the model fit in the remaining subsample of the total sample. Five folds and five repetitions were specified in this analysis. The total sample was randomly split so that 70% of the sample was used during the training phase and 30% for the testing phase. The final model accounted for 68% of the variance in grief severity. SS, pandemic-related stress, and worry were among the most important predictors of grief severity. On the whole, indices of adjustment, variables related directly to the deceased (dependence, amount of contact, happiness with the relationship), and loneliness were the strongest predictors of grief severity in this model. The other factors related to the death (who died, days since death, expectedness of death), demographic variables, and other factors related to adjustment to loss were much less important. 

These results add to existing literature that suggests the importance of worry and SS. Given that worry is malleable and responsive to change, further research is needed to improve treatment for bereaved individuals by incorporating unregulated worry in response to SS as a treatment target. Amongst bereaved populations where SS may be particularly higher, such as historically marginalized communities, acknowledging their specific SS and circumstances in treatment planning will allow for the advancement of existing interventions and better serve vulnerable communities. Full results, limitations, and implications will be discussed.