Natural language processing identified grief, divorce, financial distress, violence and family problems in 39.7 million clinical notes. The notes covered 202,988 patients in Central Finland from 2010 to 2023, with F1 scores of 0.81 to 0.95. Prevalence ranged from 0.5% for family problems to 22.1% for grief. Time-varying Cox models linked each extra event type with anxiety (HR 2.08), depression (HR 1.99) and stress-related disorders (HR 1.62). Generalised linear mixed models linked it with more comorbidity (rate ratio 1.58) and frailty (mean ratio 1.57). Violence and financial distress were recorded far below national estimates, so events are likely underrecorded.
Why it is interesting: Stressful life events are absent from structured health records but can be extracted reliably from narrative notes. Their association with frailty and comorbidity links psychosocial stress to physical health outcomes in routine care data.
Abstract — the authors' own words, CC BY
Abstract Background: Despite being harmful to both mental and physical health, stressful life events are not recorded in structured electronic health records. They appear sporadically in narrative clinical notes, where they are difficult to analyze at scale. Natural language processing (NLP) has made clinical notes more available for research, but applications to stressful life events remain limited, leaving their documentation patterns and prognostic value understudied. By applying NLP, we identified stressful life events from clinical notes and examined their associations with mental disorders, comorbidity, and frailty. Methods:
We analyzed 39.7 million clinical notes, recorded between 2010 and 2023, from 202,988 patients in Central Finland Wellbeing Services County. NLP was applied to identify stressful life events in five categories: grief, divorce, financial distress, violence, and family problems. Associations with the risk of mental disorders (stress-related disorders, anxiety, and depression) were estimated using time-varying Cox proportional hazards models, and associations with physical health outcomes (comorbidity and frailty) using generalized linear mixed models. Results:
Stressful life events were identified from clinical notes with high accuracy (F1 scores 0.81-0.95). Event prevalences ranged from 0.5% for family problems to 22.1% for grief. Most individual events were associated with mental disorders, comorbidity, and frailty. The number of different life events was associated with a higher risk of stress-related (hazard ratio [HR], 1.62; 95% confidence interval [CI], 1.58-1.67), anxiety (HR, 2.08; 95% CI, 2.03-2.14), and depressive disorders (HR, 1.99; 95% CI, 1.94-2.03), and with higher comorbidity (rate ratio, 1.58; 95% CI, 1.56-1.59) and frailty (mean ratio, 1.57; 95% CI, 1.56-1.58). Conclusions:
Stressful life events were accurately identified from clinical notes and were associated with a higher risk of mental disorders, comorbidity, and frailty. Even so, violence and financial distress, the two events with comparable national estimates, were documented substantially less often than those estimates, suggesting underrecording of stressful life events in clinical settings.
Cite this item
BibTeX
Free sign-up. A digest of the new papers twice a week, grouped by domain.