Developing a prediction model for cognitive impairment in older adults following critical illness.

Abstract

Background: New or worsening cognitive impairment or dementia is common in older adults following an episode of critical illness, and screening post-discharge is recommended for those at increased risk. There is a need for prediction models of post-ICU cognitive impairment to guide delivery of screening and support resources to those in greatest need. We sought to develop and internally validate a machine learning model for new cognitive impairment or dementia in older adults after critical illness using electronic health record (EHR) data.

Methods: Our cohort included patients > 60 years of age admitted to a large academic health system ICU in North Carolina between 2015 and 2021. Patients were included in the cohort if they were admitted to the ICU for ≥ 48 h with ≥ 2 ambulatory visits prior to hospitalization and at least one visit in the post-discharge year. We used a machine learning model, oblique random survival forests (ORSF), to examine the multivariable association of 54 structured data elements available by 3 months after discharge with incident diagnoses of cognitive impairment or dementia over 1-year.

Results: In this cohort of 8,299 adults, 22% died and 4.9% were diagnosed with dementia or cognitive impairment within one year. The ORSF model showed reasonable discrimination (c-statistic = 0.83) and stability with little difference in the model's c-statistic across time.

Conclusion: Machine learning using readily available EHR data can predict new cognitive impairment or dementia at 1-year post-ICU discharge in older adults with acceptable accuracy. Further studies are needed to understand how this tool may impact screening for cognitive impairment in the post-discharge period.