Identifying Unexpected Deaths in Long-Term Care Homes

Area of Science:

• Gerontology and Public Health
• Health Informatics
• Epidemiology

Background:

• Predicting unexpected deaths in long-term care (LTC) facilities is vital for resident safety and quality improvement.
• Identifying at-risk individuals can inform clinical decisions and policy development.
• Existing methods require evaluation for accuracy in predicting mortality events.

Purpose of the Study:

• To compare the predictive performance of logistic regression (LR), mixed-effect LR (mixLR), and XGBoost for unexpected deaths in LTC residents.
• To identify reliable models for detecting facilities with potentially higher rates of unexpected mortality.
• To establish a foundation for future comparisons of facility-level mortality differences.

Main Methods:

• Retrospective cohort study utilizing Resident Assessment Instrument Minimum Data Set (RAI MDS) data from Ontario, Canada (April 2017-March 2018).
• Application of LR, mixLR, and XGBoost algorithms to predict individual mortality within 5 to 95 days post-assessment.
• Analysis of a cohort of 106,366 LTC residents, including 22,419 deaths.

Main Results:

• XGBoost demonstrated superior calibration and discrimination (C-statistic 0.837) compared to mixLR (0.819) and LR (0.813).
• All models showed high correlation in predicting death (LR-mixLR: 0.979).
• A combined model approach identified 210 unexpected deaths (0.9% of observed deaths) with low false positives.

Conclusions:

• XGBoost offers enhanced predictive accuracy for unexpected deaths in long-term care settings.
• Combining multiple predictive models can improve the detection of facilities with elevated unexpected mortality rates.
• These models can support ongoing surveillance and quality assurance initiatives at multiple administrative levels.