Identifying Unexpected Deaths in Long-Term Care Homes

Objectives: Predicting unexpected deaths among long-term care (LTC) residents can provide valuable information to clinicians and policy makers. We study multiple methods to predict unexpected death, adjusting for individual and home-level factors, and to use as a step to compare mortality differences at the facility level in the future work. Design: We conducted a retrospective cohort study using Resident Assessment Instrument Minimum Data Set assessment data for all LTC residents in Ontario, Canada, from April 2017 to March 2018. Setting and Participants: All residents in Ontario long-term homes. We used data routinely collected as part of administrative reporting by health care providers to the funder: Ontario Ministry of Health and Long-Term Care. This project is a component of routine policy development to ensure safety of the LTC system residents. Methods: Logistic regression (LR), mixed-effect LR (mixLR), and a machine learning algorithm (XGBoost) were used to predict individual mortality over 5 to 95 days after the last available RAI assessment. Results: We identified 22,419 deaths in the cohort of 106,366 cases (mean age: 83.1 years; female: 67.7%; dementia: 68.8%; functional decline: 16.6%). XGBoost had superior calibration and discrimination (Cstatistic 0.837) over both mixLR (0.819) and LR (0.813). The models had high correlation in predicting death (LR-mixLR: 0.979, LR-XGBoost: 0.885, mixLR-XGBoost: 0.882). The inter-rater reliability between the models LR-mixLR and LR-XGBoost was 0.56 and 0.84, respectively. Using results in which all 3 models predicted probability of actual death of a resident at <5% yielded 210 unexpected deaths or 0.9% of the observed deaths. Conclusions and Implications: XGBoost outperformed other models, but the combination of 3 models provides a method to detect facilities with potentially higher rates of unexpected deaths while minimizing the possibility of false positives and could be useful for ongoing surveillance and quality assurance at the facility, regional, and national levels. Crown Copyright (C) 2021 Published by Elsevier Inc. on behalf of AMDA - The Society for Post-Acute and Long-Term Care Medicine.

Automated summary

The study focuses on predicting unexpected deaths among long-term care (LTC) residents and compares multiple methods for prediction. The XGBoost model outperforms other models in calibration and discrimination. The combination of three models can help monitor mortality differences at the facility level and is important for quality assurance and surveillance.