Falls prediction using the nursing home minimum dataset.

Area of Science:

• Gerontology
• Health Informatics
• Clinical Prediction Modeling

Background:

• Falls are a significant risk for nursing home residents, leading to injury and increased healthcare costs.
• Existing fall risk assessment tools often lack accuracy or rely on data not readily available in electronic health records.
• Predictive modeling using electronic data offers a scalable solution for proactive fall prevention in long-term care settings.

Purpose of the Study:

• To develop and validate a predictive model for fall risk in nursing home residents.
• To utilize electronically available data from over 15,000 US facilities for model development.
• To create a model with improved performance characteristics for fall prediction.

Main Methods:

• A hybrid Classification and Regression Tree (CART)-logistic approach was employed.
• The model was trained and tested on two data extracts (2011-2013 and 2016-2018) from the Long-term Care Minimum Dataset (MDS).
• Drug data from five skilled nursing facilities were integrated with MDS data.

Main Results:

• The final dataset included 3985 residents (mean age 77 years, 64% female).
• The model achieved an area under the ROC curve of 0.668 on the validation subsample.
• Antidepressant medications showed a protective association against falls in specific resident subgroups.

Conclusions:

• The developed hybrid CART-logit algorithm outperforms 22 previously evaluated fall risk tools.
• This model offers better performance for a fall prediction window of ≤90 days.
• It is uniquely designed to use features easily obtainable across nearly all US nursing facilities, enabling widespread implementation.