Predicting Falls in Long-term Care Facilities: Machine Learning Study.

Area of Science:

• Gerontology
• Health Informatics
• Machine Learning

Background:

• Electronic health records (EHRs) can enable dynamic care practices for senior fall risk.
• Short-term fall prediction models are crucial for timely interventions in senior care.

Purpose of the Study:

• To implement machine learning (ML) algorithms using EHR data for predicting 3-month fall risk.
• To evaluate ML model performance across diverse senior care facility types.

Main Methods:

• Retrospective analysis of EHR data from 2785 individuals (2007-2021).
• Assessed 3 ML models and a standard fall risk assessment.
• Examined impact of input features, training data, and prediction windows.

Main Results:

• Extreme Gradient Boosting (EGB) model achieved the highest performance (AUC 0.846).
• Key predictors included active medications, number of diseases, and vital signs (diastolic blood pressure, weight changes).
• Combining vital signs with traditional factors improved prediction accuracy.

Conclusions:

• EGB models effectively predict short-term falls using extensive EHR data.
• Integrating vital signs enhances the accuracy of fall risk surveillance.
• ML models offer dynamic, automated, and cost-effective fall predictions for senior care.