BACKGROUND
Evidence for the best choice of fall risk assessment in long-term care facilities is limited. Short-term fall predictions may enable the implementation of dynamic care practices that specifically address changes in individualized fall risk within senior care facilities. This can be achieved through the use of electronic health records (EHRs), which contain routinely collected information regarding the majority of known fall risk factors.
OBJECTIVE
We implemented machine learning algorithms that use EHR data to predict a three-month fall risk in residents from a variety of senior care facilities providing different levels of care.
METHODS
This retrospective study obtained EHR data (2007-2021) from Juniper communities’ proprietary database of 2,785 individuals primarily residing in skilled nursing facilities, independent living facilities, and assisted living facilities across the United States. We assessed the performances of three machine learning (ML)-based fall predictions models and the Juniper Communities fall risk assessment across these different facilities. The ML input features included vital signs and several known risk factors, such as history of fall, comorbidities, and medications. These features were identified within the EHR system based on relevant International Classification of Diseases codes, string searches, or keyword queries. Additional analyses were conducted to examine how the changes in the input features, training datasets, and prediction window affected the performance of these models.
RESULTS
The extreme gradient boosting (XGB) model exhibited the highest performance with an area under the receiver operating characteristic curve (AUROC) of 0.846, specificity of 0.848, and sensitivity of 0.706 while achieving the best tradeoff in balancing true positive and negative rates. The number of active medications was the most significant feature associated with fall risk, followed by a resident's number of active diseases, and several variables associated with vital signs, including diastolic blood pressure and changes in weight and respiratory rates. The combination of vital signs with traditional risk factors as input features reached a higher prediction accuracy than using either group of features alone. When reducing the prediction window to two months, the XGB model continued to exhibit the highest performance (AUROC = 0.753) in comparison to logistic regression (AUROC = 0.690), multi-layered perceptron (AUROC = 0.678) and Juniper's fall risk assessment (AUROC = 0.582).
CONCLUSIONS
This study provides novel insights into EHR-based features for predicting short-term fall risk in different types of care facilities. The integration of EHR data into fall prediction models, and particularly vital signs, yields a cost-effective and automated fall risk surveillance. Our XGB model uncovered the impact of a wide range of clinical and pathophysiological fall predictors across heterogenous cohorts while outperforming traditional fall risk assessments and standard ML techniques that are less compatible with EHR data.
CLINICALTRIAL
N/A
Comparing Machine Learning Approaches for Fall Risk Assessment
