Online Program

Return to main conference page
Thursday, May 17
Public Health/Disease
Thu, May 17, 10:00 AM - 10:45 AM
Regency Ballroom B
 

A Machine Learning Approach to Improve Fall Risk Prediction in Home Health Care (304720)

Kathryn H. Bowles, School of Nursing, University of Pennsylvania 
*Yancy Lo, Institute for Biomedical Informatics, The Perelman School of Medicine, University of Pennsylvania 
Selah F. Lynch, Institute for Biomedical Informatics, The Perelman School of Medicine, University of Pennsylvania 
Jason H. Moore, Institute for Biomedical Informatics, The Perelman School of Medicine, University of Pennsylvania 
Randal S. Olson, Institute for Biomedical Informatics, The Perelman School of Medicine, University of Pennsylvania 
Ryan J. Urbanowicz, Institute for Biomedical Informatics, The Perelman School of Medicine, University of Pennsylvania 

Keywords: fall risk prediction, electronic health records, big data, random forest

Falls are the leading cause of injuries among older adults, yet they are the top avoidable event. The risk of falls is higher among the more vulnerable home health care (HHC) population. Existing standardized fall risk assessments often require additional effort to collect and tend to have low specificity (13%). Using readily available data for all HHC patients from a comprehensive clinical, functional, behavioral, and environmental assessment, advanced modeling strategies can be harnessed to better profile and predict fall risk for HHC patients.
Our goal was to identify factors that predict and quantify fall risks using machine learning methods on a dataset of roughly 60,000 unique patients from a HHC database. Each patient is evaluated on a mandated Outcomes Assessment Information Set (OASIS), which provides rich detailed features profiling their condition upon admission to receive HHC services.
We applied a random forest algorithm to two features sets: 1. The 10-item Missouri Alliance for Home Care (MAHC-10) fall risk assessment that is commonly used in HHC, and 2. Potentially relevant features from the OASIS, curated by experienced clinicians. We compared precision, balanced accuracy, and area under the ROC curve (AUC) of our models against the baseline MAHC-10 scores. Further, we visualized risk factors and their contribution to the overall fall risk on a decision tree. Finally, we provided a ranked list of risk factors based on feature importance scores generated from the models.
Our study is the first known attempt to determine fall risk factors from the extensive OASIS results from a large sample. Our predictive model significantly outperforms the existing MAHC-10 scoring system, achieving higher specificity and higher AUC. Using our quantitative prediction of fall risks, clinicians can discuss precise risk factors and corresponding interventions with respect to each patient, thereby lowering the incidence of falls.