Tool Development
The JHFRAT for predicting in-hospital fall risk was published in 2005 and updated in 2007. The tool was based on literature, tested in potential patient scenarios, and adjusted based on group consensus. The JHFRAT instrument contains 7 assessments with associated points: age (0–3 points); fall history (0 or 5); mobility (0, 2, 4 or 6); cognition (0–7); elimination/ bowel/ urine (0, 2 or 4); patient care equipment (0–3); and medication (0, 3, 5 or 7). Total scores of 6-13 are considered moderate-risk and a score of 14 or more is considered high-risk.10,11
Epic systems introduced the ROF cognitive computing model in 2018. The model is intended to improve on the performance of the traditional manual assessments, while requiring minimal additional documentation from clinicians. The model responds to information entered in the EHR to provide an up-to-date risk assessment for falls to clinicians. Per Epic, while the statistical performance of the model is good, the primary benefit is with the ROF is the automation of the fall risk assessment, saving up to 3 minutes per patient per day. In its standard configuration scores < 35 is low risk and ≥ 70 is high risk.12
The model includes demographic variables, as well as 4 categories of clinical variables: vital signs and assessments, lab results, medications, and procedure orders. Missing values for binary variables and counts were imputed as zero, assuming the patient did not meet the criteria for the variable. Of note:
Records for patients who experienced a fall were marked as positive for the entire stay.12
The medication variables represent a count of medications in specific pharmaceutical classes administered in the last 24 hours of the admission.12
The lab variables represent the most recent lab values collected within the 72 hours of discharge.12
To address hospital falls, Kinometrix created K-FRAS, a proprietary, predictive machine learning model that uses over 45 variables, including demographic, procedural, clinical, assessment, and physiological data to provide objective and accurate fall-risk predictions. The K-FRAS was developed using 500,000 patient records, XG-boost decision tree, and Bayesian parameter tuning. Of note:
Records for patients who experienced a fall were marked as positive only for a specific period around the fall occurrence.
All variables were included for the entirety of a patient stay in determined interval periods, allowing the model to account for longitudinal patient data occurring throughout a patient’s stay.
The K-FRAS can integrate with any EHR and provide a risk prediction without any additional clinical documentation required.
