Attending Physician Shirley Ryan Ability Lab, Nothwestern McGaw Chicago, Illinois, United States
Objectives: Falls are a significant concern for lower limb prosthetic users (LLPUs).Current fall risk assessments, such as the Timed Up and Go (TUG) test, have limited predictive utility and are often validated on narrow subsets of LLPUs. This study aims to develop more accurate tools for predicting falls in LLPUs.
Design: LLPUs of various amputation levels who could walk 10 meters with or without assistance were included. Individuals with recent significant injury or cognitive impairment were excluded. Functional assessments included the TUG, Four Square Step Test (FSST), Activities-Specific Balance Confidence (ABC) Scale, and Prosthetic Limb Users Survey of Mobility (PlusM). T-tests and ROC curve analysis were used to assess the relationship between functional measures and fall status. Machine learning models, including random forest and extreme gradient boosting, were used to combine functional assessments with amputation level to predict retrospective fall status.
Results: Data from 50 LLPUs were analyzed. No significant differences were found between fallers and non-fallers on the ABC scale (p=0.157) or PlusM score (p=0.313). The TUG test showed a significant difference between fallers and non-fallers (p=0.034), with an AUC of 0.69. The FSST did not show significant differences, with an AUC of 0.67. Machine learning improved prediction, with an AUC of 0.47 and 0.75 for random forest and extreme gradient boosting, respectively. Including amputation level further improved the models, resulting in AUCs of 0.6 and 0.8, respectively.
Conclusions: While the TUG test showed statistical significance, none of the functional measures alone effectively predicted falls in LLPUs. Aggregate of functional measurements with machine learning method perform slightly better than individual measure. Including amputation levels in machine learning models improved accuracy, but even the best model struggled to accurately classify fallers versus non-fallers. In the future we will explore if incorporation of kinematics captured during these assessments can improve fall risk prediction
