Development of a Machine Learning-Based Model to Predict Timed-Up-and-Go Test in Older Adults.

Abstract

The measurement of physical frailty in elderly patients with orthopedic impairments remains a challenge due to its subjectivity, unreliability, time-consuming nature, and limited applicability to uninjured individuals. Our study aims to address this gap by developing objective, multifactorial machine models that do not rely on mobility data and subsequently validating their predictive capacity concerning the Timed-up-and-Go test (TUG test) in orthogeriatric patients. We utilized 67 multifactorial non-mobility parameters in a pre-processing phase, employing six feature selection algorithms. Subsequently, these parameters were used to train four distinct machine learning algorithms, including a generalized linear model, a support vector machine, a random forest algorithm, and an extreme gradient boost algorithm. The primary goal was to predict the time required for the TUG test without relying on mobility data. The random forest algorithm yielded the most accurate estimations of the TUG test time. The best-performing algorithm demonstrated a mean absolute error of 2.7 s, while the worst-performing algorithm exhibited an error of 7.8 s. The methodology used for variable selection appeared to exert minimal influence on the overall performance. It is essential to highlight that all the employed algorithms tended to overestimate the time for quick patients and underestimate it for slower patients. Our findings demonstrate the feasibility of predicting the TUG test time using a machine learning model that does not depend on mobility data. This establishes a basis for identifying patients at risk automatically and objectively assessing the physical capacity of currently immobilized patients. Such advancements could significantly contribute to enhancing patient care and treatment planning in orthogeriatric settings.