Identification potential of cognitive-motor dual-task gait in frailty via machine learning model

The early diagnosis and intervention of frailty play a crucial role in enhancing the quality of life for elderly individuals in their later years. Currently, the identification of frailty relies on various manual assessment scales, which are time-consuming and pose significant challenges to clinical practice.

A total of 220 participants were recruited to collect Timed Up-and-Go Test (TUGT) data, as well as single-task (ST) and cognitive-motor dual-task gait parameters. The modified Frailty Index-11 (mFI-11) scale and cognitive function assessments were completed. Machine learning (ML) methods were employed to screen gait parameters and construct a frailty diagnostic model. The receiver operating characteristic (ROC) curve was utilized to evaluate the correlation between significant gait indicators identified in the optimal model and cognitive frailty.

Among the 220 participants, the numbers of individuals classified as non-frailty, pre-frailty, and frailty were 83, 88, and 49, respectively. Within these groups, 43 were cognitively normal, and 53 exhibited cognitive frailty. In the model utilizing gait features, the SVM - Linear Kernel model exhibited the best classification performance, with an accuracy of 64.09 % and an F1 score of 64.38 %. Among the gait parameters, TUGT contributed significantly to the model while also demonstrating high predictive value for cognitive frailty (AUC = 0.8293, P < 0.0001).

The integration of cognitive-motor dual-task gait parameters with ML methods demonstrates satisfactory overall accuracy for the tri-classification diagnosis of frailty, indicating potential for community screening and auxiliary clinical diagnosis. Cognitive frailty shows significant correlations with gait parameters, particularly those assessed by TUGT.