E S Ikonnikova, A E Slotina, G A Belyankin, M A Zabello, V Yu Vishnyakov, A V Lobanov, N V Bedeneu, A A Dobrovolsky, D O Zemlyanaya, A Yu Tropynina, O A Kirichenko, A V Baidukova, N S Suponeva
The final model when analyzing video data achieved 76% accuracy in risk prediction, with an average absolute error of time parameter prediction of 1.445 s. The following parameters were identified as key biomechanical predictors of fall risk: step base width at the ankle joint level, lateral trunk sway, step base width at the knee joint level, vertical foot clearance during the swing phase, and temporal gait characteristics.
UNLABELLED: The aim of the study was to develop an artificial intelligence model for predicting the probability of falls based on the analysis of video recordings of gait patterns.
MATERIALS AND METHODS: The study involved 187 patients (median age 59 years) with neurological diseases of various etiologies who complained of impaired balance and unsteadiness while walking. All participants underwent video recording of their movements using a smartphone during a 10-meter walk test and Timed Up and Go test. Based on the analysis of 2077 steps recorded in 122 patients, an algorithm for fall risk stratification was developed using the YOLO-NAS Pose M architecture and a two-layer machine learning model.
RESULTS: The final model when analyzing video data achieved 76% accuracy in risk prediction, with an average absolute error of time parameter prediction of 1.445 s. The following parameters were identified as key biomechanical predictors of fall risk: step base width at the ankle joint level, lateral trunk sway, step base width at the knee joint level, vertical foot clearance during the swing phase, and temporal gait characteristics.