Quantitative analysis of gait and balance using deep learning on monocular videos and the timed up and go test in idiopathic normal-pressure hydrocephalus

TL;DR

This study uses deep learning on monocular videos to analyze gait and balance in idiopathic normal-pressure hydrocephalus patients, showing strong predictive power for fall risk.

Contribution

A vision-based gait analysis system using monocular videos and deep learning is validated for predicting fall risk in INPH patients.

Findings

TUG scores were negatively correlated with gait velocity, cadence, and stride length.

An automated machine learning model achieved an area under the curve of 0.979 for predicting falling risk.

Gait velocity was identified as the most important predictor of falling risk using SHapley Additive exPlanations.

Abstract

A vision-based gait analysis system using deep learning algorithms for simple monocular videos was validated to estimate temporo-spatial gait parameters in idiopathic normal-pressure hydrocephalus (INPH) patients. The Timed Up and Go (TUG) test has been used to reflect risk of falling in INPH patients. The aims of the study were (1) to investigate relationships between temporo-spatial gait parameters measured by a vision-based gait analysis system using monocular videos and TUG scores and (2) to determine whether an automated machine learning model based on these gait parameters could predict falling risk in INPH patients.

Gait data from 59 patients were collected from the vision-based system. All patients were also evaluated with the TUG test. A TUG time of ≥13.5 s was used as a cut-off to identify potential fallers.

Timed Up and Go scores were negatively correlated with gait velocity, cadence, stride length, and swing phase. TUG scores were positively correlated with step width, stride time, stance phase, double-limb support phase, stride time variability, and stride length variability. The area under the curve for predicting falling risk using the automated machine learning-based model was 0.979. We found that velocity was the most important factor in predicting falling risk with the interpretable method called SHapley Additive exPlanations.

This study identified important associations between gait parameters measured by vision-based gait analysis and TUG scores in INPH patients. An automated machine learning model based on gait parameters measured by vision-based gait analysis can predict falling risk with excellent performance in INPH patients. We suggest that our vision-based gait analysis method using monocular videos has the potential to bridge the gap between laboratory testing and clinical assessment of gait and balance in INPH patients.

Figures