JMIR Publications

ABSTRACT

Background:

Neurodegenerative diseases (NDD) are common among older adults worldwide. Early diagnosis of NDD is a difficult but rewarding task. Gait status has been shown to be an indicator of changes in the early stages of NDD and can play a significant role in diagnosis, treatment, and rehabilitation. In the past, assessments of gait either relied on complex but imprecise scales by trained doctors or required patients to wear additional equipment that made them uncomfortable. Advances in artificial intelligence may change this completely and provide a new approach to gait evaluation.

Objective:

This study sought to use cutting-edge machine learning outcomes to (1) provide patients with a non-invasive gait assessment in a completely non-contact manner, (2) provide healthcare professionals with accurate gait assessment results covering all common gait-related parameters to assist them in diagnosis and rehabilitation planning.

Methods:

The data collection included motion data from 41 different subjects aged 25-85 years, with an average age of 57.51 years, captured in motion sequences with the Kinect (a 3D camera with a sampling frequency of 30HZ). To identify the type of gait in each frame of the walk, we used Support Vector Machine (SVM) and Bi-directional Long short-term memory (Bi-LSTM) classifiers that were trained by spatio-temporal features extracted from the raw data. Then, the gait semantics could be obtained from the frame labels. Finally, all gait parameters can be calculated according to gait semantics. To ensure the best generalization performance of the model, classifiers were trained under a 10-fold cross-validation strategy. Furthermore, the proposed algorithm is also compared with the previous best heuristic method. The qualitative and quantitative feedback from the medical staff and patients in the actual medical scenarios are also widely collected for usability analysis.

Results:

Evaluations were made for 3 aspects. For the two classifiers classification result, the Bi-LSTM achieved an average precision, recall, and F1-score of 90.54%, 90.41%, and 90.38% separately, while these metrics of SVM are 86.99%, 86.62%, and 86.67%. Moreover, the Bi-LSTM-based method attained an accuracy of 93.2% when taking the gait segmentation evaluation (tolerance is set to 2) while that of the SVM-based method was only 77.5%. For the final gait parameter calculation result, the average error rate of the heuristic, SVM and Bi-LSTM is 20.91%, 5.85%, and 3.17% respectively, and the average standard deviation error of them are 24.69%, 5.45%, and 2.75%.

Conclusions:

This study proves that the Bi-LSTM-based approach can provide strong support for the accurate assessment of gait parameters to help medical staff to make early diagnoses and reasonable rehabilitation plans for patients with NDD.