COMPARISON BETWEEN TWO METHODS OF GAIT ANALYSIS IN SHALLOW WATER: MANUAL KINEMATIC ANALYSIS AND DEEP LEARNING NEURAL NETWORK

L. de Liz Alves¹, A. Ivaniski-Mello¹, M. Zimmermann Casal¹, L.A. Peyré-Tartaruga¹, F. Gomes Martinez¹

¹Federal University of Rio Grande do Sul, Physical Education, Physiotherapy and Dance Department, Porto Alegre, Brazil

Background: Shallow water walking is a common intervention that sougth to reduce cardiorespiratory deficits and movement disorders. Artificial intelligence in rehabilitation is becoming increasingly most important as the methods advance, but applications in aquatic physical therapy are rare. The use of artificial intelligence in aquatic physical therapy has huge potential due to the inherent challenges of collecting and processing data from motion analysis systems in the aquatic environment.

Purpose: We aimed to examine if fully automated measurements of motion capture system using a markerless pose estimation based on transfer learning with deep neural networks are feasible and comparable with a conventional manual digitalization.

Methods: Fifteen healthy adult men walked at 0.4 m/s in the xiphoid immersion level immersion. 100 steps from 10 participants were used to train the network and 50 steps from 5 participants were used totest the 2 methods. The following anatomical references were used for the analysis: fifth metatarsal, calcaneus, lateral malleolus, lateral epicondyle of the femur, greater trochanter of the femur and lateral of the trunk at the height of the xiphoid process. The dependent variables were: walking speed, stride frequency, stride length, duty factor and range of ankle, knee, hip and trunk motion. A Gopro™ camera (60Hz) recorded the sagittal plane of walking. For the MAN, the SkillSpector™ software was used. For the IAD, a deep neural network was trained using the DeepLabCut open-source method with 500 images during 300k interations. This trained network was used to digitalize the walking videos. Paired t test and Bland-Altman method with a simple linear regression was used to compare methods. All the data collected was processed on Matlab and the statistical procedures were performed on SPSS (α=.05).

Results: The IAD train was performed with a train error of 2,38 pixels and a test error of 4,34 pixels. The total time necessary for the data digitization by de MAN was 20 hours and 7 minutes (15 minutes per video) and for the IAD digitization was 14 minutes (16,6 seconds per video). No significant differences between the methods were found for all spatiotemporal variables, stride speed (p= 0.91), stride frequency (p = 0.73), stride length (p = 0.21), and for the range of motion from ankle (p = 0.55) and hip (p= 0.05) joints. The knee and trunk range of motion was higher in the IAD method (p< 0.001). For the Bland Altman test, all the spatiotemporal variables were statistically similar with significant linear regression values for the stride speed (p = 0.072), stride frequency (p = 0,056) and stride length (p = 0.043).

Conclusions: The present study shows promising results for the use of artificial intelligence in the underwater human movement, expressively reducing the digitization time. Fully automated measurements from motion capture systems based on artificial intelligence could facilitate the clinical implementation of visual and quantitative feedback in real setups.

Implications: Our results can contribute to the evaluating gait in shallow water as a physical activity, as well as indicating the feasibility of carrying out the evaluation of other methods through deep learning neural networks.

Funding acknowledgements: FOUNDATION TO SUPPORT RESEARCH OF RIO GRANDE DO SUL (FAPERGS) - BRAZIL.

Keywords:
Gait
Shallow water
Deep learning neural network

Topics:
Innovative technology: information management, big data and artificial intelligence
Musculoskeletal: lower limb
Health promotion & wellbeing/healthy ageing/physical activity

Did this work require ethics approval? Yes

Institution: Federal University of Rio Grande do Sul

Committee: Federal University of Rio Grande do Sul - Brazil

Ethics number: 37928