[Paper Review] Clinical Parameters Prediction for Gait Disorder Recognition

Soheil Esmaeilzadeh, Ouassim Khebzegga+1 more|arXiv (Cornell University)|May 22, 2018

Gait Recognition and Analysis1 references3 citations

TL;DR

This paper proposes a dual-method approach to predict clinical gait parameters—cadence, step length, and walking speed—for gait disorder recognition using clinical measurements and video-based joint coordinate analysis. Using SVM and neural networks on OpenPose-extracted kinematic features, the model achieves an R² of 0.50 for walking speed prediction, outperforming linear regression and demonstrating the viability of video-based gait analysis for clinical decision support.

ABSTRACT

Being able to predict clinical parameters in order to diagnose gait disorders in a patient is of great value in planning treatments. It is known that extit{decision parameters} such as cadence, step length, and walking speed are critical in the diagnosis of gait disorders in patients. This project aims to predict the decision parameters using two ways and afterwards giving advice on whether a patient needs treatment or not. In one way, we use clinically measured parameters such as Ankle Dorsiflexion, age, walking speed, step length, stride length, weight over height squared (BMI) and etc. to predict the decision parameters. In a second way, we use videos recorded from patient's walking tests in a clinic in order to extract the coordinates of the joints of the patient over time and predict the decision parameters. Finally, having the decision parameters we pre-classify gait disorder intensity of a patient and as the result make decisions on whether a patient needs treatment or not.

Motivation & Objective

• To predict key clinical gait parameters—cadence, step length, and walking speed—for gait disorder diagnosis.
• To evaluate the performance of machine learning models using both clinically measured parameters and video-derived joint coordinates.
• To develop a data cleaning pipeline for multi-view video data with occlusions, shadows, and multiple subjects.
• To assess whether video-based kinematic features can reliably predict gait parameters comparable to clinical measurements.
• To support clinical decision-making by pre-classifying gait disorder severity and treatment necessity.

Proposed method

• Collected clinical data from Gillette Children’s Specialty Healthcare, including ankle dorsiflexion, BMI, walking speed, step length, stride length, and percent stance.
• Extracted 18-joint 2D coordinates from lateral and front-view videos using OpenPose, storing them as JSON files with x, y, and confidence (c) values.
• Implemented a data cleaning strategy based on center of gravity tracking using stable joints (0,1,2,5,8,11,14,15,16,17) to separate valid patient data from shadows or multiple people.
• Normalized joint coordinates relative to hip-neck distance to reduce camera distance bias.
• Applied Fast Fourier Transform (FFT) to extract frequency-domain features, computing power density over 12 intervals per ankle (24 features total).
• Trained and compared linear regression, SVM with RBF kernel, and feedforward neural networks on the extracted features, using R² and MSE as evaluation metrics.

Experimental results

Research questions

• RQ1 Can machine learning models accurately predict gait parameters (cadence, step length, walking speed) from clinical measurements alone?
• RQ2 To what extent can video-based kinematic features extracted via OpenPose predict the same gait parameters with comparable accuracy?
• RQ3 How do different machine learning models—linear regression, SVM, and neural networks—perform in predicting gait parameters from video-derived features?
• RQ4 Why does the SVM model achieve higher R² for walking speed (0.50) than cadence (0.36), despite cadence being inherently a frequency-based parameter?
• RQ5 What challenges arise in video-based gait analysis due to shadows, multiple people, or camera occlusion, and how can they be mitigated?

Key findings

• The SVM model achieved the highest R² of 0.50 for walking speed prediction on the test set, significantly outperforming linear regression (R² = -0.60).
• For cadence prediction, the SVM achieved an R² of 0.36, while the neural network reached only 0.18 at 2000 epochs, indicating limited performance despite high training epochs.
• The neural network showed severe overfitting for step length and speed, likely due to model architecture being optimized for cadence only.
• Linear regression performed poorly overall, with negative R² values for step length (-0.20) and speed (-0.60), indicating poor predictive power.
• Video-based feature extraction via FFT and OpenPose enabled reliable prediction of gait parameters, though performance was inferior to clinical data-based models.
• Data cleaning based on center of gravity tracking successfully separated valid patient data from shadows and multiple subjects, as evidenced by clean coordinate time series in Fig. 5(c).