Background: Altered gait patterns, such as toe walking (TW), are common in various clinical conditions, including Autism Spectrum Disorder (ASD), where TW affects about 20-30% of individuals. Quantification of TW can provide valuable insights into gait abnormalities and guide interventions. A recently proposed video-based coding protocol allows TW quantification in natural settings but requires manual video review by an operator. However, a markerless gait analysis system (MGAS) could address the challenges of non-acceptance in individuals with ASD or intellectual disabilities, enabling automated kinematic analysis and detection of uncommon gait patterns like toe walking with minimal operator involvement during data processing. This approach could improve the efficiency and accuracy of gait assessments in such populations.

Objectives: This study aimed to develop a novel method for detecting toe-walking using the OpenPose module.

Methods: The present study involved 10 subjects with neurotypical development (aged 4-18 years; 7 males) who simulated toe-walking and 2 individuals with ASD (age: 5 and 8,1 years; 2 males), diagnosed according to DSM-5 criteria and presenting toe-walking. All participants were administered a validated video-based coding protocol based on standardized video recordings to quantify toe-walking. The test consisted of transporting an object (e.g., puzzle, Lego®) from a therapist to a playing table located 3 meters away and back again 10 times. All tests were conducted without shoes, albeit with socks. The subjects with neurotypical development were tested once, while the individuals with ASD were tested three times on three different days (6 acquisitions), resulting in 16 tests. Sagittal-plane video recordings of the walking trials were collected based on a setup described in the literature. Using the OpenPose module, the skeletons were extracted in the sagittal plane, and gait kinematics were computed. A two-tailed T-test was applied to identify the most relevant features: we selected features at p-value test <0.05 as inputs for the classifiers. Then, these selected features were then used to train supervised neural networks (NNs) to evaluate accuracy, sensitivity, specificity, and the area under the curve (AUC).

Results: Finally, 1140 normal and 648 tip-toe gait samples were used to train three neural network classifiers. We identified 12 relevant features, summarized in Table 1. The classifiers were trained using these selected features and a leave-one-out cross-validation process, repeated 19 times. Classifier performance metrics (mean and standard deviation for accuracy, sensitivity, and specificity) were evaluated using Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Bayesian classifier, with z-score normalization. The results, summarized in Table 2, showed that the SVM classifier obtained the best results.

Conclusions: The proposed method offers the potential for automated detection and quantification of toe walking in standardized protocols and settings. This approach provides detailed kinematic data with a fully non-intrusive method, not requiring the patient’s instrumentation. The following research steps will be dedicated to the study of a larger court of subjects to further validate the findings of this preliminary study.

Notes: 

(1)Scientific Institute IRCCS “E.Medea”, Bosisio Parini, Lecco, Italy, (2)Autism Research Unit, Villa Santa Maria Institute, Tavernerio, Como, Italy, (3)School of Medicine and Surgery, University of Milano-Bicocca, Milano, Milano, Italy, (4)PhD Program in Neuroscience, University of Milano-Bicocca, Milano, Milano, Italy, (5)Autism Research Unit, Villa Santa Maria SCS, Tavernerio, Como, Italy, (6)Department of Design, Politecnico di Milano, Milan, Italy