Introducing 3DCNN ResNets for ASD full-body kinematic assessment: a comparison with hand-crafted features

TL;DR

Findings show that end-to-end models enable less variable and context-independent ASD classification without requiring domain knowledge or task specificity, however, they also recognize the effectiveness of hand-crafted features in specific task scenarios.

Abstract

Autism Spectrum Disorder (ASD) is characterized by challenges in social communication and restricted patterns, with motor abnormalities gaining traction for early detection. However, kinematic analysis in ASD is limited, often lacking robust validation and relying on hand-crafted features for single tasks, leading to inconsistencies across studies. End-to-end models have emerged as promising methods to overcome the need for feature engineering. Our aim is to propose a newly adapted 3DCNN ResNet from and compare it to widely used hand-crafted features for motor ASD assessment. Specifically, we developed a virtual reality environment with multiple motor tasks and trained models using both approaches. We prioritized a reliable validation framework with repeated cross-validation. Results show the proposed model achieves a maximum accuracy of 85$\pm$3%, outperforming state-of-the-art end-to-end models with short 1-to-3 minute samples. Our comparative analysis with hand-crafted features shows feature-engineered models outperformed our end-to-end model in certain tasks. However, our end-to-end model achieved a higher mean AUC of 0.80$\pm$0.03. Additionally, statistical differences were found in model variance, with our end-to-end model providing more consistent results with less variability across all VR tasks, demonstrating domain generalization and reliability. These findings show that end-to-end models enable less variable and context-independent ASD classification without requiring domain knowledge or task specificity. However, they also recognize the effectiveness of hand-crafted features in specific task scenarios.

Figures (4)

• Figure 1: Visual representation of the proposed methodology. Top box represents the processing pipeline, from raw data to the data used for both the feature engineering and end-to-end approaches. Bottom left box represents the feature extraction process and the machine learning models used. Bottom right box represents the video generation process from raw data, the data augmentation process and our proposed end-to-end model.
• Figure 2: Representation of the validation strategy. Top boxes show our cross validation strategies for both the end-to-end and feature engineered models. Bottom right represents our voting system for the ensemble model, which combines task-specific model predictions. Bottom left depicts the pairwise statistical analysis used for model comparison.
• Figure 3: Mean ROC curves and standard deviations (highlighted) across models for all folds and games.
• Figure 4: Levene differences across models independently of the game