Bridging local and global features: A multi-backbone ensemble framework for ASD facial classification
TL;DR
Empirical results validate that bridging high-frequency local textures with low-frequency global contextual relationships minimizes predictive variance, offering an accurate, stable, and scalable computational screening solution for automated clinical environments.
Abstract
Early non-invasive screening technologies are a paramount priority in modern healthcare for identifying complex neurodevelopmental traits characterized by social, communicative, and behavioural challenges. Recent breakthroughs in computer vision and deep learning have established automated facial image analysis as a highly viable paradigm for objective clinical screening. This study introduces a robust, optimized weighted ensemble framework that integrates the complementary architectural strengths of convolutional and transformer-based networks for binary classification of these specialized facial trait profiles. The pipeline concurrently leverages EfficientNet-B5 for localized feature scaling, Data-Efficient Image Transformers (DeiT) for long-range global self-attention, and ConvNeXt for modernized, high-performance convolutional representations. To ensure generalization and counteract dataset selection bias, a strict stratified 5-fold cross-validation scheme is enforced, followed by an optimized out-of-fold weighted probability fusion mechanism. Experimental evaluation on a benchmark dataset demonstrates that the unified ensemble achieves a state-of-the-art classification accuracy of 95.67% and an ROC-AUC of 0.9788, significantly outperforming individual standalone baselines. These empirical results validate that bridging high-frequency local textures with low-frequency global contextual relationships minimizes predictive variance, offering an accurate, stable, and scalable computational screening solution for automated clinical environments.