Detecting self-care challenges in children through enhanced ConvNeXt optimized by modifeied neural network algorithm

Early detection of self-care challenges in children is intricate and subjective, despite their implications on development and well-being. The proposed work exhibits an improved ConvNeXt model, which is fine-tuned using a modified variant of the Neural network algorithm that detects self-care difficulties from the multimodal Child Mind Institute - Healthy brain network (HBN) dataset. The model combines focus mechanisms, bespoke loss functions, and unique feature selection to outperform by combining behavioral, neuroimaging, physiological, and audio/video data. We achieve extensive detection accuracy (97.89%) in detecting nodules and outperforming existing state-of-the-art models (ViT and EfficientNet) for the same task, with increased robustness for all age groups and severity levels. Interpretability tools like SHAP values provide visual transparency in decision-making, which is critical alongside their clinical trust. Results show the model’s capacity to emphasize relevant attributes while combating overfitting through regularization and dropout techniques. Even with this process, the results are only applicable for a temporary period, this is because data in these jurisdictions is changing and spreading over time, and the needs of society are changing as well. This not only advances the field of pediatric care by providing a scalable, objective solution, but it also establishes a reference point for multimodal AI applications in healthcare. These results highlight the promise of using AI-powered tools to reshape early intervention techniques and enhance outcomes for children struggling with self-care.