A Multi-Modal Deep Learning Framework for Chromosomal Abnormality Diagnosis Using Neuro-Inference Predictive Network

Abstract

Trisomy 21 (Down syndrome) is one of the most prevalent chromosomal abnormalities, requiring accurate and early diagnosis for effective clinical intervention. Conventional diagnostic approaches often rely on either chromosome image analysis or facial phenotype assessment independently, limiting their ability to exploit complementary genomic and phenotypic information. Existing multimodal methods also suffer from inadequate cross-domain feature alignment, inefficient heterogeneous feature fusion, limited interpretability, and suboptimal optimization, resulting in reduced prediction accuracy and poor generalization. To address these limitations, this study proposed a Fusion-Based Prediction and Optimization comprising the Neuro-Inference Predictive Network (NeIPN) and the Meta-Gradient Fusion Optimizer (MeGFO). NeIPN integrated chromosome and facial classification outputs using adaptive correlation mapping, probabilistic attention, cross-domain feature fusion, and hierarchical inference to generate a unified latent representation for Trisomy 21 prediction. Subsequently, MeGFO refined the prediction through meta-learning-based gradient adaptation, dynamic fusion weight optimization, and learning parameter fine-tuning to minimize prediction error and improve model generalization. Experimental evaluation demonstrated that the proposed framework achieved an accuracy of 97.21 with an average loss of 0.2237, outperforming conventional multimodal prediction approaches. The proposed fusion and optimization strategy provides a reliable, interpretable, and scalable decision-support framework for accurate automated Trisomy 21 detection and has strong potential for future clinical implementation.