Pediatric Language Impairment requires timely and consistent screening to enable early intervention and to improve long-term developmental outcomes. Nevertheless, current methods struggle to extract clinically relevant biomarkers from high-dimensional and heterogeneous neuroimaging data, particularly when multimodal analyses and interpretable modeling are not involved. To address these limitations, this paper presents XAI-AttnFusionNet+, an explainable multimodal deep learning model for early-stage Language Impairment screening with early childhood resting-state functional magnetic resonance imaging (rs-fMRI). The proposed model integrates spatial patterns of brain activation, dynamic functional connectivity, and clinical phenotypic characteristics through a hierarchical attention-based fusion mechanism enabling adaptive modality weighting and improved representation learning. To enhance transparency and clinical interpretability, post hoc explainability methods, such as attention visualization, Grad-CAM, and SHAP, were employed to identify discriminative brain regions and feature contributions. The framework was evaluated on a multicenter pediatric cohort of 255 subjects based on ABIDE II using a Leave-One-Subject-Out (LOSO) validation protocol and further evaluated for cross-dataset generalization using ABIDE I. In the absence of Language Impairment annotations, proxy labels were constructed based on phenotypic language indicators. The performance of XAI-AttnFusionNet+ with an accuracy of 93.3% and an AUC of 0.968 outperforms the traditional unimodal baseline models and demonstrates strong cross-site generalization. The results indicate the potential of the proposed framework as a neuroimaging-based approach for early detection of Language Impairment risk. However, its deployment in real-world settings requires further validation using clinically annotated datasets.