Mental-health risk detection seeks early signs of distress from social media posts and clinical transcripts to enable timely intervention before crises. When such risks go undetected, consequences can escalate to self-harm, long-term disability, reduced productivity, and significant societal and economic burden. Despite recent advances, detecting risk from online text remains challenging due to heterogeneous language, evolving semantics, and the sequential emergence of new datasets. Effective solutions must encode clinically meaningful cues, reason about causal relations, and adapt to new domains without forgetting prior knowledge. To address these challenges, this paper presents a Continual Neuro-Symbolic Graph Learning (CNSGL) framework that unifies symbolic reasoning, causal inference, and continual learning within a single architecture. Each post is represented as a symbolic graph linking clinically relevant tags to textual content, enriched with causal edges derived from directional Point-wise Mutual Information (PMI). A two-layer Graph Convolutional Network (GCN) encodes these graphs, and a Transformer-based attention pooler aggregates node embeddings while providing interpretable tag-level importances. Continual adaptation across datasets is achieved through the Multi-Head Freeze (MH-Freeze) strategy, which freezes a shared encoder and incrementally trains lightweight task-specific heads (small classifiers attached to the shared embedding). Experimental evaluations across six diverse mental-health datasets ranging from Reddit discourse to clinical interviews, demonstrate that MH-Freeze consistently outperforms existing continual-learning baselines in both discriminative accuracy and calibration reliability. Across six datasets, MH-Freeze achieves up to 0.925 accuracy and 0.923 F1-Score, with AUPRC $\ge 0.934$ and AUROC $\ge 0.942$, consistently surpassing all continual-learning baselines. The results confirm the framework’s ability to preserve prior knowledge, adapt to domain shifts, and maintain causal interpretability, establishing CNSGL as a promising step toward robust, explainable, and lifelong mental-health risk assessment.

Fine-grained dog breed classification presents significant challenges due to subtle inter-class differences, pose variations, and intra-class diversity. To address these complexities and limitations of traditional handcrafted approaches, a novel and efficient two-stage Deep Learning (DL) framework tailored for robust fine-grained classification is proposed. In the first stage, a lightweight object detector, YOLO v8N (You Only Look Once Version 8 Nano), is fine-tuned to localize both the head and full body of the dog from each image. In the second stage, a dual-stream Vision Transformer (ViT) architecture independently processes the detected head and body regions, enabling the extraction of region-specific, complementary features. This dual-path approach improves feature discriminability by capturing localized cues that are vital for distinguishing visually similar breeds. The proposed framework introduces several key innovations: (1) a modular and lightweight head–body detection pipeline that balances accuracy with computational efficiency, (2) a region-aware ViT model that leverages spatial attention for enhanced fine-grained recognition, and (3) a training scheme incorporating advanced augmentations and structured supervision to maximize generalization. These contributions collectively enhance model performance while maintaining deployment efficiency. Extensive experiments conducted on the Tsinghua Dogs dataset validate the effectiveness of the approach. The model achieves an accuracy of 90.04%, outperforming existing State-of-the-Art (SOTA) methods across all key evaluation metrics. Furthermore, statistical significance testing confirms the robustness of the observed improvements over multiple baselines. The proposed method presents an effective solution for breed recognition tasks and shows strong potential for broader applications, including pet surveillance, veterinary diagnostics, and cross-species classification. Notably, it achieved an accuracy of 96.85% on the Oxford-IIIT Pet dataset, demonstrating its robustness across different species and breeds.