Ocular-induced abnormal head posture (AHP) arises as a compensatory response to eye misalignment and plays a critical role in maintaining binocular vision. Early diagnosis is essential to prevent secondary complications, including facial asymmetry and musculoskeletal strain. However, current clinical assessments remain largely subjective and are further complicated by incomplete or fragmented medical records. To address these challenges, this study propose two complementary deep learning frameworks. First, AHP-CADNet proposes a clinically informed hierarchical attention architecture that integrates ocular landmarks, head pose features, and structured clinical attributes within a unified multi-task framework. The model employs intra-modal self-attention, gated cross-modal relevance modeling, and adaptive global modality weighting to capture dependency-aware relationships across modalities and support interpretable predictions. Evaluated on the PoseGaze-AHP dataset, AHP-CADNet achieves overall F1-scores of 92.9% (eye laterality), 91.8% (eye misalignment), 94.8% (diagnosis), 94.5% (AHP type), and 95.3% (AHP direction). For prism diopter estimation, it attains a mean absolute error of 3.316 prism diopters (R2 = 0.752; correlation = 0.882). Second, a masked curriculum learning–based imputation framework is proposed to address incomplete clinical documentation. Unlike conventional imputation approaches that assume static missing inputs, the framework applies task-aligned structured masking with progressively increasing difficulty, enabling robust reconstruction under realistic sparsity patterns. By leveraging domain-specific contextual representations from PubMedBERT to recover structured clinical attributes, the approach integrates unstructured textual knowledge with structured variables in a dependency-aware manner. The framework achieves predictive accuracies of 93.46%–99.78%, demonstrating reliable recovery of diagnostically relevant information under clinically plausible missing-data conditions.