A Symbolic Inference System for AI Decision Support in Multimodal Assessment of Pediatric Cerebral Visual Impairment

Cerebral visual impairment (CVI) is a heterogeneous neurodevelopmental condition in which visual dysfunction may affect acuity, visual field, ocular motility, fixation, electrophysiological responses, and functional vision. Multimodal assessment is therefore required, but the resulting evidence is often incomplete, heterogeneous, and difficult to integrate reproducibly. This computational feasibility study developed a symbolic AI decision-support model for multimodal assessment of pediatric CVI. Sixty-two children older than 18 months and younger than 18 years were encoded as patient-level facts after neuro-ophthalmological assessment and brain and orbital MRI. Structured variables included MRI status, refraction, ocular motility, visual field testing, and visual evoked potentials. An anterior-pathway-not-explanatory predicate was used as a non-weighted eligibility gate, while a priori expert-defined support weights were assigned to positive evidence domains. MRI abnormalities and ocular motility abnormalities were each present in 47/62 patients (75.8%), visual field alterations in 16/62 (25.8%), and altered visual evoked potentials in 22/62 (35.5%). At least one predefined CVI-compatible functional proof path was established in 54/62 patients (87.1%). With 𝑘=1, CVI support scores ranged from 0.700 to 0.800, reflecting the strongest retained functional path. With 𝑘=3, the maximum support score was 0.985 in patients with three retained functional paths. Equal-weight ablation (𝑤=0.75 for each functional domain) preserved logical derivability and yielded a three-path value of 0.9844, indicating that multi-path convergence was not dependent on the small differences among the expert-assigned functional weights. No independent diagnostic gold standard was available; therefore, these findings characterize symbolic derivability and provenance behavior rather than diagnostic accuracy. The proposed system provides an explicit, provenance-aware decision-support representation and a foundation for future neurosymbolic integration of quantitative clinical and instrumental data.