Explainable neuro-symbolic artificial intelligence for automated interpretation of corneal topography and early keratoconus detection.

Background: Early detection of keratoconus is essential for preventing postoperative complications in refractive surgery and preserving long-term visual function. Although artificial intelligence has demonstrated strong potential in ophthalmic image analysis, many existing models operate as black-box systems and provide limited clinical interpretability. Transparent decision support is therefore critical for safe deployment of AI in clinical practice. Methods: We propose an explainable neuro-symbolic framework for automated interpretation of corneal topography reports and refractive surgery eligibility assessment. The proposed system integrates multimodal feature extraction, a symbolic corneal knowledge graph, probabilistic reasoning, and large language model (LLM)-based report generation. Quantitative biometric parameters and corneal curvature maps extracted from IOLMaster 700 reports were processed through a hybrid convolutional neural network-Vision Transformer (CNN-ViT) module to capture spatial corneal morphology. These representations were aligned with a clinically curated knowledge graph encoding relationships between corneal parameters, disease states, and surgical decision criteria. Bayesian probabilistic inference was then applied to estimate disease likelihoods, while an ensemble LLM module generated structured bilingual clinical reports explaining the reasoning process. Results: In a prospective pilot cohort of 20 eyes, the proposed framework demonstrated strong diagnostic performance for early keratoconus detection, achieving an area under the receiver operating characteristic curve (AUC) of approximately 0.95. Sensitivity and specificity remained high across decision thresholds, and the system achieved a balanced F1 score for refractive surgery eligibility classification. Expert evaluation indicated high interpretability and clinical usefulness of the generated reports. The end-to-end pipeline required approximately 95 ± 12 s per case, supporting near-real-time clinical decision support. Conclusion: The proposed neuro-symbolic framework combines deep representation learning, structured medical knowledge, and explainable language-based reporting to provide transparent AI-assisted corneal diagnostics. Although the current results are based on a pilot cohort, the framework demonstrates the potential of integrating neural networks, knowledge graphs, and large language models for interpretable ophthalmic AI systems. Future studies using larger multicenter datasets are required to further validate clinical performance and generalizability.