Explainable AI for Corneal Topography Interpretation and Keratoconus Detection

A prospective pilot cohort of 20 eyes was used to develop a neuro-symbolic AI framework for interpreting corneal topography. The framework, authored by Mini Han Wang and last updated in April 2026, integrates deep learning, a symbolic knowledge graph, and large language models to generate bilingual clinical reports. It achieved an area under the receiver operating characteristic curve of approximately 0.95 for early keratoconus detection.

Use Cases

• Training diagnostic models for early keratoconus detection based on corneal topography parameters and curvature maps.
• Developing explainable AI systems for clinical decision support based on the described neuro-symbolic framework integrating knowledge graphs and LLMs.
• Benchmarking AI performance for refractive surgery eligibility assessment based on quantitative biometric parameters.

Strengths

• Framework demonstrated strong diagnostic performance with an AUC of approximately 0.95 in the pilot cohort.
• System is designed for high interpretability, with expert evaluation indicating clinical usefulness of generated reports.
• End-to-end pipeline operates in near-real-time, requiring approximately 95 ± 12 seconds per case.

Limitations

• Results are based on a pilot cohort of only 20 eyes, limiting statistical power and generalizability.
• Row count and column-level documentation are absent; data structure must be inferred from the description.
• The dataset is very small at 14.3 KB, indicating limited raw data scope.

Provenance

Source

Mini Han Wang via figshare.

Collection Method

Data appears to be derived from IOLMaster 700 reports in a prospective pilot study.

Time Range

Pilot study timeframe is not specified.

Freshness

Last updated 2026-04-13 05:46:44.

Geography

Geographic coverage of the pilot cohort is not specified.