Assessing GPT-4o in Cataract Surgery Decision-Making: 74 Patient Cases

74 cataract patient cases (146 eyes) were used to evaluate the GPT-4o model's appropriateness for intraocular lens (IOL) selection. The dataset includes preoperative clinical findings, biometric parameters, systemic histories, and lifestyle questionnaires. It was created by Yan Liu and last updated on 2026-05-29.

Use Cases

• Benchmarking LLM performance against human surgeons based on appropriateness rates for IOL selection.
• Analyzing how model performance varies by IOL category, such as monofocal versus trifocal toric lenses.
• Investigating the impact of ocular or systemic comorbidities on AI decision-making consistency.
• Training or fine-tuning domain-specific models for routine cataract surgery scenarios.

Strengths

• Dataset is based on 74 real-world patient cases (146 eyes) with adjudicated expert consensus.
• Analysis includes specific performance metrics: 37.84% appropriateness rate for GPT-4o and subgroup rates like 90.0% for monofocal IOLs.
• Evaluation compares model outputs against selections by both high- and low-volume human surgeons.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset is very small at 16.5 KB, indicating limited scope.

Provenance

Source

figshare

Collection Method

Retrospective clinical observational study of patients who underwent phacoemulsification and IOL implantation.

Freshness

Last updated 2026-05-29 06:20:26; freshness should be verified.