Answers are grounded in this paper and cite their sources.
Purpose Ophthalmology relies on diverse diagnostic modalities, including imaging and clinical examination. With increasing incorporation of telemedicine and artificial intelligence (AI) into clinical practice, evaluation of diagnostic accuracy is essential, though weaknesses of diagnostic accuracy metrics have been reported. Shannon information theory provides a framework to quantify the performance of diagnostic modalities by calculating their entropy (or information uncertainty) removal, measuring the extent to which these modalities reduce clinical uncertainty. Methods A literature search identified ophthalmic diagnostic accuracy studies that were adherent to STARD standards and reported sufficient data (true and false positives and negatives, respectively), or derivable metrics. Decision tree models were made from extracted variables, and entropy removal was calculated. These values were compared with conventional metrics (sensitivity, specificity, PPV, NPV, Youden index, and diagnostic odds ratio) using Pearson and Spearman correlations. Results Of 1,514 studies, 35 met inclusion criteria (21 non-AI, 14 AI-related), encompassing 27 pathologies and 55 modalities. Entropy removal correlated strongly with the Youden index and PPV. Select examples from analysis were closely examined and demonstrated that iCare tonometry reduced entropy more than non-contact tonometry (53.31% vs 40.74%). Amsler grid screening showed higher utility in healthy individuals, showing 38.58% entropy removal compared to 4.85% removal in dry AMD. Ophthalmologists outperformed neuroradiologists in detecting open globe injuries on CT (47.57% vs 7.09%). Conclusions Shannon entropy can be applied to evaluate the accuracy of ophthalmic diagnostic modalities and has the potential to enhance the assessment of clinical tools and healthcare value. Further validation is needed to confirm its clinical applicability.
No takes yet. Share an insight, caveat, or question.
Chong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: