Assessing Clinicians’ Ability to Dx Infectious Keratitis

Download PDF

How well do corneal specialists distinguish bacterial from fungal keratitis? There is significant room for improvement, according to a comparison of human clinical performance and confirmed microbiologic data.¹

“The average expert cornea clinician was only a little bit better than random chance when interpreting images of corneal ulcers,” said lead author Travis K. Redd, MD, MPH, at the Casey Eye Institute in Portland, Oregon. However, there was considerable regional vari­ation, with cornea experts in South India significantly outperforming their colleagues in other countries.

Grading the images. Sixty-six cornea specialists from 16 countries were asked to grade 100 corneal ulcer images from a clinical trial database at the Aravind Eye Care System in South India. The graders received no clinical or historical information but were told that half the cases had been microbiologically proven to be bac­terial keratitis, the other half fungal. Ten images (five bacterial; five fungal) were presented twice to each grader for assessment of test-retest reliability.

Results showed that the experts’ areas under the curve (AUC) were highly variable. These measures of the usefulness of a test ranged from 0.39 to 0.82, with a mean of 0.61. This means that some graders did worse than a coin toss at predicting the cause of infec­tion, while others were reasonably accurate. The mean AUC figure suggests that the average expert cornea clinician was only a little better than random chance in interpreting the images, said Dr. Redd.

Regional variations. A subgroup analysis revealed that the clinicians who practice in India were more ac­curate in identifying fungal ulcers than their colleagues located elsewhere (76% vs. 49% accuracy rate, respec­tively). No comparable geographic difference was noted with bacterial ulcers.

The subgroup finding suggests a possible difference in the appearance of ulcers in different regions. Thus, familiarity with one’s local ulcer morphology possibly confers better accuracy at predicting the underlying cause of infection, Dr. Redd said. Alternatively, the dif­ference may be attributable to Indian experts’ greater familiarity with fungal keratitis, which is less common in other regions, including the United States and Europe.

While additional studies are needed to compare expert performance, the finding emphasizes the impor­tance of regionality in predicting etiology of corneal ulcers and has implications for designing an artificial model (AI) model, Dr. Redd said. “Careful regional evaluation will be required before these models can be implemented. AI models trained on data from one location may not generalize to other regions.”

Building on the findings. The researchers recently published initial results of an AI model that demon­strated “superhuman performance, even surpassing the Indian experts in differentiating bacterial and fungal keratitis,” said Dr. Redd.² Now they are working to incorporate clinical history and expert opinion into the model. They’re also developing models to inter­pret smartphone images, which may allow for earlier initiation of antimicrobial therapy and improved visual outcomes for patients.

—Miriam Karmel

___________________________

1 Redd TK et al. Ophthalmology. 2022;129(2):227-229.

2 Redd TK et al. Ophthalmol Science. Published online Jan. 28, 2022.

___________________________

Relevant financial disclosures—Dr. Redd: NEI: S; Research to Prevent Blindness: S.

For full disclosures and the disclosure key, see below.

• Biomarkers Predict Anti-VEGF Response in AMD

• OCT Pins Down MOGAD and MS Diagnosis in ON

• AI Quantifies Loss of Retinal Cells in Stargardt