Seventy-nine of 80 photographs were included for analysis. We excluded one photograph from the fungal trial because it depicted a mixed bacterial and fungal infection. One of the clinicians from the Aravind group did not complete the survey and was excluded from the analysis regarding differentiation of Gram stain, genus, and species.
Of the 39 smear-proven fungal ulcers, 30 grew in culture and could be further identified. The most common causative organisms were Fusarium spp (n = 18, 45%) and Aspergillus spp (n = 10, 25%). Of the 40 bacterial ulcers, the most common causative organisms were Streptococcus pneumoniae (n = 20, 50%) and Pseudomonas aeruginosa (n = 10, 25%) (). As not all organisms can be speciated based on laboratory protocol, species data were included whenever available.
Clinicians were able to accurately distinguish bacterial from fungal etiology 66% (95% confidence interval, 63% to 68%) of the time, a result significantly better than chance alone (P < 0.001, permutation test) (). The clinicians from the Proctor were able to differentiate the etiologies in 63% (95% CI, 59% to 68%) of photographs, whereas the Aravind group was successful in 68% (95% CI, 64% to 73%) of cases. The two clinical groups (Aravind and Proctor) had a different success rate in distinguishing bacterial from fungal keratitis (P = 0.041, permutation test). Examining photographs of bacterial infections only, the probability of correct identification was 69% (95% CI, 62% to 76%), whereas for fungal infections, the probability of correct identification was 62% (95% CI, 55% to 69%).
Data were available regarding the Gram stain for all 40 bacterial ulcers, the genus for 40 bacterial ulcers and 30 fungal ulcers, and the species for 34 bacterial ulcers and 10 Aspergillus ulcers. Clinicians accurately predicted Gram stain results of the 40 bacterial ulcers 46% (95% CI, 40% to 53%) of the time, and were able to correctly identify the genus and species of these bacterial ulcers in 23% (95% CI, 17% to 30%) and 24% (95% CI, 16% to 31%) of cases, respectively (). When examining the 39 cases of fungal keratitis, the clinicians predicted genus in 27% of cases (95% CI, 21% to 33%) and species in 7.9% of cases (95% CI, 1.7% to 16.0%). When analyzing all available data for bacterial and fungal ulcers together, clinicians accurately predicted genus for 25% (95% CI, 20% to 29%) of 70 ulcers and species for 20% (95% CI, 15% to 25%) of 44 ulcers.
We analyzed the results of the aggregated bacterial and fungal ulcers for both the Proctor and Aravind groups. The Proctor group accurately predicted Gram stain results, genus, and species in 47% (95% CI, 37% to 56%), 22% (95% CI, 17% to 28%), and 15% (95% CI, 8.7% to 22%) of cases, respectively. The clinicians from the Aravind Eye Hospital identified Gram stain results, genus, and species accurately 46% (95% CI, 37% to 56%), 29% (95% CI, 22% to 35%), and 27% (95% CI, 17% to 36%) of the time, respectively.
We found no evidence that the two groups differed with respect to the ability to predict Gram stain results for bacterial ulcers (P > 0.99, permutation test) or genus for fungal ulcers (P = 0.31). The Aravind group was significantly better at predicting bacterial genus (P = 0.013), bacterial species (P = 0.0024), and Aspergillus species (P = 0.041).
The clinical signs most commonly used by clinicians to differentiate ulcers were infiltrate border appearance, surrounding stromal haze, and the presence (or absence) of hypopyon. The same clinical signs were listed as most helpful in the graders' clinical decision regardless of whether the photograph was correctly identified as bacterial or fungal.
The clinical signs noted most often were central location of the infiltrate, the absence of a plaque, well-delineated borders, and moderate surrounding stromal haze. The predictive abilities of clinical signs were analyzed with respect to the known etiology based on laboratory testing, not the clinical determination. The presence of irregular/feathery borders was strongly associated with fungal keratitis (P = 0.002, logistic regression) (). This clinical sign was present in 14 of 39 fungal photographs and 4 of 40 bacterial photographs. Although a wreath infiltrate was rarely seen, its presence was associated with bacterial keratitis (P = 0.005). The presence of an epithelial plaque was never agreed on by most raters as being present in any one photograph, yet it remained statistically associated with an underlying bacterial etiology (P = 0.02).