Prognostic Value of a Classification and Regression Tree Model in Patients with Open-Globe Injuries

Abstract

Objective: To evaluate the accuracy of the Classification and Regression Tree (CART) model in prognosticating visual outcomes of patients with open-globe injuries

Methods: This was a retrospective, single-center, cohort study of patients with open-globe injuries seen over a two-year period. Purposive sampling of hospital medical records was done to collect data from both in- and out-patient cases. The CART algorithm was utilized to determine the predicted visual outcome for each case, and the accuracy of prognostication was measured by computing for sensitivity, specificity, positive predictive value, and negative predictive value. The area under the receiver operating characteristic curve was used to check its discriminatory capability.

Results: A total of 65 eyes (65 patients) with the following diagnoses based on the Birmingham Eye Trauma Terminology (BETT) classification were included: penetrating eye injury (n=58), globe rupture (n=2), and intra- ocular foreign body (n=5). Majority were male patients (81.5%) in the 17-39 year age group (40%). The sensitivity and specificity of CART were 100% (95% CI 93.6 to 100%) and 77.8% (95% CI 40 to 97.2%), respectively, with an overall accuracy of 96.9% (95% CI 89.3 to 99.6%). Area under the curve (AUC) was statistically significant at 0.89 (95% CI 0.79 to 0.95), indicating that the CART model can discriminate vision survival versus no vision.

Conclusion: The CART model demonstrated high accuracy in prognosticating visual outcomes after an open- globe injury in the local setting. It may be used as a helpful tool to guide treatment decisions in open-globe injuries.