Evaluation of novel Bayesian algorithms for ophthalmic presentations to aid at point of care

Authors: Amy Basilious ,Alexander Deans, Chris Govas, Robin Deans, Cindy Hutnik. 

Author Disclosure Block: A. Basilious: None. A. Deans: None. C. Govas: None. R. Deans: None. C. Hutnik: None.

Abstract Body:

Purpose: Ophthalmic conditions present diagnostic challenges for healthcare providers, resulting in unnecessary referrals and delayed care. A dynamic algorithm that simplifies and communicates a clear medical workup has the potential to improve the value and efficiency of care. This project assesses the utility of novel decision-making algorithms for ophthalmic presentations.

Study Design: Prospective multicenter study

Methods: This study includes patients with 1) Red Eye or 2) Vision Loss presenting to family physicians and emergency departments in Windsor, Ontario and urgent care in London, Ontario. These patients were first evaluated by family physicians or emergency staff who rendered a “referrer diagnosis”.

Questionnaires based on history and physical exam were also completed at point of care by these providers. Patients were then referred for assessment by an ophthalmology staff and/or resident who attributed a “gold standard diagnosis”. Post-visit, data from the questionnaires were entered into the algorithm, which provided an “algorithm differential”. Referrer diagnoses and the algorithm’s top 3 diagnoses were compared to the gold standard to determine diagnostic accuracy.

Results: This study included data from 215 patient assessments. In the Windsor sample, referrer diagnoses were correct in 35% of red eye cases (n=48).The algorithm’s top diagnosis was correct in 71% of these cases, increasing to 94% and 98% when the top 2 and top 3 diagnoses, respectively, were included. Referrer diagnoses were correct in 29% of vision loss cases (n=66). The algorithm’s top diagnosis was correct in 70% of these cases, increasing to 85% with the top 2 diagnoses and 86% with the top 3. In the London sample, referrer diagnoses were correct in 71% of red eye cases (n=48). The algorithm’s top diagnosis was correct in 58% of these cases, increasing to 69% and 75% when the top 2 and top 3 diagnoses, respectively, were included. Referrer diagnoses were correct in 49% of vision loss cases (n=53).The algorithm’s top diagnosis was correct in 45% of these cases, increasing to 66% with the top 2 diagnoses included and 72% with the top 3.

Conclusions: These dynamic, Bayesian algorithms successfully improved diagnostic accuracy, using only clinical tools and information collected by non-ophthalmologists. This tool has the potential to optimize patient outcomes by improving the triage of patients. Refer rer diagnostic accuracy differed between the Windsor and London datasets, perhaps suggesting that the algorithms may provide the greatest benefit in settings such as family physician practices and community hospitals.