Artificial intelligence virtual assistants in primary eye care practice

Abstract

Purpose: To propose a novel artificial intelligence (AI)-based virtual assistant trained on tabular clinical data that can provide decision-making support in primary eye care practice and optometry education programmes.

Method: Anonymised clinical data from 1125 complete optometric examinations (2250 eyes; 63% women, 37% men) were used to train different machine learning algorithm models to predict eye examination classification (refractive, binocular vision dysfunction, ocular disorder or any combination of these three options). After modelling, adjustment, mining and preprocessing (one-hot encoding and SMOTE techniques), 75 input (preliminary data, history, oculomotor test and ocular examinations) and three output (refractive, binocular vision status and eye disease) features were defined. The data were split into training (80%) and test (20%) sets. Five machine learning algorithms were trained, and the best algorithms were subjected to fivefold cross-validation. Model performance was evaluated for accuracy, precision, sensitivity, F1 score and specificity.

Results: The random forest algorithm was the best for classifying eye examination results with a performance >95.2% (based on 35 input features from preliminary data and history), to propose a subclassification of ocular disorders with a performance >98.1% (based on 65 features from preliminary data, history and ocular examinations) and to differentiate binocular vision dysfunctions with a performance >99.7% (based on 30 features from preliminary data and oculomotor tests). These models were integrated into a responsive web application, available in three languages, allowing intuitive access to the AI models via conventional clinical terms.

Conclusions: An AI-based virtual assistant that performed well in predicting patient classification, eye disorders or binocular vision dysfunction has been developed with potential use in primary eye care practice and education programmes.

Keywords: artificial intelligence; clinical decision support; machine learning; optometry; virtual assistant.

FIGURE 1

Names and formats of the features used to classify the eyes. ISNT, inferior, superior, nasal, temporal; NPC, near point of convergence.

FIGURE 2

Diagram of the data processing and model validation steps. M1, model to predict the most likely case classification; M2, Model to predict pathology in different ocular segments; M3, Model to predict any binocular vision dysfunction type; Nb, balanced normalised; Nu, unbalanced normalised; Ob, original balanced; Ou, original unbalanced.

FIGURE 3

Summary of the most impactful features in each model. The top panel presents the Model (M1) used to predict the most likely case classification, the middle Model (M2) used to predict pathology in different ocular segments and the bottom Model (M3) used to predict any binocular vision dysfunction type. Left (a): An ordered overview of feature importance is presented in each model (dashed line marking the point where at least 95% importance is reached), and right (b): A breakdown of the top 15 features most relevant in each model is presented. Conj, conjunctiva; CT, cover test; ISNT, inferior, superior, nasal, temporal; NPC, near point of convergence; OD, optic disc; VA, visual acuity.

FIGURE 4

Summary of the random forest (RF) algorithm performance metrics for subclassification. (a) Results for Model 1 subject subclassification. (b) Precision–recall curve of Model 2 for segment subclassification of eye diseases. (c) Confusion matrix of Model 3 for subclassification of binocular vision dysfunction.

FIGURE 5

Interface of the web‐based application designed to use the developed artificial intelligence (AI) algorithms. Vision Care Helper Intelligence (VICHI) is open access and available at https://www.visioncare.digital/vichi. The figure shows a sequence of screens, starting with the search for available solutions, going through the input data entry and modulation (conversion of nomenclature to model data format) and ending with the prediction result of the consulted model.