Timely and accurate identification of eye disease is essential to avoid permanent vision loss and to begin immediate treatment. Methods: We examined a Random Forest classifier that determines eye-disease status given clinical patient records. Confusion-matrix results detected 1,834 diseased and 1,743 non-diseased subjects in diagnostic classification. The model is highly accurate with extremely low false positive (just 4) meaning that there are rare false alarms. However, the 419 false negatives demonstrate a relatively low sensitivity for early or subtle disease. The ROC curve gives some more insight into this, and the classifier is 0.74 AUC, which is moderate discriminative ability, and also shows that the model is performing consistently above random guessing across thresholds. The above further supported by the classification report shows that this model has an overall accuracy of ~0.89 with pretty high F1-scores of 0.89 for No Eye Disease class and 0.90 for Eye Disease class. In conclusion, the Random Forest model performs very accurately and consistently, but still requires additional optimization with respect to sensitivity and early detection of the disease.

 

Keywords: Eye Disease Prediction, Random Forest, Machine Learning, Clinical Decision Support