Predicting visual acuity with machine learning in treated ocular trauma patients

Area of Science:

• Ophthalmology
• Medical Artificial Intelligence
• Computer Vision

Background:

• Ocular trauma presents a significant challenge in predicting visual outcomes.
• Accurate prediction of best-corrected visual acuity (BCVA) is crucial for managing patients with ocular trauma.
• Current methods may not fully capture the complexity of visual recovery post-trauma.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting BCVA in patients with ocular trauma.
• To assess the accuracy of ML algorithms in classifying visual function after treatment.
• To identify key clinical and ocular features influencing visual outcomes in trauma patients.

Main Methods:

• Utilized a dataset of 850 patients (1589 eyes) with ocular trauma, with a separate test set of 60 patients (100 eyes).
• Employed four ML algorithms (Extreme Gradient Boosting, support vector regression, Bayesian ridge, random forest regressor) for BCVA prediction and four for classification.
• Integrated clinical data with ocular parameters from optical coherence tomography and fundus photographs.

Main Results:

• ML models demonstrated significant correlation between predicted and actual BCVA values (Pearson correlation coefficient > 0.6).
• Achieved low Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) in both traumatic and combined groups.
• Test dataset showed high accuracy (0.90) and precision (0.92) for BCVA classification, with sensitivity (0.83) and specificity (0.95).

Conclusions:

• Machine learning models provide accurate predictions of BCVA in patients with treated ocular trauma.
• These models are valuable tools for identifying visual dysfunction and guiding patient management.
• The integration of diverse data sources enhances predictive capabilities for visual outcomes.