Establishment of a prediction tool for ocular trauma patients with machine learning algorithm

Area of Science:

• Ophthalmology
• Medical Informatics
• Machine Learning

Background:

• Open globe injury poses a significant threat to vision, necessitating accurate prognostic tools.
• Predicting final visual acuity is crucial for patient management and treatment planning.

Purpose of the Study:

• To develop and validate a highly accurate predictive model for final visual acuity in open globe injury patients.
• To identify key factors influencing the prognosis of open globe injuries.

Main Methods:

• A retrospective review of 171 open globe injury patients' data was conducted.
• Supervised classification algorithms, including a two-class boosted decision tree, were employed using Microsoft Azure Machine Learning Studio.
• Cross-validation, permutation feature importance, and synthetic minority over-sampling technique were utilized to enhance model performance.

Main Results:

• The two-class boosted decision tree model achieved high predictive performance (accuracy: 0.925, AUC: 0.971).
• The top 14 prognostic factors identified include retinal detachment, laceration location, initial visual acuity, and iris damage.
• The model demonstrated strong accuracy, precision, recall, and F1 score.

Conclusions:

• A highly accurate and accessible model for predicting final visual acuity in open globe injury was developed.
• This prognostic tool can aid clinicians and patients, potentially reducing socioeconomic burden.
• Further multicenter validation is recommended for global applicability.