Ensemble Pruning for Glaucoma Detection in an Unbalanced Data Set.

Area of Science:

• Ophthalmology
• Computer Science
• Machine Learning

Background:

• Random forests are powerful ensemble classifiers, but can be computationally intensive.
• Ensemble pruning reduces computational cost and memory demand without sacrificing performance.
• Early glaucoma detection using topographical data presents unique challenges due to low prevalence.

Purpose of the Study:

• To evaluate ensemble pruning strategies for glaucoma detection in imbalanced datasets.
• To assess the impact of pruning on classification performance metrics like AUC and Brier score.
• To investigate the influence of glaucoma prevalence on pruning strategy effectiveness.

Main Methods:

• Utilized topographical eye data from 254 healthy controls and 55 glaucoma patients.
• Compared pruned random forest ensembles (30-80 trees) against a full 1000-tree ensemble.
• Employed pruning strategies based on prediction accuracy, uncertainty weighting, and tree similarity.
• Conducted a simulation study to analyze the effect of varying glaucoma prevalence.

Main Results:

• Pruning strategies significantly improved Area Under the ROC Curve (AUC) and reduced Brier scores for glaucoma classification.
• Optimized sub-ensembles of 30-80 trees outperformed the full 1000-tree random forest.
• Glaucoma prevalence was identified as a critical factor, with lower prevalence negatively impacting pruning strategy performance.

Conclusions:

• Ensemble pruning effectively reduces memory requirements for random forest-based glaucoma detection.
• Pruning strategies maintain or improve classification performance in unbalanced datasets.
• This approach is beneficial for early glaucoma detection in at-risk populations.