Validating Variational Bayes Linear Regression Method With Multi-Central Datasets.

Area of Science:

• Ophthalmology
• Medical Statistics
• Machine Learning

Background:

• Glaucoma is a progressive optic neuropathy characterized by visual field (VF) loss.
• Accurate prediction of VF progression is crucial for timely intervention and management.
• Traditional statistical methods may have limitations in capturing the complexities of disease progression.

Purpose of the Study:

• To validate the predictive accuracy of variational Bayes linear regression (VBLR).
• To compare VBLR's performance against ordinary least squared linear regression (OLSLR).
• To assess VBLR's utility in predicting future visual field changes using external datasets.

Main Methods:

• Two external datasets (JAMDIG and DIGS) were used for validation.
• Both VBLR and OLSLR were applied to predict total deviation (TD) values of future visual fields.
• Prediction accuracy was quantified using the root mean squared error (RMSE).

Main Results:

• VBLR demonstrated significantly lower RMSEs compared to OLSLR across all tested visual field series (P < 0.01).
• VBLR achieved RMSEs between 5.0 and 3.7 dB for JAMDIG and 4.6 and 3.6 dB for DIGS.
• No statistically significant difference was found in VBLR's performance between the JAMDIG and DIGS datasets.

Conclusions:

• VBLR significantly outperforms OLSLR in predicting visual field progression.
• VBLR shows potential as a valuable tool in clinical settings for glaucoma management.
• The model's robustness is supported by consistent performance across independent datasets.