Augmenting Kalman Filter Machine Learning Models with Data from OCT to Predict Future Visual Field Loss: An Analysis...

Area of Science:

• Ophthalmology
• Medical Informatics
• Artificial Intelligence

Background:

• Glaucoma diagnosis and progression monitoring rely on visual field perimetry and optical coherence tomography (OCT) measurements.
• Machine learning (ML) models, including Kalman filters (KF), are being explored to predict glaucoma progression.
• The influence of demographic factors, such as race, on ML model performance requires further investigation.

Purpose of the Study:

• To evaluate if incorporating global retinal nerve fiber layer (RNFL) data enhances the predictive accuracy of Kalman filter (KF) models for glaucoma progression.
• To determine the impact of racial composition in training and testing datasets on the performance of KF models for glaucoma prediction.

Main Methods:

• A retrospective longitudinal cohort study was conducted using data from patients with open-angle glaucoma (OAG) or glaucoma suspects.
• Two KF models were developed: KF-TP (tonometry and perimetry) and KF-TPO (tonometry, perimetry, and global RNFL data).
• These KF models were compared against linear regression (LR) models for predicting 36-month changes in mean deviation (MD) and pattern standard deviation, also assessing performance based on racial alignment of training and testing sets.

Main Results:

• KF models demonstrated higher predictive accuracy (73.5% for KF-TP, 71.2% for KF-TPO) compared to LR models (57.5%, 58.0%) for patients with OAG.
• The addition of global RNFL data (KF-TPO vs. KF-TP) did not significantly improve predictive accuracy (P=0.20).
• Aligning the race of training and testing sets improved mean absolute prediction error by 0.39 dB for KF-TP and 0.48 dB for KF-TPO.

Conclusions:

• Kalman filter models show promise for predicting glaucoma progression, outperforming linear regression models.
• Incorporating global RNFL data offers marginal benefits to KF model predictive accuracy in this cohort.
• Optimizing ML model performance for glaucoma prediction necessitates careful consideration of racial diversity in training and testing datasets.