Kalman Filtering Based Machine Learning to Predict Future Mean Deviation Values for Patients with Glaucoma -- Enhancing Existing Models Using Data from Optical Coherence Tomography. A Study Using Data from ADAGES and DIGS.

Abstract

Purpose : To assess whether it is possible to improve the accuracy of predicting future values of mean deviation (MD) for glaucoma suspects and patients with open-angle glaucoma (OAG) by enhancing our previously developed machine learning based Kalman filtering (ML-KF) approach with data on retinal nerve fiber layer (rNFL) thickness from optical coherence tomography (OCT).

Methods : We identified 109 glaucoma suspects and 438 patients with OAG enrolled in ADAGES/DIGS who received tonometry, perimetry, and rNFL OCT approximately every 6 months for 6 years. We parameterized, calibrated, and validated 2 ML-KF models to predict MD on standard automated perimetry 3 years into the future and compared them to actual MD values obtained at study visits. Our previously created ML-KF (KF_A) considered only past values of MD, PSD, and IOP from the patient while a new ML-KF (KF_A+OCT) used these inputs plus past rNFL thickness measurements. Both were compared against 2 traditional linear regression (LR) forecasting models. We compared the error distribution and RSME of the 4 models to determine which was most accurate.

Results : When tested on the 438 patients with OAG, the proportions of MD errors within 1.0 dB of the actual value at 3 years into the future for KF_A, KF_A+OCT, LR1, and LR2 were 43%, 46%, 31%, and 30%, respectively, and the RMSEs were 3.87, 3.88, 4.46, and 4.34, respectively. When the same comparisons were made on the 109 glaucoma suspects, results were 70%, 76%, 52%, and 52% for KF_A, KF_A+OCT, LR1, and LR2, and the RMSEs were 1.05, 0.94, 1.50, and 1.50, respectively.

Conclusions : Our ML-KF models are capable of predicting values of MD 3 years into the future to within 1.0 dB of the actual value for nearly half of the ADAGES/DIGS study participants and within 2.5 dB of the actual value for 83%. The addition of rNFL data from OCT into our ML-KF models enhanced their predictive ability for glaucoma suspects more than for patients with OAG.

This is a 2020 ARVO Annual Meeting abstract.