Kalman Filtering Can Forecast NTG Disease Trajectory
American Journal of Ophthalmology, March 2019
Unlike most types of open-angle glaucoma, normal-tension glaucoma (NTG) often includes dense visual field (VF) loss, which may occur close to central fixation and early in the disease. Because common activities such as reading and driving can be difficult for patients with central or paracentral VF loss, it would help to personalize the forecasting of disease trajectory, allowing identification of patients at high risk for progression before the damage occurs. The Kalman filtering (KF) algorithm, which has been used for decades by the aerospace industry to guide planes and shuttles, recently has been applied to the trajectory of chronic diseases. The KF model accounts for underlying disease dynamics among patient populations, as well as unique patient-specific characteristics. Personalized forecasts are derived, which can be updated whenever the patient has additional testing. Garcia et al. previously tested the model in patients with high-tension open-angle glaucoma and found it effective in forecasting disease progression. In a new study, they established its utility for patients with NTG.
Initially, the authors validated a KF model, named KF-NTG, to forecast mean deviation (MD) and other parameters. The algorithm was used for 263 eyes (263 Japanese patients) with NTG. The proportion of patients with MD forecasts within 0.5, 1.0, and 2.5 dB of the actual values was determined, and the root mean squared error (RMSE) was calculated for each forecast. Results of KF-NTG were compared with those of the KF model used for patients with high-tension OAG. Of this group, 242 eyes had enough data to forecast two years into the future.
Twenty-four months in advance, KF-NTG was able to forecast MD values that fell within 0.5, 1.0, or 2.5 dB of actual values for 78 eyes (32.2%), 122 eyes (50.4%), and 211 eyes (87.2%), respectively. The percentage of eyes with forecasted MD values within 2.5 dB of actual values (87.2%) was similar to that with the model for high-tension OAG (86.0%) and with the null model (86.4%), and much better than data from two linear regression models (72.7%-74.0%). KF-NTG achieved a lower RMSE than the other models in this study, denoting the superiority of its performance.
These findings suggest that KF holds promise for personalizing disease trajectory forecasts. The authors continue to refine their KF models by incorporating additional variables and validation studies.
The original article can be found here.