Candidate for the Lee B. Lusted Student Prize Competition
Purpose: We investigated how using filtered longitudinal data as input for logistic regression to predict glaucoma progression affects the classification ability of the logistic regression function.
Methods: A Kalman filter was developed to reduce the process and measurement noise present in longitudinal data from the Collaborative Initial Glaucoma Treatment Study (CIGTS), a randomized clinical trial of patients with early to moderate open angle glaucoma (OAG). These filtered repeated measures estimates were then used as data input for logistic regression via generalized estimating equations in order to predict OAG progression in patients. Analysis of the receiver operating characteristic (ROC) curve was used to compare this Kalman filter-based model against the standard methodology of using raw observations from the clinical trial as data input for logistic regression.
Results: The Kalman filter-based model resulted in higher specificity and sensitivity compared to the standard raw observations logistic regression model. The area under the ROC curve (AUC) for the Kalman filter-based model was 0.953 while the AUC for the raw observations model was 0.890.
Conclusions: Kalman filtering for estimating the true value of disease-related variables has been shown to improve the progression identification ability of logistic regression functions as compared to the standard approach of using raw data. This approach is applicable to any chronic disease which is subject to noisy observations and requires longitudinal follow-up for effective disease management.
