COMPARISON OF CONTROL ALGORITHMS FOR SCHEDULING TESTING VISITS

Purpose: We developed and compared the performance of two novel dynamic control algorithms for determining personalized time between testing for patients diagnosed with open angle glaucoma (OAG) against fixed interval monitoring schedules.

Methods:  We developed a Kalman filter which combines population disease dynamics with the individual patient's health to predict the patient's future health state.  Logistic regression was then used to map the Kalman filter's Gaussian confidence region for the forecasted health states to a probability that a patient is experiencing OAG progression.  Two control algorithms (for scheduling the times at which tests will be taken) were created: (1) test when the worst-case point of the confidence region exceeds a threshold on the probability of progression; or (2) test when a proportion of the confidence region exceeds a progression threshold.  These algorithms were compared against fixed interval scheduling using longitudinal data from 571 patients who were enrolled in the Advanced Glaucoma Intervention Study (AGIS) and Collaborative Initial Glaucoma Treatment Study (CIGTS) randomized clinical trials.

Results: Our control algorithms achieved Pareto dominance over the fixed yearly monitoring schedule with respect to efficiency in progression detection and detection delay. Using the same average number of scheduled tests as the fixed yearly schedule for each patient, both algorithms increased efficiency in identifying OAG progression by 29% (p<0.0001) and detected progression 57% earlier (p=0.02).  Furthermore, the algorithm's performance using the conservative single point of progression was near identical to the algorithm's performance using the more robust proportion of the region method.

Conclusions: We proposed and validated two new and robust algorithms for designing personalized monitoring schedules based on the patient's disease dynamics (which are updated as each new test is performed). Our two control algorithms reallocate monitoring visits to schedule testing closer to the time when progression occurs and to avoid testing when no significant change in the patient's health state is predicted.  We demonstrate that equivalent solutions may be obtained by considering two optimization based approaches: a single point of progression threshold or a region crossing the progression threshold.  More broadly, the novel modeling framework we developed can be applied to determine personalized monitoring schedules for a variety of chronic diseases which require longitudinal monitoring.