Development of an Algorithm for Real‐time Delivery of Automated Intervention Boosters to Support Long‐term Use of an Electronic Memory Aid
Maureen Schmitter‐Edgecombe, Catherine Luna, Sarah Tomaszewski Farias, Diane J Cook- Psychiatry and Mental health
- Cellular and Molecular Neuroscience
- Geriatrics and Gerontology
- Neurology (clinical)
- Developmental Neuroscience
- Health Policy
- Epidemiology
Abstract
Background
Electronic memory aids can facilitate positive health outcomes in older individuals experiencing memory difficulties. Despite this, long‐term maintenance of device use is often limited. Delivery of automated boosters has the potential to enhance use of these aids and retention of compensation training goals, but questions remain regarding optimal timing of boosters. For example, if delivered too frequently recipients will habituate. Here we describe our process for developing an automated schedule that will detect when use of an electronic memory aid has started to decline so that an automated booster can be delivered. We start by unobtrusively observing use patterns and generating initial delivery rules based on clinical judgement, and then apply these rules to develop an automated algorithm.
Methods
We used three months of post‐training data from 28 older adults with mild cognitive impairment who learned to use an electronic memory and management aid (EMMA) application. Two clinicians examined graphs of each users’ daily interactions with EMMA to identify specific days when device interactions had declined and iteratively developed a set of rules that identified the best time points to deliver boosters. The algorithm was then applied to data from an entirely different cohort of 50 older adults engaged in an intervention utilizing the EMMA app. We then compared when the clinician would have boosted vs when the algorithm would deliver a boost.
Results
Following 4 iterations of clinicians examining graphs with possible boost points from the first set of users, an algorithm was developed. When applied to the second data set, the algorithm identified 371 boost points, with a false positive rate of 11% and false negative rate of 3%. It was also deemed that the timing of approximately 28% of the correctly identified boost points by the algorithm were off by +/‐ 2 days.
Conclusion
Next steps will be to further modify the algorithm and then formally test the impact of boosts at different time points to empirically evaluate their impact on usage. Our ultimate goal is to deliver boosters in real time and as early as possible when usage wanes to improve intervention outcomes.