A Statistical Reasoning System for Medication Prompting

We describe our experience building and using a reasoning system for providing context-based prompts to elders to take their medication. We describe the process of specification, design, implementation and use of our system. We chose a simple Dynamic Bayesian Network as our representation. We analyze the design space for the model in some detail. A key challenge in using the model was the overhead of labeling the data. We analyze the impact of a variety of options to ease labeling, and highlight in particular the utility of simple clustering before labeling. A key choice in the design of such reasoning systems is that between statistical and deterministic rule-based approaches. We evaluate a simple rule-based system on our data and discuss some of its pros and cons when compared to the statistical (Bayesian) approach in a practical setting. We discuss challenges to reasoning arising from failures of data collection procedures and calibration drift. The system was deployed among 6 subjects over a period of 12 weeks, and resulted in adherence improving from 56% on average with no prompting to 63% with state of the art context-unaware prompts to 74% with our context-aware prompts.