Belief-Network Based Reminder Systems that Learn

Janine E. Janosky, Ph.D. - Statistician

Reminder systems are expert systems that send messages to physicians whenever they detect potential errors or deficiencies in the management of patients. Current reminder systems, implemented as rule-based expert systems, cannot represent explicitly the uncertain data and knowledge that is characteristic of the medical domain. For these reasons, the first broad objective of the proposed research is to investigate the potential of an alternative type of reminder system-one based on decision theory and belief networks.

The second broad objective of the proposed research is to investigate techniques for the collection and use of feedback from recipients of reminders. Recent advances in technology and in electronic medical record systems now permit reminder systems to potentially collect large numbers of feedback data directly from clinicians-at almost no cost-about the appropriateness of the reminders that they send. The confluence of this development with the recent development of belief-network-learning algorithms makes possible the use of such data to improve the performance of a reminder system. These learning algorithms provide a mechanism to use feedback data to alter the knowledge encoded in a reminder system, provided that the reminder system is based on belief networks.

The specific aims of this project are:

• (1) to build a belief-network-based reminder system, and a rule-based system to serve as a control;
• (2) to evaluate the performance of the belief-network-based system in a randomized-controlled comparison with the rule-based system;
• (3) to evaluate the effect of learning from feedback data, by comparing the belief-network system before learning with the same system after learning, using a randomized-controlled study design;
• (4) to investigate alternative time-utility functions, the use of patient-specific utilities, and the use of time-dependent utilities to select automatically among alternative reminder-delivery mechanisms having different degrees of intrusiveness for clinicians.