Using belief networks to enhance sharing of medical knowledge between sites with variations in data accuracy

Differences in data definition between sites are a known obstacle to sharing of reminder-system rule sets. We identify another data characteristic--data accuracy--with implications for sharing. We reviewed the literature on data accuracy and found reports of high error rates for many data classes used by reminder systems (e.g., problem lists). The accuracy of other, equally important, data classes had not been characterized. Wide variations in accuracy between sites has been observed, suggesting that such differences may pose a previously unrecognized barrier to sharing of reminder rules. We propose a belief-network model for encoding reminder rules that explicitly models site-specific data accuracy and we discuss how encoding knowledge in this format may lower the cost and effort required to share reminder rules between sites.