The goal of this project is to refine and evaluate techniques that automatically construct, from clinical databases, Bayesian belief networks that can be used as diagnostic and prognostic aids. The amount of clinical information stored in databases has increased markedly in the last two decades, and it seems likely that this trend will continue. Belief networks are able to represent the probabilistic dependencies among clinical variables in a relatively general manner. Researchers have developed algorithms for performing probabilistic inference using belief networks, and they have applied these algorithms to perform medical diagnosis and prognosis. Although advances have been made in developing the theory and application of belief networks, the manual construction of these networks often remains a difficult, time-consuming task. The automated generation of belief networks from high-quality databases may facilitate significantly the construction of diagnostic and prognostic systems, which can serve as clinical decision aids, after their accuracy and usefulness are validated. The long-range goal of this research is to advance our understanding and development of probabilistic systems that can serve as useful diagnostic and prognostic tools for physicians. Such systems can serve as one method for disseminating the clinical knowledge captured in high-quality databases, such as those developed from PORT studies. Within this context, the specific aims of the current, proposed research project are to: * refine and extend current methods for automatically constructing belief networks from large databases; * test the diagnostic and prognostic accuracy of systems that are based on belief networks constructed automatically from high quality databases, compared to several standard statistical techniques; * test whether a combination of automated and expert-based methods for constructing belief networks will yield diagnostic and prognostic systems that are more accurate than systems that are based on belief networks that are constructed automatically. These three aims will be pursued using large, high-quality clinical- research databases at the University of Pittsburgh that contain information on patients with syncope and patients in a PORT study with community-acquired-pneumonia.
| Aliferis, C F; Cooper, G F (1998) Temporal representation design principles: an assessment in the domain of liver transplantation. Proc AMIA Symp :170-4 |
| Cooper, G F; Aliferis, C F; Ambrosino, R et al. (1997) An evaluation of machine-learning methods for predicting pneumonia mortality. Artif Intell Med 9:107-38 |
| Aliferis, C F; Cooper, G F (1995) A new formalism for temporal modeling in medical decision-support systems. Proc Annu Symp Comput Appl Med Care :213-7 |
| Wagner, M M; Cooper, G F (1995) Evaluation of a belief-network-based reminder system that learns from utility feedback. Proc Annu Symp Comput Appl Med Care :666-72 |
