Knowledge of the causal relationships between clinical actions and patient outcomes is of paramount importance in improving the quality and cost-effectiveness of health care. Randomized controlled clinical trials (RCTs) provide the most reliable method we have for establishing and quantifying clinical causal relationships. While we can use RCTs to provide in theory an ideal test of causality, in practice they may be infeasible or impractical. The amount of available observational clinical data is much greater than the amount of experimental data, and observational data is being increasingly recorded in hospital information systems. We need reliable methods that make use of observational data to augment our understanding of the relationships between clinical actions and patient outcomes. Most current observational techniques rely on an assumption that all significant confounders of actions and outcomes are measured and controlled for. This assumption is difficult to test, and thus, the validity of the resulting causal conclusions remain in question. We have developed a new computer-based analytic method which assumes that different clinician groups see a representative sample of a particular patient population. Clinicians may be grouped according to a variety of criteria, including past resource use. If the patients treated by such clinician groups differ significantly according to their resource use or to the outcomes of their patients, then their differences are due to practice variations. In that case, we can analyze the specific variations in clinical actions that characterize the groups. By assuming that all significant clinical actions have been measured, we also can constrain in general the causal influence of the actions on clinical outcomes. We propose to investigate the new analytic modeling method using retrospective data on patients with community-acquired pneumonia, myocardial infarction, asthma, chronic obstructive pulmonary disease, congestive heart failure, cellulitis, and transient ischemic attacks. Our study will focus initially on the records of patients who present with these conditions to the Emergency Department. We will use patient data from information Systems at the University of Pittsburgh Medical Center (UPMC). We plan to test the assumption of representative group samples by conducting observational field studies in clinical environments at UPMC. If the proposed method is able to estimate reliably the influence of clinician actions on patient outcomes, it will provide a major new tool to assist in improving the quality and cost-effectiveness of clinical care.
| Lustgarten, Jonathan L; Visweswaran, Shyam; Gopalakrishnan, Vanathi et al. (2011) Application of an efficient Bayesian discretization method to biomedical data. BMC Bioinformatics 12:309 |
| Gopalakrishnan, Vanathi; Lustgarten, Jonathan L; Visweswaran, Shyam et al. (2010) Bayesian rule learning for biomedical data mining. Bioinformatics 26:668-75 |
| Mani, S; Cooper, G F (2000) Causal discovery from medical textual data. Proc AMIA Symp :542-6 |
| Kayaalp, M; Cooper, G F; Clermont, G (2000) Predicting ICU mortality: a comparison of stationary and nonstationary temporal models. Proc AMIA Symp :418-22 |
| Mani, S; Cooper, G F (1999) A study in causal discovery from population-based infant birth and death records. Proc AMIA Symp :315-9 |
