PubMedDate not supplied
Within the last few years, risk assessment has become an integral part of clinical practice, particularly for thoracic surgery and interventional procedures. Risk assessment statistical models are being used in medical decision making, quality improvement tools, and as aids to patient counseling. This literature review was conducted to evaluate the types of predictive models and outcomes measures that have been examined, and methods used in development, validation, and application of these models. A Medline search performed to identify articles (limited to human studies) published in English from 1980 to 1999 resulted in 89 articles, of which 71 were evaluable. Populations studied for model development included patients undergoing coronary artery bypass graft (CABG), percutaneous transluminal coronary revascularization (PTCR), cardiac catheterization, or stenting procedures and patients with angina or stroke. The models were equally developed from a single center versus multicenter and from retrospective databases versus prospective studies. In terms of model perspectives, only three of the models measured cost or cost-effectiveness as the outcome; the remainder considered only clinical outcomes. The most commonly reported types of predictive models were developed using logistic regression and Bayesian techniques, followed by neural networks, rule-based artificial intelligence, simultaneous equation system, and multiple linear regression. Factors to consider when developing or evaluating a predictive model include uniformity of definitions of outcomes, uniformity of definitions of variables, completeness of data, number and frequency of variables, timeliness and source of data, development population characteristics, development and testing (validation) cohorts, and calibration and discrimination. Application of these models to an individual patient can spur quality improvement efforts that can lead to dramatic, system-wide improvements in outcomes.