Tuesday, October 20, 2009
Grand Ballroom, Salons 1 & 2 (Renaissance Hollywood Hotel)
Michael W. Kattan, PhD1, Changhong Yu, MS1, Brian J. Wells, MD, MS1 and C. Greg Hagerty, PhD2, (1)Cleveland Clinic, Cleveland, OH, (2)University of Medicine and Dentistry of New Jersey - Robert Wood Johnson Medical School, New Brunswick, NJ
Purpose: We have developed a web-based, prediction model deployment application for a broad class of risk models that statisticians and physicians can use free of charge, using a systematic methodology. Medical decision-making relies on the predicted probabilities of risk according to alternative treatments and patient characteristics. Model-based risk assessments can be deployed as clinician and patient-friendly computerized tools; however the implementation of such tools can be expensive and error-prone.
Methods: We took a web services approach to develop tools that can be accessed on a wide variety of platforms. The system presents a “Risk Calculator Constructor” interface that allows a statistician to provide a formula for each outcome in the prediction model. The tool automatically populates a template which itemizes the predictors appearing in the equation. The statistician can modify the acceptable bounds for each of the predictors and attach descriptive labels. The system then constructs the corresponding “Risk Calculator” that can be accessed on the web both by personal computers (PC) and handheld devices (eg. BlackBerry). The system also provides a web link that can be emailed to the decision makers. The calculator consists of a form which prompts the user for the values of the predictors and presents the resulting risks of outcomes.
Results: The system is very user friendly for both statisticians and clinicians. The interface for clinicians is capable of specifying a risk calculator with no web-programming expertise and automatically identifies a variety of errors in formulae. The system handles any standard regression models (ordinary least squares, logistic, and Cox), and the resulting risk calculator interface is clear and easily navigable on the PC and Blackberry. Furthermore, as the tool requires very little additional ongoing support, we are now hosting it free of charge.
Conclusion: Our web based application automatically creates an easy to use risk calculator from model-based predictions and should facilitate the distribution of comparative effectiveness tools. This system makes it easy to compare predicted outcomes for individual patients so that decision makers can explicitly see the benefits and harms involved in alternate healthcare scenarios. Moreover, the methodology employs a systematic representation, allowing a wide-variety of models to be catalogued and reviewed.
Candidate for the Lee B. Lusted Student Prize Competition
