RECEIVER OPERATING CHARACTERISTIC METHODOLOGY SUPPORT

点击翻译按钮获取中文摘要

Receiver Operating Characteristic (ROC) methodology has become well established as an important tool for addressing decision-making uncertainties in medicine and in other disciplines. Int evaluates how well a decision strategy classifies retrospective dichotomous (bivalent), or fuzzy (multivalued) events, and it provides a rational basis for designing decision strategies that classify prospective events. Prevalence and error cost factors are easily incorporated into ROC-based decision designs. The purpose of this project is to conduct continuing research and development in ROC methodology as applicable to biomedical research, to publicize practical extensions of ROC methodology derived from this research and development, and to provide computational service, support, and guidance in ROC methodology to the NIH intramural research community. Toward these ends we have conducted research and performed experimentation with ROC methodology as it applies to modern biomedical research objectives. A user-friendly, DOS-based software package called ROCLAB has been produced that computes ROC functions and their useful derived features for discrete and fuzzy class membership data. Decision strategies that account for uncertainties related to prevalence, false classification costs, and fuzzy class membership are easily constructed with ROCLAB. Some results of these studies have been published and presented publicly. ROCLAB has been installed in the DCRT Scientific Computing Resource Center (SCRC) and the program package is available for distribution to interested persons. ROC methodology remains an important tool in biomedical research. Enhancing ROC methodology to support biomedical research and distributing ROC computational tools on modern computer platforms are vital services to offer the biomedical research community. Furthermore, ROC methodology has close associations with neural network methodology, which is fast becoming an important computational tool in the biomedical sciences. Future plans include enhancing ROCLAB as useful new features are developed in response to user feedback; applying the enhanced ROCLAB to new biomedical research problems as appropriate; exploring combining covariates to produce enhanced decision rules; and exploring the use of ROC methods in conjunction with neural network methodology.