NEW ROC METHODOLOGY FOR EVALUATION OF DIAGNOSTIC SYSTEMS

诊断成像系统和程序的评价是一个重要的 医疗保健的成本和质量因素。多尔夫曼-贝尔鲍姆-梅斯公司 （1992）开发了一种新的方法，提供了一个全面的 分析接收机工作特性的框架 (RQC)多个读者的研究，考虑到这两个 读者-样本和案例-样本变异。广泛的长期目标 是评估、完善和扩大 Dorfman/Berbaum/梅斯（DBM）多读者ROC方法， 诊断成像系统的准确性评价。显著 在这种方法能够广泛应用之前，仍有一些问题有待解决。 有信心地采用，以评估诊断系统。具体 本研究的目的是：（1）确定为什么经验 在零假设下，标称错误率有时会有所不同， 离散评级格式，并调查减少 这种差异在广泛的实验条件下， 有效性和增加权力;（2）确定为什么经验和 对于连续评级格式，标称错误率有时不同， 研究在大范围内减少这种差异的程序， 实验条件的优化，提高效度，增加把握度;（3） 评估DBM多读者方法在估计 各种效应量差异的置信区间;（4）开发 并评估估计最低读者人数和病例的程序 检测具有临床意义的差异所需的样本量， 适当程度的功效和较小的I类错误概率;以及 (5)将协变量纳入DBM多读者方法，以及 评估此扩展的有效性。这些目标将得到解决 通过蒙特卡罗模拟和理论分析。

Evaluation of diagnostic imaging systems and procedures is an important factor in the cost and quality of health care. Dorfman, Berbaum & Metz (1992) developed a new methodology that provides a comprehensive framework within which to analyze receiver operating characteristic (RQC) studies with multiple readers that takes into account both reader-sample and case-sample variation. The broad, long-term objective is to assess, perfect, and extend the useful range of application of the Dorfman/Berbaum/Metz (DBM) multireader ROC methodology for the evaluation of accuracy of diagnostic imaging systems. Significant problems remain to be solved before this approach can be broadly adopted with confidence to evaluate diagnostic systems. The specific aims of the proposed research project are: (1) determine why empirical and nominal error rates sometimes differ under the null hypothesis for the discrete rating format, and investigate procedures for reducing this difference in a broad range of experimental conditions to improve the validity and increase power; (2) determine why empirical and nominal error rates sometimes differ for continuous rating format and to investigate procedures for reducing this difference in a broad range of experimental conditions to improve validity and increase power; (3) evaluate the validity of the DBM multireader methodology in estimating confidence intervals for various effect-size differences; (4) develop and evaluate procedures for estimating the minimum reader and case sample sizes needed to detect a clinically meaningful difference with a fair degree of power and a small probability of a Type l error; and (5) incorporate covariates into the DBM multireader methodology, and evaluate the validity of this extension. These aims will be addressed by both Monte Carlo simulations and theoretical analysis.