Kleshinski James, Khuder Sadik A, Shapiro Joseph I, Gold Jeffrey P
Department of Medicine, The University of Toledo College of Medicine, Health Science Campus, Toledo, OH 43614-2598, USA.
Adv Health Sci Educ Theory Pract. 2009 Mar;14(1):69-78. doi: 10.1007/s10459-007-9087-x. Epub 2007 Nov 7.
To examine the predictive ability of preadmission variables on United States Medical Licensing Examinations (USMLE) step 1 and step 2 performance, incorporating the use of a neural network model.
Preadmission data were collected on matriculants from 1998 to 2004. Linear regression analysis was first used to identify predictors of performance on step 1 and step 2. A generalized regression neural network (GRNN) as well as a feed forward neural network (FFNN) was then developed in an effort to more accurately predict step 1 and step 2 scores from these preadmission data.
Statistically significant predictors for step 1 and step 2 included science grade point average (SGPA), the biologic science (BS) section of the Medical College Admissions Test (MCAT), college selectivity, race, and age of the applicant. Neural networks were found to predict a significant portion of the variance, and the FFNN demonstrated some superiority over that obtained with linear regression models as well as the GRNN.
The results have implications that could impact the selection of applicants to medical school and the neural networks that we developed could be used in a prospective manner.
运用神经网络模型，研究入学前变量对美国医师执照考试（USMLE）第一步和第二步成绩的预测能力。
收集了1998年至2004年入学学生的入学前数据。首先使用线性回归分析来确定第一步和第二步成绩的预测因素。然后开发了广义回归神经网络（GRNN）以及前馈神经网络（FFNN），以便根据这些入学前数据更准确地预测第一步和第二步的分数。
第一步和第二步成绩在统计学上具有显著意义的预测因素包括科学平均绩点（SGPA）、医学院入学考试（MCAT）的生物科学（BS）部分、大学选择性、种族和申请人年龄。发现神经网络可以预测很大一部分方差，并且前馈神经网络显示出比线性回归模型和广义回归神经网络所获得的结果具有一定优势。
研究结果具有可能影响医学院校申请人选拔的意义，并且我们开发的神经网络可以前瞻性地使用。
