MCAT is a program of the Association of American Medical Colleges

MCAT AMCAS ERAS NRMP Publications More Resources	Shopping Cart Site Map

About the MCAT Exam

Registration

Examinees with Disabilities

Lacher, D.A.; Richardson, B.L. Predicting First-Year Medical School Performance From MCAT Scores and Premedical Grade-Point Average Using Neural Networks. Paper presented at the Research in Medical Education Conference; October 1994. Boston, Mass.

PURPOSE: This study was designed to compare neural networks and multiple linear regression approaches in their ability to predict medical school performance and the selection of significant predictor variables.

METHODS: Performance measures consisted of grades and a cumulative first-year scores for 107 medical students. The independent variables included in the study were undergraduate GPA and MCAT scores. Multiple linear regression and back propagation neural network analyses were performed. The authors developed a back propagation neural network consisting of 7 input neurons, 5 hidden neurons, and one output neuron for the first-year medical school grades. Cross-validation was accomplished by selecting 85 (80%) students on which to develop the regression equation and neural network and subsequently predicting the grades of the remainng 22 (20%) students.

RESULTS: When all input variables were used, the neural network predicted grades (R=0.75) better than multiple linear regression (R=0.64). Stepwise multiple linear regression selected the science GPA and MCAT Biological Sciences score as the best predictor variables while neural network sensitivity analysis revealed that the total GPA and the Writing Sample scores were the best predictor variables. Cross-validation analysis revealed that multiple linear regression gave more consistent predictions of first-year performance (R ranged from 0.62 to 0.73) than neural networks (R ranged from 0.38 to 0.68).

CONCLUSIONS: The authors concluded that their findings indicated neural networks better predicted medical school grades than did multiple linear regression but were less consistent in cross-validation analysis. Additionally, they determined that each type of analysis identified different variables as being most valuable in predicting first year medical school grades.