Neural network and linear regression models in residency selection

Abstract

For many years, multiple linear regression models have been used at a residency program to generate preliminary rank lists of residency applicants. These lists are then used by the admissions committee as an aid in developing a final ranking to submit to the National Residency Match Program (NRMP). A study was undertaken to compare predictions made using linear regression with those generated by a newer technique, an artificial neural network. A prospective cohort design was used. Seventy-four applicants to an emergency medicine program were evaluated by faculty and resident interviewers with regard to medical school grades, autobiography, interviews, letters of recommendation, and National Board scores. Normalization of these scores (by linear transformation of interviewer means) was used to correct for differences among interviewers. Multivariate linear regression and neural network models were developed using data from the previous 5 years' applicants. These models were used to forecast provisional rank orderings of the candidates. These rankings were combined into a single hybrid list that was used by the admissions committee as the starting point for development of the final rank list by consensus. Each model's predictions were tested for goodness of fit against the final NRMP rank using Wilks' test. Using the final submitted NRMP rank order as the dependent variable, the neural network yielded a correlation coefficient of 0.77 and an R 2 of 59.4%. The linear regression model exhibited a correlation coefficient of 0.74 and an R 2 of 54.0%. No significant difference was found ( χ 2 = 1.08, P = .7). A neural network performs as well as a linear regression model when used for forecasting the rank order of residency applicants.