Room: ePoster Forums
Purpose: Historically, the decision to launch education and training programs at an institution is solely based on perceived need and heuristic methods. In this work we propose using data as evidence to design and develop an informed and successful medical physics graduate program.
Methods: Graduate program data from 2013-2017 for 53 medical physics graduate programs accredited by the CAMPEP were used in this study. Data from each program was processed to establish the number of applicants, enrollments and graduations per year. Using the program URL, key information such as the number of credit hours and core credits was collected. Additional binary variables, such as whether the program requires thesis research for Masters and if a Ph.D. is also offered by the program, were included. A categorical variable, program â€œsuccessâ€? was established as the ratio of graduations to employments for each of the last five years for all programs. Using the open-source statistical computing environment of R, two different predictive statistical learning models were developed for machine learning: a) Classification-based decision trees and b) Ensemble approach of random forests. Both approaches used a subset of data to learn â€œsuccessâ€? to form individual models. The two models were then implemented on a validation dataset to evaluate accuracy.
Results: The decision tree model using required credit hours, core courses, applicants, ability to offer Ph.D., required thesis research, enrollments, and graduations as independent predictors produced receiver operator characteristics with area of 99% and 88% on the learning and validation dataset respectively. For the random forest model the accuracy was at 100% for the learning data and 76% for the validation set.
Conclusion: We have successfully demonstrated the proof of principle for using data analytics tools to produce intelligent models that can predict the success of a medical physics graduate education program with reasonable accuracy.