Room: AAPM ePoster Library
Purpose: evaluate the effectiveness of various machine learning algorithms in their ability to predict electron output factors using various clinical parameters. Machine learning models have begun working their way into the clinic, so eliminating the underperforming models from consideration can save time and money.
Methods: data set of 1166 unique, measured output factors and their corresponding clinical parameters were used to train, test, and validate machine learning algorithms including linear ridge regression, support vector machines, decision trees, boosted decision trees, random forests, Gaussian processes, and multi-layer perceptrons. Models were generated by adjusting hyper-parameters and then evaluating performance on training and test sets using R-squared as the metric. Once models performed sufficiently well on both training and testing sets, they were validated by comparing their predictions to measurements on a previously untouched data set.
Results: results of each model were evaluated using maximum percent error, mean percent error, and R-squared value. As models increased in complexity, so too did their predictive power. Linear ridge regression performed the worst, with a max percent error of 17.02%, mean percent error of 1.97%, and an R-squared value of .882. Multi-layer perceptrons performed the best, with a max percent error of 6.54%, a mean percent error of 1.35%, and an R-squared value of .952.
Conclusion: machine learning algorithms used in this work had high variance in their performance metrics. Gaussian process regression, multi-layer perceptrons, and boosted decision trees outperformed the others by a significant margin, and should therefore be considered as the first choice of machine learning models when attempting to solve this clinical problem.
Not Applicable / None Entered.