Exploring L1 and L2 Regularization Techniques For Combining Learning and Feature Selection on Cancer Imaging Data

Presentation Date: 

Monday, December 11, 2017

Location: 

Massachusetts Institute of Technology

 We developed a statistical learning approach to predict head and neck cancer response to radiation therapy using PET medical imaging data. PET imaging looks at glucose metabolism, which allows cancer cells to proliferate, with the help of a tracer. Glucose metabolism is known to be highly indicative of treatment response and patient survival. We have utilized publically available clinical trial data from The Cancer Imaging Archive (http://doi.org/10.7937/K9/TCIA.2017.umz8dv6s) as our training data. Inherently, imaging data of this type has hundreds of potential features; however, clinical studies like this typically have around 100 patients. Furthermore, imaging data features are highly correlated. Due to this “trifecta” of low sample size, high dimensionality, and high correlation there is significant room to explore regularization learning methods to understand this problem. We analyzed efficacy of these combinations through embedded variable selection schemes by looking at their ability to predict patient survival outcomes from diagnostic PET imaging. The data set contained 161 patients with complete clinical and imaging data, including 35 patients with non-survival due to cancer progression; we split the data set into a training set (n=100 with 20 non-survivors) and a validation set (n=61). We identified 28 clinical features from patient health records and 42 PET imaging features from the diagnostic imaging. PET images are split into 100 2D slices, and each imaging feature was calculated per slice. We calculated the mean, standard deviation, min, and max feature value over all of a patient’s slice, resulting in 168 total imaging features. We fit lasso, elastic net, l1-svm, and 1l-random forest models to the data in order to predict patient survival. We utilized 10-fold cross validation on the training set to tune hyper parameters. We used the resulting tuned models on the validation set to assess model misclassification, sensitivity, and specificity. The four regularization methods had the following cross validation errors: elastic net (13%), l1 penalized random forest (13%), lasso (17%), and l1-SVM (22%). Lasso, elastic net, and l1-SVM had similar sensitivity scores of 0.913, 0.913, and 0.956, respectively. L1-random forest, however, has sensitivity 0.76 suggesting it is misclassifying patients with positive survival at a higher rate than the other methods. All four methods have fairly low specificity – 0.4 for l1-random forest, 0.267 for elastic net and lasso, and 0.067 for l1-svm, suggesting that many patients with death due to HNSCC are misclassified as surviving. One possible reason for this is that the feature weights were learned predominantly  using surviving patients because there were only 20 patients with death due to HNSCC in the training set. Using a resampling approach to include more patients with death due to HNSCC compared to surviving patients may help to solve this issue. Finally, we found that the overlap between features selected by the different methods is small, suggesting that many different combinations of the features lead to similar classification ability. Our work suggests that PET imaging does have some predictive ability with regards to patient survival; however additional features, patients, and imaging modalities are likely needed to improve prediction power.