APPLYING FRONTAL CORTEX METABOLITES QUANTIFIED BY 7-TESLA 1H-MRS TO PREDICT MULTIPLE SCLEROSIS SUBTYPE THROUGH RECURSIVE PARTITIONING AND CONDITIONAL INFERENCE TREES
MR SCIENCE Lab | Columbia University Medical Center
June 2017 - August 2018
Multiple sclerosis (MS) is an autoimmune disease that impairs the central nervous system by attacking the myelin sheaths of neurons. It affects more than 2.3 million people worldwide. One potential key to understanding the metabolic differences between the brains of patients with relapsing-remitting MS (RR-MS) and progressive MS (P-MS); and potentially treating MS is investigating the relationship between different metabolites and the types of MS. Data on the different metabolite levels of patients with relaxing-remitting MS, progressive MS, and controls have been previously obtained from the frontal cortex of adults using 1H-MRS at 7 Tesla. We created classification trees and performed tree-boosting on these data in order to find the metabolites that best categorize the different MS subtypes as accurately as possible. The results of our conditional inference trees exhibited 44% accuracy in predicting whether an individual had relapsing-remitting, progressive, or no MS; they also showed that glycine and glutamine were the strongest factors in this classification. The results of our recursive partitioning trees showed better accuracy, at 72%, and reinforcing the metabolites with the most influence to the classification. Our data highlight the importance of continued investigation in the potential applications of machine learning in analyzing clinical data.
Multiple Sclerosis (MS) is an autoimmune disease that impairs the nervous system by attacking the myelin sheaths of nerve cells1. Although it affects more than 2.3 million people worldwide, the only way it is diagnosed is either through the McDonald Criteria or through a differential diagnosis. The McDonald Criteria falls short in that it cannot be applied to patients who do not present symptoms during an MR scan. A potential supplement to this is using levels of different metabolites potentially implicated in MS pathogenesis as a predictor for the subtype of MS that a patient has. In this study, we used machine learning to analyze the metabolite levels measured in both RR-MS and Progressive P-MS patients. Machine learning is unique in that it finds patterns in the data not easily found by humans. Furthermore, it is unbiased and is able to learn from previous iterations of the algorithms it uses.
The accuracy of the recursive partitioning tree was 50.32% after cross-validation on the training set and 38.89% after being applied to the test set. The accuracy of the conditional inference tree was 51.02% after cross-validation on the training set and 44.44% after being applied to the test set. The generalized boosted regression model reported a 45.95% accuracy after cross-validation on the training set and a 27.78% accuracy after being applied to the test set.
Furthermore, after using both functions to classify the cases into only healthy patients and those with MS—as opposed to differentiating between RR-MS and P-MS—the recursive partitioning tree had a 64% accuracy after cross-validation on the training set, and a 72% accuracy after applying it to the test set. The conditional inference tree had a 64% accuracy after cross-validation on the training set and a 50% accuracy after applying it to the test set.
Gradient boosting machine algorithms showed glycine had the most influence on the classification of patient types, with a relative influence of 41.46%. Out of ten metabolites tested, total creatine and scyllo-inositol had no influence on the trees created.