S Samadzadeh, R Abolfazli, B Khorsand, J Zahiri… – EUROPEAN JOURNAL OF NEUROLOGY, 2018

Sara Samadzadeh, Roya Abolfazli, Babak Khorsand, Javad Zahiri, Seyed Shahriar Arab, Siamak Najafinia, Christian Morcinek, Peter Rieckmann
Publication date
This study was carried out in cooperation between two hospitals (a Multicenter Study). Written informed consent from study participants were obtained. The sample population consisted of 90 healthy control and 182 RRMS patients that were consecutive nonselected patients diagnosed with MS according to the McDonald criteria (2010). Pupillary parameters were obtained and other characteristics such as Physician-confirmed history of optic neuritis were also extracted from patients’ records. Of a cohort of 182 RRMS patients, pupillometry parameters of 100 randomly selected subjects equal to the size of healthy control were selected. For decreasing the side effect of selecting 100 random negative data, we make 100 datasets and train models each time with one of these datasets and report the mean of evaluation measures as a result. For partitioning the data into train and test, a 10-fold cross validation procedure is used. Data is partitioned into 10 equal parts. Each time nine partitions are used for training and one remaining partition is used for testing the model. Average of evaluation measures of the ten testing sets are reported as final.
We have used seven based learners including linear, poly and radial SVM, Random Forest (RF), Decision and Cart tree, K Nearest Neighbors (KNN). For combining the results of the based learners, Ensemble learning is used. In Ensemble learning we use two different methods for predicting the results:• Majority voting: The most frequent prediction is reported as final prediction.• Meta learner: Sending the results of all models as an input of a meta-learner and report the prediction of meta-learner as the final result …


تصویر ثابت