Bootstrapped Multinomial Logistic Regression on Apnea Detection Using ECG Data

In designing a classification system, one of the most important considerations is how optimal the classifier will adapt and give best generalization when it is given data from unknown model distribution. Unlike linear regression, logistic regression has no simple formula to assess its generalization ability. In such cases, bootstrapping offers an advantage over analytical methods thanks to its simplicity. This paper presents an analysis of bootstrapped multinomial logistic regression applied on apnea detection using ECG data. We examine multinomial logistic regression to detect or recognize multi-class apnea categories (heavy, middle, healthy). We show that for generally complex and highly unstructured medical data such as ECG for apnea detection, bootstrapping gives more meaningful assessment to detect over-fitting than k-fold cross validation. The bootstrapping also gives higher classification accuracy prediction over k-fold cross validation for the same training data proportion.