Exploring conditional missing patterns for automated bacteria identification using MALDI-TOF MS data

Abstract

Training classifiers for automated bacterial identification using MALDI-TOF fingerprints requires addressing class-conditional missingness patterns (CMPs). A CMP is a non-missing-at-random pattern that provides evidence for classification. One possible strategy to handle CMPs is feature stratification. This work evaluated the effectiveness of stratification in training naive Bayes classifiers for the proposed task through two experiments. The first experiment compared the predictive performance of categorical naive Bayes classifiers trained on stratified/discretized features with the performance of a Gaussian naive Bayes fitted on imputed data. The second experiment assessed the impact of class imbalance on the differences in the performance of Gaussian and categorical naive Bayes classifiers. The ANOVA results suggest that feature stratification can induce more accurate classifiers. Correlation analysis shows that class imbalance has a low influence on the difference in the performances of classifiers.