STATISTICAL EVALUATION OF AN EXTENSION OF LINEAR DISCRIMINANT ANALYSIS IN MULTICLASS SETTINGS WITH AN APPLICATION FOR DETECTION OF CERVICAL NEOPLASIA

FRANCES A. BRITO, UTHealth School of Public Health

Abstract

Although a pathologists’ review of Papanicolaou smear cell samples has been successful in decreasing cervical cancer incidence, it is often costly and time-consuming. Quantitative cytology is a promising semi-automated method that measures cell features for further analysis or classification. There have been several advancements in classification algorithms, but many do not account for the nested data structure seen in quantitative cytology. Further, histologic diagnoses are separated into five or more classes, yet, multi class classification has not been investigated. Here, we compare the predictive performance of macrolevel discriminant analysis (MDA) to traditional discriminant analysis methods in multi-class settings on cervical quantitative cytology data and simulated data sets. MDA had similar overall classification accuracy and area under the ROC curve results to linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA) when applied to cervical quantitative cytology data. However, MDA has a tremendous advantage over LDA and QDA methods when a macrolevel (patient or individual) effect is assumed and when one class is composed of a mixture of gaussian distributions.