Multiclass classification under prioritized error control and specific error costs with applications to dementia classification

点击翻译按钮获取中文摘要

Project Summary Fifty million people worldwide have dementia, and there are nearly 10 million new cases every year. The subtypes of dementia include Alzheimer's disease (AD), vascular dementia, Parkinson's disease (PD), dementia with Lewy bodies, and a group of diseases that contribute to frontotemporal dementia. Early and accurate diagnosis of the dementia cause is crucial because it can lead to the timely provision of symptomatic treatment and avoidance of medications that may worsen symptoms and assist in developing and evaluating new drugs and gaining access to effective treatments when they become available. There has thus been extensive research on developing accurate classifiers (e.g., linear discriminant analysis, support vector machines, multiclass logistic regression, random forests, boosting, convolutional neural network) to automatically classify dementia into different classes. In most existing work, the classifiers are designed to maximize overall accuracy. Since different types of classification error may have different consequences (costs), it is highly desirable to develop multiclass classifiers with prioritized error control. There is a rich literature on binary classifiers that minimize the false negative rate, with false positive rate controlled under a particular level. A prominent example is the Neyman-Pearson classification framework. However, the extension of the framework to multiclass classification, in conjunction with the desired prioritized error controls, while of vital importance, remains largely unknown. This project fills this gap by developing a general framework for multiclass classification with controls for misclassification errors, while imposing various (relative) costs for another set of misclassification error types. This can be viewed as a unification of cost-sensitive learning and the Neyman-Pearson classification in the multiclass setting. The new methodological development will optimize clinical diagnosis in the multiclass setting and expedite drug discovery through more efficient clinical trials. Our specific aims are: Aim 1. To propose a flexible framework that includes prioritized error control requirements as well as costs for various classification error types, and develop an efficient umbrella algorithm to solve the associated constrained optimization problem. Aim 2. To study the feasibility of the optimization problem and the properties of the umbrella algorithm. Aim 3. To evaluate the proposed algorithm via extensive simulation studies, apply it to dementia subtype classification, and develop a publicly available R package.