Individualized subtypes prognosis of MCI using deep‐learning based jointly optimized clustering‐classification method

Abstract

Background The heterogeneity of Alzheimer’s disease is important theme for studying the epidemiology of the disease. Practically, understanding disease phenotypes helps deciding treatment before proceeding to dementia. Generally, it is known that AD’s subtype can be divided into three. Method We used 660 mild cognitive impairment (MCI) patients’ T1‐weighted magnetic resonance images (MRI) from Alzheimer’s Disease Neuroimaging Initiative (ADNI). Inputs are atrophy vectors which mean excluding the positive part and resigned W‐scores calculated from 81,924 vertex‐level cortical‐thicknesses extracted from T1. Unlike existing clustering‐only or separated clustering‐classification models, we used a model which is jointly optimized with respect to both clustering and classification. The model is comprised of three parts: 1) feature extractor, 2) clustering module, and 3) classifier. In first part, we extract important features using simple neural network. Then, we do clustering using the deep‐features and obtain pseudo‐labels in latent space. Finally, we implement classification at last layer with the answers which are assigned from clustering module. It is end‐to‐end training being able to find the solution of combined objective function. Results Results were represented from 10 fold cross‐validation. Average log‐likelihood of Gaussian mixture and normalized mutual information of labels between adjacent iterations are used as clustering performance metric, cross‐entropy loss and accuracy as classification. It is difficult for such models to converge according to referenced studies. But, in this study, all the metrics were well‐converged. Atrophy maps show group comparison results of cortical thicknesses between each subtype and cognitive normal (CN) patient with FDR correction and W‐score averaged in each subtype. The demographic table about Aβ positivity and dementia conversion percentages without missing information and neuropsychological test scores table of each subtypes showed that subtype‐I has highest degree of disease severity, then subtype‐II and finally III. External validation with 662 aMCI patients’ data from Samsung Medical Center (SMC) used in similar study represented bigger gap of dementia conversion percentages between subtype‐I and II. Conclusions In this study, we made clustering‐classification model for individualized subtype prediction. It is end‐to‐end training method optimizing clustering and classification loss simultaneously. The results showed our model exceeds existing subtype clustering model.