Data Imputation of Brain MRI Features with Enhanced Multinomial Logistic Regression for Alzheimer's Disease Classification

Abstract

Alzheimer's disease (AD) is an immedicable neurodegenerative disease, while the collective cognitive functions deteriorate over time and turns out to be fatal. The analysis of brain's features depicted by Magnetic Resonance Image (MRI) for early diagnosis of AD has led to the development of new diagnostic and therapeutic strategies. The problem of missing data in the experimental datasets predominantly impedes the accuracy of any classification model. However, many of the existing classification techniques do not engage any mechanism to handle such missing data. In this study, we propose a novel Enhanced Multinomial Logistic Regression (EMLR) technique, with imputation to diagnose the stages for AD. This study provides a general classification framework on the stages of AD as mild AD, moderate AD, non-demented, and very mild AD. The input dataset is treated with data imputation to fill in the missing values, transformed by one-hot encoding into vectors, and is fed into the optimized MLR model. The experiments were performed for multi-class classification of AD on OASIS brain MRI dataset features using for simple multinomial logistic regression (MLR) with and without data imputation and EMLR with data imputation and the results are recorded. The classification accuracy of EMLR model has outperformed its competitive classifiers with 88.39% accuracy. Hence, this model can be further developed into a reliable and robust interactive expert system for AD detection and classification.