An Automated Multi-phase Dilated CNN for Pattern Classification in Detection of Dementia

Abstract

This research work proposed an automated Multi-phase Dilated Convolutional Neural Network (MDCNN) based pattern classification approach for early detection of dementia. In this, there are thirteen early symptoms related features are extracted from dementia dataset and their patterns are classified to detect five dementia severity classes like normal, very mild, mild, moderate and severe. First, the imbalanced numbers of samples in each dementia classes’ from a dataset are pre-processed or rebalanced using Synthetic Minority Oversampling Technique (SMOTE) for getting more reliable and accurate performance. This rebalanced outcome is further processed using proposed MDCNN architecture to extract features automatically from input data. These outcome features are further flattened and fed into sigmoid activation to predict targeted class labels. From experimental evaluation, this proposed dementia detection method achieves 15% higher accuracy than machine learning methods and 7% higher accuracy than deep learning methods.