Multiclass prediction of Alzheimer’s disease using balanced multimodal data and deep ensemble learning

Discerning Alzheimer disease (AD) in its early stages, before the clinical symptoms, can help mitigate its progression and slow down the brain tissue damage through appropriate treatment. In this regard, a computer-aided diagnosis (CAD) system can be useful in the neurologist toolbox, as it may help make a timely diagnosis and optimize the treatment of Alzheimer’s patients. Over the past decade, researchers have vested a lot of efforts on deep learning (DL) techniques for AD diagnosis, but few contributions have exploited the benefit of ensemble classification (EC) in DL. Despite the advances, the scientific community is still looking for reliable and robust approaches to identify early stages of AD. The paper aims to propose a complete CAD system that takes into account heterogeneous and unbalanced datasets to generate a robust classifier of AD stages. An original approach is presented at the end of the deep ensemble cooperation to predict AD and which allows managing the effect of class imbalance. Well-known convolutional neural network models (CNNs) are used as feature extractors with evolutionary-based hyper-parametric optimization to find the adequate architecture and enhanced momentum based optimizer to compile the CNNs. EC techniques are used as classifiers to reduce the probability of selecting a bad classifier among the basic models using an original weighted probability criterion. The ensemble classifiers integrated an original resampling strategy to rebalance the distribution of classes which enhanced performance of the minority class. The efficiency of the CAD system is evaluated on multimodal fusion data from the public OASIS and ADNI datasets, using 5-fold cross validation and nine confusion matrix-based metrics. The obtained success rate is encouraging when compared with state-of-the-art machine learning models, with all the implemented deep EC models showing acceptable performance. The experience has shown sensitivity rates of (94.05, 82.1, 96) % for Healthy control (HC), Mild cognitive impairment and AD stages of ADNI dataset, and (97.35, 85.35, 90.05, 93.92, 94.4) % for HC, Very mild and Mild impairment as well as Moderate and Severe dementia stages of OASIS dataset. Overall accuracy rates of (95, 99.1) % and (94.7, 98.2) % are obtained for multi-class and binary classification of ADNI and OASIS datasets respectively. The performance results of the test data are more precise and reliable and are superior than most of the reviewed state-of-the-art works. Moreover, the CAD system maintains good performance using an external ABIL dataset with an accuracy of 97.02% which reflect a generalization capability of the proposed approach justifying its adoption in real-world clinical contexts.