Early diagnosis of Alzheimer’s Disease: A multiclass deep learning framework with modified k-Sparse Autoencoder classification

The growing rate of Alzheimer’s Disease (AD) is a significant concern among clinicians who see the urgent need for an accurate and effective computer-aided diagnosis (CAD) method. Especially with sufferers of AD above the age of 65 years, there is a gradual loss of memory, judgment, and functional abilities, including mobility in some cases. The Bayesian method and several other machine learning methods have failed to produce desired results in diagnosing AD. Because of workflow constraints, researchers have not been able to work directly with multiple classes but have been forced to work with binary classifications, despite the likely effects on making accurate assessments.

In the present research, a deep learning method framework based on a modified k-Sparse Autoencoder (mKSA) classification is proposed to assist in diagnosing AD. This is a classification problem. The first step was to tune the mKSA method itself to find the exact suitable sparsity level (m) to produce suitable results on various openly available datasets such as the Mixed National Institute of Standard and Technology Database (MNIST) and NYU Object Recognition Benchmark (NORB).

Then with the use of a deep neural network (DNN) framework, the mKSA is applied to the Alzheimer Disease Neuro-imaging Initiative (ADNI) data sets. Employing an optimised mKSA method based DNN on a large ADNI dataset improves classification accuracy and speed. This then provides the clinician with accurate AD status information. The advantages and limits of the proposed framework are discussed here.

The proposed framework comprises different phases, namely data acquisition; feature extraction; feature selection; and a classification phase for effective classification of Alzheimer subjects. When the mKSA classifier sparsity level is optimised, an improved result in the classification task is observed. The proposed technique is implemented in Python, and a set of experiments is conducted. The mKSA technique is validated using a benchmark dataset of Alzheimer’s Disease, and results are compared with those of representative techniques in the field, namely, ZMS and k-sparse techniques. Analysis and comparison of the reported results indicate that the proposed technique is applicable and in fact superior for accurate detection of Alzheimer disease in the field in terms of detection accuracy, specificity and sensitivity.

The present work amounts to a comprehensive step forward in improving accuracy in detecting Alzheimer’s Disease among subjects, with an experimental result that shows a 2 to 5 percent improvement over the previous methods used on ADNI datasets. The research also provides clues in relation to several research issues that need the immediate attention of the research community in this field.