Development of a three tiered cognitive hybrid machine learning algorithm for effective diagnosis of Alzheimer’s disease

Abstract

Alzheimer's disease (AD) is one of the most frequent neurodegenerative disorders in the elderly subjects. Since early detection can prevent or delay cognitive decline in the older subjects, it is desirable to develop effectual protocols for the diagnosis of the disease. Most of the existing diagnostic tools fail to improvise timely disease prognosis in susceptible patients. Keeping this fact into consideration, we developed a cognitive-based 3-tiered machine learning (ML) algorithm employing baseline characteristics to predict AD or mild cognitive impairment (MCI) to construct psychometric test results. Earlier machine learning based AD diagnosis methods used a binary or multinomial classification technique. We relied on the development of a sophisticated hybrid cognitive ML algorithm that provides an accurate and precise prediction of the disease. We built an ML model using cognitive and demographic data. The prediction method consisted of a three-step process. Alzheimer’s Disease Neuroimaging Initiative (ADNI) database was used to develop a novel prediction algorithm. Considering the fact that nineteen ML and deep learning classifiers could not adequately classify ADNI data, we created a 2-layer model stacking procedure. Model stacking outperformed six ML classifier combinations, including Logistic Regression, Naïve Bayes, Support Vector Machine, Decision Trees, Random Forest, and eXtreme Gradient Boosting. The performance of the as-proposed model was evaluated employing seven performance assessment measures and four classification error indicators. Each model was evaluated in three separate strategical assessment modules. In the first experiment, XGB, Random Forest, and SVM achieved 89.63% accuracy, while Random Forest achieved 93.90% accuracy in the second experiment. Experiment 2 improved the classification and performance of overall prediction. In the third experiment, hybrid modeling, the accuracy increased significantly, with experiment 1 giving 90.24% accuracy and experiment 2 yielding 95.12% accuracy. The as-proposed model successfully predicted early AD and MCI in an effective manner. We were able to reduce nineteen classifiers into four classifiers (from experiment-1) and six classifiers (from experiment-2) and subsequently into one meta-learner (19 → 4 → 1 and 19 → 6 → 1), with high predictive power. Finally, we performed a thorough comparative analysis of different ADNI datasets to validate our findings.