Improved Alzheimer’s disease detection by MRI using multimodal machine learning algorithms

Dementia is one of the huge medical problems that have challenged the public health sector around the world. Moreover, it generally occurred in older adults (age > 60). Shockingly, there are no legitimate drugs to fix this sickness, and once in a while it will directly influence individual memory abilities and diminish the human capacity to perform day by day exercises. Many health experts and computing scientists were performing research works on this issue for the most recent twenty years. All things considered, there is an immediate requirement for finding the relative characteristics that can figure out the identification of dementia. The motive behind the works presented in this thesis is to propose the sophisticated supervised machine learning model in the prediction and classification of AD in elder people. For that, we conducted different experiments on open access brain image information including demographic MRI data of 373 scan sessions of 150 patients. In the first two works, we applied single ML models called support vectors and pruned decision trees for the prediction of dementia on the same dataset. In the first experiment with SVM, we achieved 70% of the prediction accuracy of late-stage dementia. Classification of true dementia subjects (precision) is calculated as 75%. Similarly, in the second experiment with J48 pruned decision trees, the accuracy was improved to the value of 88.73%. Classification of true dementia cases with this model was comprehensively done and achieved 92.4% of precision. To enhance this work, rather than single modelling we employed multi-modelling approaches. In the comparative analysis of the machine learning study, we applied the feature reduction technique called principal component analysis. This approach identifies the high correlated features in the dataset that are closely associated with dementia type. By doing the simultaneous application of three models such as KNN, LR, and SVM, it has been possible to identify an ideal model for the classification of dementia subjects. When compared with support vectors, KNN and LR models comprehensively classified AD subjects with 97.6% and 98.3% of accuracy respectively. These values are relatively higher than the previous experiments. However, because of the AD severity in older adults, it should be mandatory to not leave true AD positives. For the classification of true AD subjects among total subjects, we enhanced the model accuracy by introducing three independent experiments. In this work, we incorporated two new models called Naïve Bayes and Artificial Neural Networks along support vectors and KNN. In the first experiment, models were independently developed with manual feature selection. The experimental outcome suggested that KNN 3 is the optimal model solution because of 91.32% of classification accuracy. In the second experiment, the same models were tested with limited features (with high correlation). SVM was produced a high 96.12% of classification accuracy and NB produced a 98.21% classification rate of true AD subjects. Ultimately, in the third experiment, we mixed these four models and created a new model called hybrid type modelling. Hybrid model performance is validated AU-ROC curve value which is 0.991 (i.e., 99.1% of classification accuracy) has achieved. All these experimental results suggested that the ensemble modelling approach with wrapping is an optimal solution in the classification of AD subjects.