Hybrid Approach for Alzheimer’s Disease Prediction Using MRI: Deep Feature Extraction and Machine Learning Models

Since Alzheimer's disease (AD) is a neurodegenerative brain ailments characterized by a progressive loss of cognitive abilities, early identification is essential to providing appropriate treatment. Transfer learning on MRI scans is investigated in this study in order to organize AD into four stages: no impairment, very mild, mild, and moderate impairment. To extract significant visual features, pre-trained convolutional neural networks (VGG-16, VGG-19, InceptionNet, and ResNet50) were used. Four distinct machine learning methods were subsequently employed to classify these features: Random Forest, Gradient Boosting, Neural Networks, and Logistic Regression. Metrics like F1-score, AUC, recall, accuracy, and precision were used to assess the predictive models. Neural Networks consistently delivered superior results across all feature sets, with Gradient Boosting performing nearly as well. Logistic Regression and Random Forest also yielded reliable outcomes, though slightly less effective. Among all evaluated models and classifiers, VGG16 combined with Gradient Boosting achieves the highest classification accuracy of 95.7%, along with strong AUC (0.987) and F1-score (0.957). While ResNet50 also performs well, VGG16 with Gradient Boosting emerges as the best overall in terms of raw accuracy and consistency.