Hybrid stacking of Squeeze Net features and ML models for accurate Alzheimer’s diagnosis

Alzheimer’s Disease (AD), a leading cause of dementia, demands early diagnosis for effective intervention. This study introduces a hybrid framework combining SqueezeNet—a lightweight convolutional neural network—with ensemble stacking of machine learning (ML) models to enhance AD detection using MRI scans. Leveraging a Kaggle-sourced dataset of 3,382 images across four classes (Mild, Moderate, Very Mild Dementia, and Non-Dementia), SqueezeNet extracts efficient features, reducing computational demands while preserving diagnostic accuracy. A 10-fold cross-validated stacking ensemble integrates predictions from Multi-Layer Perceptrons (70 × 70 and 100 × 100 architectures), XGBoost, Cat Boost, and AdaBoost, optimizing robustness. The proposed model achieves state-of-the-art performance with 94.9% AUC and 89% accuracy, outperforming individual models (e.g., XGBoost: 83.2% accuracy) and addressing class imbalance through rigorous validation. Squeeze Net’s fire module architecture enables efficient computation with fewer parameters, making it suitable for clinical environments with limited resources. A 10-fold cross-validation was used. This helps make the model strong and reliable. The model’s performance was measured using four metrics: F1-score, recall, accuracy and precision. The stacking model performed better than single classifiers. This framework offers a scalable, reliable, and cost-effective tool for real-time clinical AD detection and classification.