PredatorNet: a clinically trustworthy deep learning framework for multi-class Alzheimer's disease classification...

MRI-based staging of Alzheimer's Disease (AD) requires not only high classification accuracy but also clinically reliable probability estimates, transparent evidence for decisions, and robustness to shifts in data distribution. We propose PredatorNet, a multi-class AD classification framework with three key modules: (i) EagleAttention, which uses spatial self-attention to focus on diagnostically salient regions; (ii) RegionImportance, which produces an explicit spatial weighting map to support interpretability; and (iii) WolfPackFusion, which fuses complementary feature pathways to improve prediction stability. The training process requires us to use weighted random sampling because we need to handle class imbalance. We establish the Clinical Trust Score (CTS) as a metric that evaluates clinical effectiveness through the combination of seven performance benchmarks. The test set results show PredatorNet achieves 96.64% accuracy with macro Receiver Operating Characteristic Area Under the Curve of 0.9950 and mean sensitivity of 0.9803 and mean specificity of 0.9876 across 1,280 MRI scans. The model shows strong calibration (Calibration Reliability = 0.9912) and stable explanations (Regional Importance Stability Index = 1.0000). The evaluation method achieves an Out-Of-Distribution (OOD) stability score of 0.2810, which demonstrates its ability to handle distributional shifts. The CTS results show a value of 0.9491 without anatomical weighting and a value of 0.8652 under stricter anatomy-based validation. The Gradient-Weighted Class Activation Mapping visualizations establish a connection between learned importance patterns and neuro-anatomical regions relevant to clinical practice, which supports clinically oriented reporting.