Advancing Alzheimer’s Disease Prediction with Optimized Deep and Active Learning Approaches

Abstract Alzheimer’s Disease (AD) is a progressive neurological disorder primarily affecting cognitive and memory functions. Artificial Intelligence (AI)-based early detection systems are needed to address the increasing incidence of AD. However, the traditional detection systems suffer from several challenges, including class imbalance, insufficient hyperparameter tuning, the high cost of data annotation, and the lack of interpretability, resulting in the poor predictive performance. To solve these limitations, Gaussian noise was used to oversample instances of the minority class in this study. In addition, we introduce three new models: Convolutional Highway Neural Network (CHNN), Uncertainty based Active Learning Convolutional Highway Neural Network (UAL-CHNN) and Particle Swarm Optimization based Convolutional Highway Neural Network (PSO-CHN). The CHNN adopts convolutional neural network and highway network to make the model simultaneously learn local spatial dependency and long range feature interaction. PSO-CHNN further expands this architecture, PSO is used for optimizing hyperparameters so that the model can learn more effectively and stably. And the UAL-CHNN combines UAL with CHNN by using the active learning for reducing the annotation cost while improving both efficiency and predictive accuracy via selectively labeling the most uncertain ones. Moreover, the generalizability of proposed models across various data splits is ensured by employing 10-fold cross-validation to further validate their robustness. The superiority of proposed models over traditional architectures is further demonstrated by the Wilcoxon signed-rank test, which confirms statistical significance. To provide feature importance analysis and models’ interpretability, the explainable AI techniques, local interpretable model-agnostic explanations and Shapley additive explanations are utilized. The proposed models CHNN, UAL-CHNN, and PSO-CHNN outperform the base and existing models, achieving improvements of 2.63%, 6.58%, and 5.62% in accuracy; 5.47%, 8.39%, and 7.91% in precision; 2.75%, 5.15%, and 4.91% in F1-score; 5.45%, 14.87%, and 12.81% in Matthews correlation coefficient; 5.79%, 15.90%, and 13.82% in Cohen’s Kappa; and a reduction of 13.41%, 33.54%, and 28.66% in hamming loss, respectively. These findings confirm that proposed models can improve the identification of AD at the early stage, through resolving crucial problems in conventional methods. Integrating explainable AI methodologies improves transparency of the model while ensuring its trust and clinical impact.