Alzheimer's Disease Surveillance Using CDC Data and Hybrid Ensemble Models

Alzheimer’s disease (AD) is a problem that will be very hard to deal with globally in public health, but most AI (Artificial Intelligence) solutions for AD prediction are still based on neuroimaging methods (e.g., MRI and PET) that are very expensive. This restricts the use of AI for monitoring at the population level. Hence, there is a need for predictive frameworks that can rely on public health data, which are available in tabular format, to forecast the trends of diseases like AD affecting the population. In this research, we put forward a Dual-Stream Hybrid Ensemble Framework for the adoption of Alzheimer's disease predictive methods by longitudinal CDC data (2015–2022). The research works on checking the ability of modern deep learning architectures to outperform the strong Gradient Boosting Machine baselines on complex epidemiological tabular datasets. The framework consists of a Feature Tokenizer Transformer (FT-Transformer) joined with three leading GBMs (XGBoost, LightGBM, CatBoost) and a deep Multi-Layer Perceptron (MLP). To maintain the validity of the scientific findings, we impose a strict “Chinese Wall” feature segregation rule to avoid data leakage by disabling the use of context-specific proxy identifiers in the binary classification stream. Furthermore, Probability Target Encoding and Cost-Sensitive Learning methods are put into practice to tackle high-cardinality categorical features and severe class imbalance issues (~1:9.5). The outcomes of the experiments indicate that the FT-Transformer provides the best regression performance with an R-squared value of 0.9451, while XGBoost gives a lower value of 0.8752, and the final weighted regression ensemble provides an R-squared value of 0.9394. For classification under leakage constraints, the cost-sensitive ensemble reaches high sensitivity at the cost of precision at Recall = 89.75%, thereby indicating a demonstrates utility for identifying AD/SCD-related surveillance strata even though metadata proxies have been removed. To sum it up, the new research results give empirical evidence that deep learning by transformers can outperform classic boosting methods for public health monitoring of Alzheimer’s disease if the right tokenization and training restrictions are applied.