A Robust Optimized Clinical Ensemble Framework for Reliable Detection of Alzheimer's Disease with Threshold Sweep

Early detection of the Alzheimer's disease (AD) is critical because it is imperative to prevent further development of the disease, provide care to the patient and slow the progression of the disease. Traditional diagnostic methods rely on MRI and PET images, which are not only expensive but also have major infrastructural requirements and cannot be implemented globally in the primary care and community setting. This study introduces an ensemble of optimized frameworks of learning based on XGBoost and CatBoost, which are tuned according to the isotonic regression and optimized using a hyperparameter optimizer such as Optuna. This approach solely relies on cheap clinical variables in the ADNI dataset such as age, MMSE scores, APOE genotype, education, and demographic variables. A threshold study is conducted over 90 operating points to obtain 0.27 as an ideal choice cutoff, which can give the best precision while still retaining recall of 0.90 or above. The model has an AUC of 0.9446, accuracy of 0.8175, precision of 0.5435, recall of 0.9259, and F1 of 0.6849 on the test set held out. Comparative trials indicate that the proposed calibrated ensemble with threshold optimization is much more effective than any other models with traditional decision thresholds, especially when it comes to the identification of AD-positive cases. This system can be used to generate clinical interpretability based on feature importance analysis, has low computational requirements (i.e., no GPUs) and can run on regular clinical equipment. All these characteristics make it suitable in large-scale screening of AD in telemedicine, rural hospital facilities, and primary healthcare services.