Feature Selection Analysis and Robustness to Missing Data for Sensor-Based Prognostic Modeling of Alzheimer’s Disease Progression

Feature selection (FS) plays a critical role in sensor-based predictive modeling for Alzheimer’s disease (AD), where heterogeneous clinical and neuroimaging measurements generate high-dimensional data with varying degrees of missingness due to incomplete clinical assessment of patients. Effective dimensionality reduction is essential to improve model interpretability, robustness, and generalization performance in sensor-driven healthcare applications. However, a systematic analysis of the interplay between FS strategies, missing-data handling, and prognostic modeling in sensor-derived AD data remains underexplored. In this study, we present a comprehensive and methodologically rigorous evaluation framework for AD prediction using multimodal data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort. We jointly investigate multiple FS techniques and prognostic models on imputed datasets, systematically varying the number of top-ranked features. To ensure robustness, a K-fold cross-validation (CV) procedure is adopted and only features consistently selected across folds (intersection-based stability criterion) are retained. These stable feature subsets are subsequently evaluated on a test set. To further assess robustness to incomplete sensor measurements, we conduct a sensitivity analysis by varying the tolerated missingness thresholds for feature inclusion, reflecting realistic scenarios of incomplete clinical data availability. In this phase, XGBoost is employed both as a prognostic model and as an embedded FS method, exploiting its native capability to handle missing values and to provide feature-importance rankings based on predictive contribution. The proposed framework enables a systematic assessment of FS stability, predictive performance, and resilience to missing sensor data. Results provide practical methodological guidelines for the development of reliable and generalizable sensor-driven prognostic models for AD in real-world clinical environments.