Abstract

Early identification of Mild Cognitive Impairment remains a critical challenge in ageing populations, particularly where access to neuroimaging and specialist assessment is limited. Behavioural digital biomarkers derived from unobtrusive smart home monitoring offer a scalable and non invasive alternative; however, model development in this context is constrained by small, imbalanced datasets and minority class scarcity. This study presents a methodological evaluation of imbalance aware machine learning strategies for Mild Cognitive Impairment detection using nocturia derived behavioural features under extreme minority representation. A unified experimental framework integrates classical models, ensemble methods, autoencoder based transfer learning, adaptive synthetic augmentation, and active learning within a leakage controlled cross validation protocol. Performance is assessed using diagnostic metrics and 99% confidence intervals to quantify uncertainty. Results indicate that augmentation and tree based ensemble strategies improve minority detection compared to baseline classifiers, although performance variability remains substantial. The highest achieved minority-class Sensitivity reached 0.53 under ADASYN-based augmentation, with substantial variability across folds (standard deviation up to 0.39), reflecting instability inherent to extreme minority scarcity. Rather than claiming clinical validation, this work characterises reproducibility and uncertainty under constrained data conditions, providing a reproducible framework for scalable Internet of Things based cognitive screening research.