Wearable Sensor-Based Mild Cognitive Impairment Identification: A Multi-Domain Gait Analysis Approach With...

Mild Cognitive Impairment (MCI) involves cognitive decline beyond age-expected norms without impairing daily function. Crucially, it often precedes dementia, making early identification vital for timely intervention. MCI identification remains challenging with conventional clinical methods. Existing studies have shown a significant association between motion behavior and cognitive performance, highlighting the promise of Inertial Measurement Units (IMUs) for motion-based portable MCI screening. Although some studies in the field of MCI identification have explored feature extraction based on IMU data, these studies mostly focus on time-domain and frequency-domain features and have not systematically analyzed features from multiple domains, which may limit a comprehensive understanding of MCI-related movement patterns for screening. Additionally, the reliance on feature importance rankings in these studies often lacks the interpretability needed for direct application in clinical decision-making, thus limiting their clinical applicability. This study proposes an integrated framework combining multi-domain gait analysis with explainable machine learning. Waist-mounted IMU was utilized for data collection during standardized Short Physical Performance Battery (SPPB) assessments in older adults. Comprehensive features were extracted across four domains, including time, frequency, nonlinear and time-frequency. A hybrid feature selection strategy was then employed for reducing dimensionality while preserving both linear and nonlinear indicators relevant to MCI. Additionally, association rule mining further identified key diagnostic feature combinations strongly linked to MCI. We conducted a field study and collected real-world data for performance evaluation. Our proposed approach achieved a mean test area under the ROC curve (AUC) of 0.897 and a mean test accuracy of 0.828 with the SVM classifier. The approach demonstrates significant potential for cost-effective and generalizable IMU-based MCI screening tools.