Wearable sensing for quantifying cognitive and balance functions in naturalistic movements of older adults with mild cognitive impairment in therapeutic environments

Mild cognitive impairment (MCI) is a clinically important stage preceding Alzheimer's disease and related dementias, in which cognitive and balance functions are commonly evaluated using standard clinical assessments such as the Montreal Cognitive Assessment (MoCA) and Mini-Balance Evaluation Systems Test (Mini-BESTest). These assessments are administered episodically by clinicians and may miss functional changes during everyday movement. Recent studies and prior work in the Charlie and Harriet Shaffer Cognitive Empowerment Program (CEP), a therapeutic environment supporting lifestyle intervention and naturalistic social interaction, suggest that wearable and passive behavioral sensing can monitor movement patterns associated with cognitive and balance function in older adults with MCI. However, it remains unclear whether passive waist-mounted IMU data collected during naturalistic movement and social interaction can quantify clinician-rated cognitive and balance outcomes, particularly at the subdomain level, in an interpretable and demographically fair manner. To address this gap, we analyzed weekly IMU recordings collected over 6 months from 44 older adults with MCI in the CEP and trained tree-based ensemble regression models to estimate MoCA and Mini-BESTest total and subdomain scores, with interpretability and demographic fairness evaluation. Our models achieved RMSEs of 3.677 for MoCA and 3.672 for Mini-BESTest, benchmarked against Minimal Detectable Change and Minimal Clinically Important Difference thresholds. Feature importance analysis showed distinct movement signal properties across assessments, with general movement intensity features most informative for MoCA and temporal gait features led by cadence most informative for Mini-BESTest. Demographic bias analysis identified sex-related model bias, mitigated through post-processing while maintaining performance. This study supports the feasibility of wearable-based estimation of clinical assessment scores in older adults with MCI during naturalistic activity, with comparable performance between sexes after bias mitigation. This advances the validation of passive sensing for home monitoring to support clinical decision-making and personalized interventions.