End-to-End Classification of Cognitive Impairment Using Daily-Life Gait Data

We investigated the feasibility of using commercial smart devices to prescreen cognitive impairments for timely intervention. A total of 125 individuals aged 50 and older were recruited from a local clinic and categorized into groups based on cognitive diagnosis: Mild Alzheimer's Disease (AD), Mild Cognitive Impairment (MCI), and Subjective Cognitive Decline (SCD).This study focused on distinguishing individuals requiring treatment (AD and MCI) from those without cognitive impairment (SCD) using gait-related data collected from daily life. This classification task was particularly challenging, as the SCD group shares cognitive symptoms with MCI and the data was collected in non-controlled, real-world conditions.Participants used smartphones and smartwatches for one month to collect accelerometer and gyroscope data during daily activities. We preprocessed walking segments and trained a deep learning classifier to differentiate between the two groups. The model achieved an area under the curve (AUC) of 0.70, demonstrating the potential of wearable-based gait analysis for cognitive impairment detection. A sensor ablation study revealed that wrist-worn gyroscope data alone achieved a comparable AUC of 0.70, suggesting a moderate association between arm motion and cognitive impairment. The results further indicated that integrating smartphone accelerometers and smartwatch gyroscopes enhanced classification performance.This study is part of a broader multimodal research initiative that integrates gait analysis with voice recordings, phone usage patterns, and other behavioral data to improve cognitive impairment classification. Future work will explore multi-modal fusion techniques to enhance accuracy and reliability, with the long-term goal of developing accessible, real-world screening tools for early detection.