Deep Learning-Based Detection of Cognitive Impairment from Passive Smartphone Sensing with Routine-Aware Augmentation and Demographic Personalization

Early detection of cognitive impairment is critical for timely diagnosis and intervention, yet infrequent clinical assessments often lack the sensitivity and temporal resolution to capture subtle cognitive declines in older adults. Passive smartphone sensing has emerged as a promising approach for naturalistic and continuous cognitive monitoring. Building on this potential, we implemented a Long Short-Term Memory (LSTM) model to detect cognitive impairment from sequences of daily behavioral features, derived from multimodal sensing data collected in an ongoing one-year study of older adults. Our key contributions are two techniques to enhance model generalizability across participants: (1) routine-aware augmentation, which generates synthetic sequences by replacing each day with behaviorally similar alternatives, and (2) demographic personalization, which reweights training samples to emphasize those from individuals demographically similar to the test participant. Evaluated on 6-month data from 36 older adults, these techniques jointly improved the Area Under the Precision-Recall Curve (AUPRC) of the model trained on sensing and demographic features from 0.637 to 0.766, highlighting the potential of scalable monitoring of cognitive impairment in aging populations with passive sensing.

💡 Research Summary

The early detection of cognitive impairment is a paramount challenge in modern healthcare, particularly as the global aging population increases. Traditional clinical assessments, while accurate, are often episodic and lack the temporal resolution required to detect the subtle, progressive changes in behavior that characterize early-stage cognitive decline. To address this limitation, this paper proposes a continuous and unobtrusive monitoring framework using passive smartphone sensing. By leveraging the multimodal data generated by everyday smartphone usage—such as mobility, activity, and sleep patterns—the researchers aim to provide a naturalistic way to track cognitive health without the burden of frequent clinical visits.

The core of the proposed methodology is a Long Short-Term Memory (LSTM) neural network, which is specifically designed to capture the temporal dependencies inherent in daily behavioral sequences. However, a significant hurdle in applying deep learning to such tasks is the high degree of inter-individual variability; every person has a unique daily routine, and demographic factors like age and gender significantly influence behavioral patterns. To enhance the model’s generalizability and accuracy across diverse users, the authors introduce two novel techniques: Routine-aware Augmentation and Demographic Personalization.

Routine-aware Augmentation is a sophisticated data augmentation strategy that goes beyond simple noise injection. Instead of creating random perturbations, it generates synthetic sequences by replacing existing days in a user’s timeline with alternative days that exhibit behaviorally similar routines. This approach ensures that the augmented data remains biologically and behaviorally plausible, helping the model learn robust features of cognitive decline that are invariant to minor daily fluctuations.

Complementing this, Demographic Personalization employs a reweighting mechanism during the training phase. By assigning higher importance to training samples that share demographic characteristics with the target test participant, the model effectively “personalizes” its predictive capability. This allows the LSTM to prioritize patterns that are most relevant to the specific demographic context of the individual being monitored.

The effectiveness of these techniques was validated using six months of sensing data from a cohort of 36 older adults. The results demonstrated a substantial improvement in the Area Under the Precision-Recall Curve (AUPRC), which rose from 0.637 to 0.766 when both techniques were implemented. This significant performance boost underscores the potential of combining domain-specific augmentation with demographic-aware training to create scalable, personalized, and highly accurate digital biomarkers for cognitive health monitoring. This research paves the way for the widespread deployment of smartphone-based continuous monitoring systems in large-scale geriatric care.