The global burden of Alzheimer's disease and related major neurocognitive disorders (ADRD) is substantial and projected to increase significantly with the aging population. Early detection of cognitive decline is crucial for timely intervention and improved quality of life. Current methods for predicting cognitive decline rely on clinical factors such as age, gender, education, and comorbidities, which have limited predictive power. This study explores the potential of using free-living accelerometry data, a readily available and non-invasive measure, to improve the prediction of cognitive decline. Accelerometry captures detailed movement patterns, which may reflect underlying changes in physical and cognitive function. While previous research has shown links between physical activity and cognition, few studies have leveraged the high-resolution data provided by accelerometers to identify unique patterns predictive of cognitive decline. This study aimed to develop and validate a model that uses accelerometry data to forecast cognitive decline in older adults, aiming to improve upon existing methods and offer a new tool for early detection and intervention.

Existing research highlights several demographic, clinical, and environmental factors associated with cognitive decline, including age, gender, education, body mass index, socioeconomic status, history of stroke or diabetes, social environment (network size, social activity, loneliness), and genetic factors such as APOE carrier status. Meta-analyses demonstrate the predictive value of social factors. However, these factors alone often fail to provide accurate predictions for individual trajectories. Wearable sensors, particularly accelerometers, have shown promise in assessing age-related conditions like frailty, disability, and social disengagement, but their use in predicting cognitive decline remains under-explored. Cross-sectional studies showed associations between activity levels and processing speed or executive functioning, and longitudinal studies indicated that higher levels of moderate-to-vigorous physical activity (MVPA) were associated with lower risk of cognitive impairment. However, consistency of these findings across different racial/ethnic groups was inconsistent. The use of high-resolution accelerometry data offers potential for improved prediction by enabling pattern recognition that may differentiate health risks across individuals. Despite the potential benefits, translation of accelerometry into routine clinical care has faced challenges related to inconsistent wear protocols, device body location, and data processing methods.

This study used data from two independent cohorts of older adults without dementia. The first cohort, the FACE Aging study, consisted of 115 participants who wore hip-mounted accelerometers for 7 consecutive days. The second cohort, from the National Social Life, Health, and Aging Project (NSHAP), comprised 575 participants who wore wrist-mounted accelerometers for 72 continuous hours. Both cohorts underwent cognitive assessments using the Montreal Cognitive Assessment (MoCA) at baseline and at follow-up (1 year for the hip cohort and 5 years for the wrist cohort). Accelerometry data were processed to generate a comprehensive set of 98 features reflecting movement patterns, including statistical and harmonic features of counts per minute (CPM) and vector magnitude count (VMC). These features were used, along with demographic and clinical covariates, to train three models using XGBoost: CDPred (demographics and clinical characteristics), CDPred-4 (adding CPM and VMC), and CDPred-4+ (including all accelerometry features). Model performance was evaluated using accuracy and AUC (Area Under the Curve) on hold-out samples. Feature importance was assessed to identify the most influential predictors of cognitive decline.

The study found that the CDPred-4+ model, which incorporated all accelerometry features, significantly outperformed simpler models. For the hip accelerometry cohort (1-year follow-up), the CDPred-4+ model achieved over 85% accuracy in predicting cognitive decline, whereas the simpler models (CDPred and CDPred-4) had accuracies of 75% and 78%, respectively. The AUC for the CDPred-4+ model was 0.86. In the wrist accelerometry cohort (5-year follow-up), the CDPred-4+ model predicted cognitive decline with nearly 70% accuracy (compared to 66% and 67% for the other models), with an AUC of 0.73. Many accelerometry features were more important in distinguishing individuals who experienced cognitive decline than many of the demographic and clinical characteristics, highlighting the unique predictive information contained in the movement patterns. The model's robustness was demonstrated across different wear protocols, device locations (hip vs. wrist), and cohorts.

The findings demonstrate that free-living accelerometry data significantly improve the prediction of cognitive decline in older adults compared to relying on demographic and clinical factors alone. The high accuracy achieved by the model, especially in the shorter-term prediction using hip accelerometry, suggests its potential for clinical application in identifying individuals at high risk of cognitive decline. The model's relative success in predicting 5-year cognitive decline using wrist accelerometry, despite the shorter wear period, highlights the utility of wrist-worn devices for long-term monitoring. The study's strengths lie in its use of two independent cohorts with different device locations and follow-up periods, which enhances the generalizability of the findings. The identified importance of accelerometry-derived features underscores the valuable information contained in movement patterns as indicators of underlying health status.

This study demonstrates the potential of free-living accelerometry to enhance the prediction of cognitive decline in older adults. The developed models show promising accuracy in both short-term and long-term predictions. Future research should focus on refining the models by incorporating additional sensor data (e.g., heart rate), expanding the diversity of the study population, and evaluating the model's impact on clinical decision-making and intervention strategies.

The study acknowledges several limitations. It is not possible to entirely rule out overlap between the two cohorts, although the different recruitment strategies and locations make this unlikely to be a significant issue. The sample sizes, while substantial, may not be large enough to fully assess the model's predictive ability across various sociodemographic subgroups. While effect sizes for different racial groups were examined, the study lacks the power to draw definitive conclusions for these subgroups. Further, the model's accuracy, while high, is not perfect, indicating room for improvement by including additional data sources and more sophisticated analysis techniques.