JMIR Publications

ABSTRACT

Background:

As the global population ages, the economic burden of dementia is increasing. Social isolation, encompassing limited social interaction and loneliness, negatively impacts cognitive function and is a significant risk factor for dementia. Populations with subjective cognitive decline (SCD) and mild cognitive impairment (MCI) represent pre-dementia stages where functional decline can be reversible. Thus, accurately predicting social isolation among these at-risk groups is essential, as early identification and intervention can mitigate the risk of further cognitive decline.

Objective:

To develop and validate a model using machine learning (ML) models to predict social interaction and loneliness levels among older adults in pre-dementia.

Methods:

The study included 99 community-dwelling older adults aged 65 and above in the pre-dementia stage. Social interaction frequency and loneliness levels were measured four times a day using mobile Ecological Momentary Assessment (EMA) over two weeks. Actigraphy data was grouped into four categories: sleep quantity, sleep quality, physical movement, and sedentary behavior. This data was analyzed across four different periods. Demographic and health-related survey data collected at baseline were incorporated into the analysis. ML models, including Logistic Regression (LR), Random Forest (RF), Gradient Boosting Machine (GBM), and eXtreme Gradient Boosting (XGBoost) were employed to develop and validate predictive models for groups with low social interaction levels and high loneliness levels.

Results:

Out of 99 participants, 43 were classified into the low social interaction level group, and 37 were classified into the high loneliness level group. RF model was most suitable for predicting low levels of social interaction (accuracy: 0.849; precision: 0.837; specificity: 0.857; and AUC: 0.935). Time-specific analysis revealed that the total activity count in the morning was lower in the group with a low social interaction level than in the group with high social interaction levels (p = .002). The GBM model was most suitable for predicting high loneliness levels (accuracy: 0.838; precision: 0.871; specificity: 0.784; and AUC: 0.887). Time-specific analysis indicated that the sleep quality (fragmentation index) during the night in the high levels of loneliness group was worse than the group with low levels of loneliness (p = .04).

Conclusions:

This study demonstrated the potential of ML-based predictive models using data collected from mobile EMA and wearable actigraphy for detecting vulnerable groups regarding the social interaction frequency and loneliness levels among older adults with SCD and MCI. Our findings highlight the importance of physical movement as a predictor of low social interaction levels and sleep quality for loneliness. The algorithms can be clinically utilized to identify high-risk individuals and provide timely interventions to enhance morning time physical movement and nighttime sleep quality to address social isolation. Ultimately, this can help prevent the progression of cognitive and physical impairments in older adults at risk of dementia.