Data-driven clustering approach to identify novel clusters of high cognitive impairment risk among Chinese community-dwelling elderly people with normal cognition: A national cohort study

Abstract

Background: Cognitive impairment is a highly heterogeneous disorder that necessitates further investigation into the distinct characteristics of populations at varying risk levels of cognitive impairment. Using a large-scale registry cohort of elderly individuals, we applied a data-driven approach to identify novel clusters based on diverse sociodemographic features.

Methods: A prospective cohort of 6398 elderly people from the Chinese Longitudinal Healthy Longevity Survey, followed between 2008-14, was used to develop and validate the model. Participants were aged ≥60 years, community-dwelling, and the Chinese version of the Mini-Mental State Examination (MMSE) score ≥18 were included. Sixty-nine sociodemographic features were included in the analysis. The total population was divided into two-thirds for the derivation cohort (n = 4265) and one-third for the validation cohort (n = 2133). In the derivation cohort, an unsupervised Gaussian mixture model was applied to categorise participants into distinct clusters. A classifier was developed based on the most important 10 factors and was applied to categorise participants into their corresponding clusters in a validation cohort. The difference in the three-year risk of cognitive impairment was compared across the clusters.

Results: We identified four clusters with distinct features in the derivation cohort. Cluster 1 was associated with the worst life independence, longest sleep duration, and the oldest age. Cluster 2 demonstrated the highest loneliness, characterised by non-marital status and living alone. Cluster 3 was characterised by the lowest sense of loneliness and the highest proportions in marital status and family co-residence. Cluster 4 demonstrated heightened engagement in exercise and leisure activity, along with independent decision-making, hygiene, and a diverse diet. In comparison to Cluster 4, Cluster 1 exhibited the highest three-year cognitive impairment risk (adjusted odds ratio (aOR) = 3.31; 95% confidence interval (CI) = 1.81-6.05), followed by Cluster 2 and Cluster 3 after adjustment for baseline MMSE, residence, sex, age, years of education, drinking, smoking, hypertension, diabetes, heart disease and stroke or cardiovascular diseases.

Conclusions: A data-driven approach can be instrumental in identifying individuals at high risk of cognitive impairment among cognitively normal elderly populations. Based on various sociodemographic features, these clusters can suggest individualised intervention plans.

Figure 1

Study flowchart. Panel A. Patient selection. Two-thirds of the 6398 participants included in the analysis (n = 4265) were randomly assigned to the derivation cohort, while one-third (n = 2133) were randomly assigned to the validation cohort. Panel B. Study design. We included 69 features and identified the top 20 most important using LightGBM. Utilising these 20 features, the GMM was conducted in the derivation cohort to categorise participants into four clusters. The 10 most important features from the LightGBM model were selected to build another prediction model in the derivation cohort, which was applied to classify participants in the validation cohort. GMM – Gaussian Mixture Model, LightGBM – light gradient boosted machine

Figure 2

Importance ranking of features. Panel A. Importance ranking of features according to light gradient boosted machine model. The 25 most important features are depicted. Panel B. SHAP values for 20 features. ADL – activities of daily living, BMI – body mass index, IADL – instrumental activities of daily living, SHAP – SHapley Additive exPlanations

Figure 3

Dendrogram and heat map for unsupervised hierarchical clustering. Dendrogram and heat map for unsupervised hierarchical clustering in four clusters based on all the features. Panel A. Derivation cohort. Panel B. Validation cohort.

Figure 4

The predictive model accurately classifies the participants into their inherent clusters. Panel A. The receiver operating characteristic curves of prediction models. The performance of prediction models in assigning every participant to one of the four clusters. Panels B–C. The radar plots represent profiles of the four clusters identified in the derivation cohort (B) and validation cohort (C) based on 10 key features. Ten axes represented z-values for 10 key features. Panels D–E. The bar chart depicts the proportion of cognitive impairment at the three-year follow-up for each cluster in the derivation cohort (D) and validation cohort (E).