Predicting Cognitive Impairment and Dementia: A Machine Learning Approach

INTRODUCTION

In 2018, it was estimated that 50 million persons were living with dementia worldwide, with an associated annual care costs of approximately 1 trillion USD [1]. As such, dementia poses a major global challenge for health, well-being, and social care [2, 3]. Overcoming this challenge will require confronting considerable gaps in our understanding of the multifactorial etiology of dementia and the lack of a disease-modifying treatment. Even if effective treatments are developed, prevention and risk reduction will still be fundamental strategies [2, 5]. The World Health Organization (WHO) has recently released guidelines on risk reduction of cognitive decline and dementia [3] that point to multiple predictors of disease onset and progression [2, 5–12]. Such a multifactorial etiology prompts questions about the relative importance of individual predictors to be targeted for prevention.

To date only a few empirical studies have compared the relative importance of predictors, and these approaches have mostly relied on meta-analytic methodologies. Barnes and colleagues [13], for example, developed a late-life dementia risk index that identified age and cognitive performance as the most salient predictive factors. The Alzheimer’s Association concluded that the strongest predictors are age, family history, apolipoprotein E (APOE) ɛ4, midlife obesity, midlife hypertension, smoking, diabetes, education, and physical activity [5]. The risk ranking of a mixed-methods study included depression, midlife hypertension, physical inactivity, diabetes, midlife obesity, hyperlipidemia, and smoking [8]. The Lancet Commission on Dementia Prevention, Intervention, and Care estimated that approximately 35% of dementia is attributable to a combination of childhood education, midlife hypertension, midlife obesity, hearing loss, late-life depression, diabetes, physical inactivity, smoking, and social isolation [2].

Evidence from these prior studies shows relative consistencies among the identified risk factors (e.g., smoking, diabetes, physical inactivity, midlife hypertension) [2, 8]. However, results of meta-analytic approaches are vulnerable to sources of bias related to methodological differences across the summarized research (e.g., differences in data quality and sample selection criteria, heterogeneity across instruments assumed to measure the same predictors, potential publication bias). Moreover, interactions among risk factors are notoriously difficult to test in a comprehensive way using classical parametric regression techniques. Therefore, the aggregation of results from such studies may exclude variables that act mainly through indirect pathways (e.g., via amplification of other risk factors’ effects on the given outcome). A clearer picture of such associations may emerge when risk factors are concurrently evaluated using data from the same (population representative) sample. Nevertheless, we were unable to identify any such study in which multiple risk factors for cognitive impairment and dementia were comparatively tested for differences in predictive strength.

To address this limitation in the literature, we used a machine learning approach (random forest survival analysis; RFSA) to evaluate the relative importance of 52 predictors for cognitive impairment and dementia in a large, population representative sample of adults (N = 9,979; aged 50–98 years) assessed over a period of up to 10 years. RFSA is an extension of regression trees, a non-parametric statistical method in which observations are recursively partitioned to identify variables most strongly associated with the outcome of interest [14–16]. RFSA aggregates estimates of predictor-outcome strength across multiple regression trees to estimate each predictor’s variable importance (VIMP) and relative importance (I _rel ) with respect to other tested predictors. Unlike more common approaches based on regression, RFSA implicitly adjusts for all possible linear, non-linear, and higher-order interaction effects. It also provides built-in safeguards against multicollinearity (i.e., predictors so closely interrelated that their effects cannot be disentangled) and model overfit (i.e., model estimates that cannot be replicated in independent samples because they relate also to the noise present in the sample analyzed). However, RFSA was not developed within a standard probabilistic framework. Therefore, we first applied RFSA to a random half subsample (n₁= 4,990) of the original sample, to assess the most important risk factors for cognitive impairment and dementia. Then, in the remaining random half subsample (n₂= 4,989), we examined these most important risk predictors using Cox proportional hazards analysis (Cox PH [17]), which is better suited to interpretation of effect sizes based on a known statistical distribution. This multi-analytic (split sample) methodology allowed us to comprehensively test the relative importance of numerous, interrelated risk factors and to estimate effect sizes for the strongest predictors of cognitive impairment and dementia.

Based on the consensus of aforementioned literature, we expected health-related variables such as diabetes, physical inactivity, and smoking to be the most important risk factors for cognitive impairment and dementia risk. We also expected education [2, 5] and emotional distress [2, 8] to be implicated.

METHODS

RESULTS

The findings of RFSA and Cox PH are shown in Table 2 (cognitive impairment) and Table 3 (dementia). Figure 1 summarizes the most important predictors as determined by the combined methodology of machine learning and semi-parametric survival analysis. Table 4 presents the VIMP ranking of all predictors. The Supplementary Material (S2–S6) contains the findings of the sensitivity analyses. Across all sensitivity analyses, the VIMP rankings were consistent for both cognitive impairment and dementia, and correlation coefficients indicated high consistency (r = 0.67–1.00) between the relative importance of main and sensitivity analyses ran with “randomForestSRC”.

OPEN IN VIEWER

Fig.1

The most important predictors for cognitive impairment (Part A) and dementia (Part B). The relative variable importance (as determined by random forest survival analysis) is graphed on the y-axis, ranging from 0 (lowest importance) to 1 (highest importance). The hazard ratios (as determined by the Cox PH survival analysis) are shown on the x-axis, ranging from 0 to 3. The shape of the marker indicates whether the factors were significantly related to an increased (round) or decreased (square) risk.

DISCUSSION

This study used a combined methodology of machine learning and semi-parametric survival analyses to estimate the relative importance of 52 variables in predicting cognitive impairment and dementia in a large sample of older adults. Results showed that African Americans and individuals with high scores of emotional distress were at relatively highest risk for developing cognitive impairment and dementia. Additionally, health variables (worse subjective health, increasing change in BMI) and sociodemographic variables (lower education, Hispanic ethnicity) were comparatively strong risk factors for cognitive impairment. More commonly studied and identified risk factors from the literature, such as cardiovascular variables and polygenic scores, appeared to be of lesser importance when evaluated concurrently with the above predictors. Multiple post-hoc sensitivity analyses pointed to the robustness of these results.

The finding that African American was among the most important risk factors for both cognitive impairment and dementia accords with past studies [30, 31] that reported a greater dementia risk for African Americans compared to Whites and other ethnic groups. Specific factors contributing to this outcome are difficult to pinpoint, but they are likely related to underlying sociodemographic conditions that affect access to health care [32, 33]. Several of these risk factors (e.g., education, wealth) showed a similarly high relative importance in the RFSA; however, African American was overall among the most important predictors—probably because for predictive purposes it spanned multiple other risk dimensions. Of note, lower wealth was the second strongest risk factor for cognitive impairment in RFSA but was non-significant in Cox PH. It could be that African American had a direct effect, whereas wealth had an indirect effect on cognitive impairment (e.g., via complicated interactions not included in Cox PH but considered in RFSA). Or, it seems also possible that the predictive overlap of wealth and African American did not affect estimates of relative importance in RFSA (because RFSA is robust against multicollinearity), whereas in Cox PH, predictive power shared between wealth and African American was largely subsumed by the latter.

A post-hoc analysis revealed that, indeed, African Americans reported significantly lower wealth (M = $208,000; SD = $391,000) compared to Whites (M = $590,000; SD = $847,000), t(973)  = –17.21, p < 0.001, and this supports the idea of economic disparity as a key factor contributing to risk differences between African Americans and Whites. Lower financial status could limit access to health care and may also negatively impact variables beneficial for building cognitive reserve, such as access to educational resources and environmental settings conducive to mental wellness [34]. Future studies may examine whether contextual factors (e.g., neighborhood as indicator for socioeconomic status) provide additional predictive information concerning African Americans’ elevated risk for cognitive impairment.

Emotional distress was related to increased risk of cognitive impairment and dementia. However, at present, the corresponding literature is somewhat fragmented by the use of numerous negative emotionality measures [5, 36]. In this study, we used factor analytical techniques to create a composite score of emotional distress (i.e., neuroticism, hostility, anxiety, negative affect, hopelessness, pessimism, depression, loneliness, and perceived constraints) that captures a core latent construct underlying these highly interrelated measures. Emotional distress may lead to social and cognitive disengagement, which in turn may exacerbate cognitive decline [2, 9]. Additionally, higher emotional distress is associated with higher cortisol levels [37], which over time may contribute to brain atrophy, loss of cognitive function, and ultimately dementia [38].

Additionally, two health variables (poor subjective health, increasing change in BMI) and two sociodemographic variables (lower education, Hispanic ethnicity) were comparatively strong predictors for cognitive impairment. Subjective health has been shown to be a reliable and robust predictor of mortality risk [14], but its prognostic capacity for cognitive impairment has been rarely examined [39]. Subjective health may reflect complex interrelations between mental, social, functional, and biological aspects of an individual [40]. This higher-order integration might, compared with more narrowly defined risk factors, be especially important for predicting cognitive impairment in older adults [14].

As for BMI, its relation with cognitive impairment/dementia appears to vary with age: A comparatively higher BMI in midlife may increase risk, whereas a lower BMI in later life may be a marker of dementia [41–43]. Our results show that gain in BMI (positive slope) in later life predicts increased risk of cognitive impairment. This suggests that future research may benefit by considering not only stable between-person differences in BMI at a given time or age but also within-person changes in BMI as predictive of increased cognitive risk.

Education had a protective effect on cognitive health, as previously shown in other studies [2, 45]. Higher education may impact cognitive health through multiple and complex pathways [46]. Specifically, higher education is associated with a healthy lifestyle, secure income, supportive relationships, and better general health [47], any of which may confer increased protection against cognitive decline. Considering its potential modifiability, prioritizing education in future research and public policy related to cognitive decline appears very much worthwhile. Along these lines, improved access to education [45] and evaluating whether education after secondary school provides additional protection against cognitive decline [4] are important goals. In addition to education, Hispanic ethnicity was among the most important sociodemographic risk factor for cognitive impairment. The risk for Hispanics was found to be lower than for African Americans but higher than for Whites [30, 49].

To summarize, African American and emotional distress appeared as key risk factors for both cognitive impairment and dementia across both analyses (Fig. 1). Risk factors identified by previous ranking approaches [2, 13], such as smoking, physical inactivity, or hypertension, as well as polygenic scores, were of less importance. It seems likely that these more narrowly defined risk factors were here subsumed by higher-order factors, such as emotional distress or subjective health, as the latter reflect complex interrelations between various aspects of an individual [40].