Highlight of the study
Among the four different Mach-L methods, RF, SGB, and XGBoost outperformed MLR, identifying education level, age, frailty score, FPG and BMI as the key risk factors for detecting abnormal CFA scores, in descending order of importance.
Mach-L methods have several common characteristics: 1. They do not need hypotheses or assumptions such as normally distributed data sets. 2. They can capture non-linear relationships better than MLR. 3. They can iterate until the best fitting model is obtained. While Mach-L methods have been equated to a ‘black box’, in that their internal operations are not easily perceived, they do outperform MLR in terms of error frequency.
Relationships between education level and CFA score
Our results show that education level is the most important risk factor for CFA, with lower scores significantly associated with lower educational attainment, a finding in line with most major studies. For example, the PAQUID project followed 3675 non-dementia participants for 5 years, finding that the hazard ratio for dementia in no-education and primary-school education participants had significantly higher risk for developing dementia (respectively 1.83 and 1.49 times greater risk their more educated counterparts) [37]. A 6-year longitudinal study in Japan of 51,186 individuals from 346 communities found that low community-level educational attainment was also associated with higher incidence of dementia [38]. At present, it is generally agreed that this positive relationship between cognitive function and education level can be explained by the fact that those with lower education typically have less physical and social resources within their communities. Moreover, low educational level is also related to relatively unhealthy lifestyles and lack of immediate health support or bonding social capital [39]. These are all the plausible underlying causes to explain this relationship.
Relationship between age and CFA score
Consistent with other major studies, age is found to be the second important factor related to CFA score, as aging can cause brain degeneration and injury [40]. The Rotterdam Study of 7,046 participants found that the incidence of dementia increased from 0.6 to 97.2 per 1,000 person-years from the youngest to the oldest 5-year age category [41]. A meta-analysis of 13 studies prepared by Gao et al., also found that dementia increased with age [42]. However, it is important to note that the underlying causes of poor cognitive function are different in younger and older persons. For younger people, the main pathological feature of dementia is more typically related to neocortical neuritis plaques, as opposed to cerebral atrophy for those aged over 95 [43].
Relationship between frailty score and CFA score
Frailty score was found to be the third most important factor for CFA. It is generally recognized that both physical and cognitive function decrease with age. In a cohort of 5,038 participants aged ≥ 55, Szlejf et al., found a negative relationship between sarcopenia and cognitive function (β = -0.20, 95% confidence interval = -0.38; -0.01, p = 0.03) after adjusting for other confounding factors [44]. While their study is cross-sectional, it still provides important evidence given the inclusion of middle-age adults. However, their use of a categorical analysis is less persuasive than a continuous variable analysis. Another study of 665 Chinese older adults (age between 60 to 95 years old) also using MoCA also found a negative correlation between sarcopenia and cognitive ability [45]. Different from the previous study, linear regression was applied and showed that low handgrip strength was associated with worse global cognitive function [45]. The present study also presents a positive correlation (β = 0.243, p < 0.001). The underlying pathophysiology for this relationship could be explained by adverse effects of chronic inflammation, impaired hypothalamic-pituitary axis, poor energy metabolism and oxidative stress [46].
Relationships between FPG and CFA score
The relationship between glucose level and cognitive function remains controversial. In the present study, FPG level was found to be negatively correlated with CFA score in simple correlation, which corresponds with the finding of Yau et al. that older T2D patients with poor glucose control had better functional outcomes. They concluded that, in this age group, glucose control should not be too strict [47]. However, other studies have published opposite findings. Using the same MoCA measurement, Shimoda et al. found that diabetes patients had were more likely to have a MoCA score ≤ 25 (3.2) [48]. However, they did not use linear regression which could quantify the effects of glucose on the MoCA. Zaslavsky et al., studied in 316 participants over the age of 80, also confirming a positive correlation between glucose control and cognitive function (odds ratio, 0.18 points lower). However, this relationship attenuates in older groups. From their results, we might conclude that age plays a role in this relationship, which supports the findings of Yau et al. In the present study, the relationship between FPG and CFA was not significant in simple correlation. However, using Mach-L, FPG was identified as the last important factor to affect CFA. As mentioned in the methods section, the errors were all smaller in all four Mach-L, thus we suggest that Mach-L results are more reliable. Future studies with larger samples and longer time of follow-up are needed.
Relationships between body fat, BMI and CFA score
It is interesting to note that both body fat percentage and BMI are the 5th and 6th important risks for low CFA score in T2D patients. This indicates that BMI and body fat are two independent factors and have different impacts on the pathophysiology of low CFA scores. It should be noted that body fat is the ‘genius’ fat composition of the human body. However, measuring body fat requires specialized equipment, whereas BMI is more easily obtained and is only an ‘estimation’ of human body fat based on body weight and height. This presents a significant drawback for BMI. For instance while bodybuilders have high body weight, most of their body composition is lean body mass. Waist circumference is another important indicator for body fat since it can be regarded as reflecting abdominal visceral fat which is more relevant to actual body fat. This is supported by Flegal et al., who found that WC and BMI are significantly more closely correlated with each other than with percentage body fat (P < 0.0001 for all sex-age groups [49]. Percentage body fat tends to be significantly more correlated with WC than with BMI in men but significantly more correlated with BMI than with WC in women (P < 0.0001). West et al., presented solid evidence for the role of body fat on cognitive function, finding that higher waist circumference was associated with future dementia after 8 year follow-up [50]. At the same time, directly measuring body fat with dual-energy x-ray absorptiometry, the Cardiovascular Health Study-Cognition Study found that higher body fat in men was significantly associated with increased dementia but only marginally associated in women in a cohort of 344 (non-diabetic) participants [51].
As for BMI, its relationship is opposite to that of body fat. Hu et al., followed 44,660 American T2D patients for 3.9 years, finding that higher BMI is associated with lower risk for dementia compared with normal BMI (< 25 kg/m2) [52]. A study in Korea also reached the same conclusion that all-cause dementia risk is lower in people with higher BMI (18.5—23 kg/m2) in T2D patients over the age of 40. The most generally accepted explanation for this correlation is that underweight is commonly associated with poor nutritional status which might result from the poor food intake and digestion [53]. However, the contradictory findings between BMI and body fat require further study with larger cohorts and more precise methods.
The present study is the first to re-evaluates the common risk factors of dementia, particularly in T2D patients using Mach-L approaches. While Mach-L has been criticized for its lack of operational transparency, it still effectively captures non-linear relationships between variables, making it highly useful for medical research. In the future, the use of multivariate adaptive regression splines could potentially provide greater operational insight and visualization.
Despite the improved understanding of the relative weights of risk factors for CFA score provided by Mach-L methods, the present study is still subject to certain limitations. First, the study is based on a relatively small sample, and further studies are needed with larger populations. Second, cross-sectional studies are less persuasive than longitudinal ones, and follow-up with T2D patients over a longer period will supply more information about the impact of these risks on CFA score. Thirdly, the methods used in the present study might be difficult or challengeable to other study group. However, the six most important impact factors identified are reasonable and consistent with previous findings. Lastly, while our study included the Montreal Cognitive Assessment, some participants opted out of the assessment for various reasons, potentially resulting in selection bias, thus caution must be taken when interpretating our results.
In conclusion, the four Mach-L methods could outperform MLR in our present study. Education level, age, frailty score, FPG, body fat, and BMI, were found to the be most important factors related to CFA in an older Chinese T2D cohort. Further study with a longitudinal design is warranted.
