Dizziness in Cognitive Impairment (DCI) has recently been described as a common syndrome in elderly patients with chronic dizziness, balance problems and instability of stance and gait [
4]. However, testing for cognitive impairment is still often neglected in the current diagnostic algorithms for dizzy patients. Therefore, we aimed to establish a prediction model for cognitive function and impairment in elderly patients, based on commonly available clinical features from demographics, vestibular and gait assessment, as well as brain imaging. We demonstrate that prediction models considering features from all four routine assessment categories yielded a high and robust classification of cognitive impairment with an AUC of 0.84
\(\pm\) 0.14 and were able to explain about 50% of the cognitive variance in our patient cohort. We further show that considering clinical features from one routine assessment category only, considerably limits model performance. The proposed model could help to appropriately and efficiently select dizzy patients in need of neuropsychological testing and training.
Rationale behind the model feature selection
The machine learning approach in this study selected features well discussed for their potential in the prediction of the risk of cognitive decline. The parameters with a significant effect in the model ranked from highest to lowest predictive contribution are age > CVRF (demographics), Romberg’s sign > caloric response (vestibular), TUG > dual task cost velocity > FGA > dual task cost cognitive (gait) and GCA-F > MTA > IOA > FAZEKAS (imaging).
Demographic factors such as older age as well as CVRF are associated not only with vascular but also neurodegenerative dementia [
26‐
28]. Previous studies were able, for instance, to show the influence of older age on fluid cognition (i.e., the ability to process and learn new information and adapt to new circumstances) such as processing speed, working memory and executive cognitive function [
8].
Vestibular contribution to cognitive functioning has been attracting increasing attention in recent years. For instance, brain networks involved in postural stability such as the cingulo-opercular, fronto-parietal, and somatosensory-motor networks are known to play a substantial role in normal cognitive performance [
29]. Therefore, it is plausible that patients with morphological or vascular cortical pathologies within these shared brain networks not only exhibit a dynamic balance disturbance but also cognitive dysfunction, as shown in the recently described DCI syndrome [
4]. Kido et al. for example, showed that postural instability is related to pathological cognitive decline and also neurodegenerative diseases [
10]. In line with the recent study introducing the DCI syndrome [
4], the degree of caloric response in the current study displayed a negative correlation with the MoCA scores (i.e., the lower the caloric response, the higher the MoCA scores). A possible explanation for this observation could be that lower caloric responses filter out patients with clear peripheral vestibular deficits leaving those with central pathologies, which in case of cortical affection of the multisensory vestibular processing network can also manifest with a concomitant cognitive dysfunction [
4]. In addition, central pathologies may display a higher caloric response, based on disinhibition of cortico-cerebellar networks. At first sight, the finding of reduced peripheral vestibular function in cognitively unimpaired patients seems to partially contradict the previously reported deficits in visuo-spatial domains in patients with bilateral vestibulopathy [
30]. However, these deficits in single cognitive subdomains are rather mild and may not be reflected optimally by MoCA screening. Importantly, the current study does not claim any causality between vestibular test results and the cognitive status, but only reports markers associated with presence of a cognitive impairment.
Gait patterns have also been shown to give insights into cognitive dysfunction and higher cortical and subcortical network pathologies [
11]. A complex network of brain regions is involved in the control of locomotion and varies depending on gait speed and cognitive demand [
31]. Cognitively impaired patients exhibit higher locomotor costs in dual-task conditions and an increased variability of time-related gait parameters [
11]. These gait pattern changes even precede cognitive decline and might therefore be used as risk markers for early identification of dementia [
12]. This condition has been previously termed motor cognitive risk syndrome [
32]. The TUG test is also associated with cognitive functions, especially executive control, memory and processing speed [
33]. Similarly, Pavlou et al. also depicted a correlation of cognitive function with FGA as well as dual task costs in chronic dizzy patients [
34].
Brain imaging parameters such as focal cortical brain atrophy and white matter lesion load are well-known predictors for cognitive impairment [
4]. Among the included visual atrophy scales, the GCA-F (addressing frontal lobe atrophy) is well-ascribed to higher cognitive functions including but not limited to language, working memory, problem solving, decision-making and behavior [
35]. The MTA scale, on the other hand, assesses mesio-temporal atrophy, a well-established region for episodic and long-term memory and, more recently, for perception and attention [
36]. There is also growing evidence highlighting the role of the insula in general cognitive functions such as language, perception, attention and working memory [
4].
Comparison with other cognition prediction models
The prediction models from this study performed comparably and even superior to other previously reported algorithms in the risk prediction of cognitive impairment, although differences in patient collectives limit a direct comparison. A gait feature-based model for detecting cognitive dysfunction in the elderly utilizing a single wearable inertia sensor yielded an AUC of 0.73–0.88 [
37]. A standardized evaluation of multiple cognition prediction algorithms based on imaging data showed that the best AUC achieved was about 0.79 [
38]. In comparison, our study revealed an AUC of 0.66 for a combination of gait parameters and an AUC of 0.71 for a set of semi-quantitative white-matter lesion and atrophy imaging parameters (Fig.
3A). A recently proposed machine learning algorithm for detecting cognitive impairment based on multiple features reflecting demographic, clinical, psychological and lifestyle aspects achieved an AUC of 0.73–0.83 [
39,
40]. In our binary logistic regression model, the combination of different feature categories resulted in a superior predictive value (AUC of 0.84). It should be noted that we applied features to the model, which were already available from routine data sources and could be quantified without major data preprocessing. The proposed models therefore should be suitable for application in a clinical setting of risk prediction. A combination of feature sources was superior to single feature source for our cohort.