A Brief Neuropsychological Battery for Measuring Cognitive Functions Associated with Obesity

Abstract

Objective

Although ample evidence links obesity to cognitive dysfunction, the trajectory of cognitive change, the underlying mechanisms, and the involvement of related factors, such as metabolic disease and diet remain unclear. To support further investigations of body mass index (BMI) and cognition, we aimed to create a concise test battery to be used in future trials.

Methods

We regressed 20 neurocognitive measures on BMI in the Human Connectome Project Healthy Young Adult S1200 data release using Linear Mixed Models and adjusting for major confounders. We then identified measures using LASSO regression analysis to select tests most strongly associated with BMI. To guide further test selection we visualized the explained variance for each variable in the final model.

Results

BMI was negatively associated with 7 neurocognitive measures. Variable selection yielded a model including years of education and -in order of model weight- Delay Discounting, the Relational Task, Penn Progressive Matrices, Oral Reading Recognition, the Variable Short Penn Line Orientation and Penn Word Memory.

Conclusion

This resulted in a ~ 40--minutes test battery for the BMI-cognition relationship in young adults that can be used in trials investigating the interrelationship between obesity and cognitive performance.

Introduction

Increasing evidence links obesity with neurocognitive impairment (1, 2). While the influence of midlife obesity on risk of dementia has been established (3, 4), there is also evidence that adiposity is associated with cognitive function in adolescence and young adulthood. Some impairments are likely pre-existing and contribute to initial risk (1, 5), while others may worsen or emerge as a consequence of pathophysiological processes associated with obesity (6). For example, impaired glucose tolerance and inflammation are detrimental to neurocognitive function (6–9) and there is now firm evidence that even short term dietary manipulations can impact cognition (10, 11). Future work is needed to disentangle cause from consequence, identify specific mechanisms, better understand the trajectory of cognitive change and evaluate the possibility of reversal (6).

A major barrier to achieving this goal is that the nature of cognitive change or “neuropsychological profile” is unclear (1, 2) and administration of a comprehensive neuropsychological battery is often not feasible due to time constraints. The literature reveals significant inconsistency in cognitive functions associated with obesity (1, 2, 12). Two systematic reviews showed that despite some evidence for impairment in most cognitive domains, the evidence was generally inconclusive (1, 2). One review indicated that only some tasks of response inhibition, working memory, and decision making are consistently associated with BMI and have reliable psychometrics (1), while the other found no evidence for working memory impairment in obesity and reported conclusive evidence for impairment in the domains of visual reconstruction and set-shifting (2). A meta-analysis of Yang et al. (12) showed a significant effect of obesity on six executive functioning domains: inhibition, cognitive flexibility, working memory, decision making, verbal fluency, and planning. Notably however, a moderator analysis showed that the effects of obesity where not significantly moderated by BMI.

The inconsistencies in the literature may be partially due to methodological limitations, including small samples sizes, limited exclusion criteria (e.g. no exclusion of persons with history of neurologic disease, psychiatric disease or substance abuse) or and/ or absence of an appropriate comparison group (2). Limited adjustment for confounding factors makes it impossible to ascertain whether reported associations of obesity with cognition are independent or an indirect consequence of related factors that may also affect cognition (e.g. co-morbidities, diet, age or socio-economic status (SES)). Moreover, few studies include comprehensive neuropsychological testing, making it difficult to determine if reductions in cognitive test performance span multiple domains or if they represent specific impairments in the context of a globally intact profile.

In the current study we overcome some of these limitations by capitalizing on the neurocognitive data from the WU-Minn consortium Human Connectome Project (HCP) which includes 25 neurocognitive measures from over 1200 healthy young adults, as well as critical demographic data, including BMI, sex, age, socioeconomic status and measurement of a number of important confounders including depression, sleep and alcohol use (13). The choice of this sample should aid us in our goal to disentangle effects of BMI from those of confounding factors such as age-related cognitive decline, co-morbidities (such as T2D) and SES. This approach was adopted in a recent analysis in the Human Connectome Project S900 release and indicates that brain morphometric measures as well as cognitive performance in multiple domains are correlated with BMI and that these traits are largely heritable(5). Extending these findings, we adopted a data driven approach to inform the construction of a concise test battery optimized for assessing the relationship of BMI with cognitive function.

We first evaluated the association between BMI and neurocognitive outcomes in separate univariate Linear Mixed Models (LMMs). Then, we applied LASSO regression, a variable selection method using regression modeling to aggressively remove correlated variables thereby selecting neurocognitive variables with the best unique prediction capabilities for BMI. For each test in the resulting model, we determined the explained variance to provide a guide for further test reduction if needed. This resulted in a test battery with a maximum length of 38 minutes including up to 6 tests: Delay Discounting, the Relational Task, Penn Progressive Matrices Test, Oral Reading Recognition Test, Variable Short Penn Line Orientation Test, and Penn Word Memory test.

Methods

Data collection

Data were obtained on December 4^th, 2018 from the HCP Young Adult S1200 data release containing data from 1206 adults aged 22–35 years (13). Detailed inclusion and exclusion criteria are reported by van Essen et al. (13). Briefly, the cohort includes sibling and twin pairs as well as individuals who smoked, were overweight, or had a history of heavy drinking or recreational drug use without having experienced severe symptoms. The HCP Young Adult study excluded individuals having severe neurodevelopmental disorders or that were born prematurely (e.g. autism), documented neuropsychiatric disorders (e.g. schizophrenia or depression) or neurologic disorders (e.g. Parkinson’s disease), and chronic illnesses such as Type 2 Diabetes (T2D) or high blood pressure. In addition, we used the following exclusion criteria for our analysis: missing BMI data, being underweight (BMI< 18.5), and having a positive alcohol or drug tests on any of the test days. The remaining sample for our analysis included data of 1027 subjects from 426 different families (Figure 1). From the dataset we extracted BMI and neurocognitive test data (detailed below). We also identified and extracted 8 variables in the dataset that were potential confounders owing to their association with cognitive performance or BMI and on recommendations in the literature (2): age, gender, race, ethnicity, family ID, yearly household income, years of education, depression score (Achenbach Adult Self Report, ASR, DSM IV), sleep quality score (Pittsburgh Sleep Questionnaire Total Score), and whether participants met DSM IV criteria for alcohol abuse (see 18, 19 for a detailed description of these parameters).

Figure 1:

Flow diagram of study inclusion and data-analysis.

LASSO = least absolute shrinkage and selection operator regression analysis, NPT= Neuropsychological Test outcomes

Neurocognitive test data

Neurocognitive testing was performed over two days at Washington University as part of a broad behavioral test battery and two neuroimaging sessions (15). Parameters for our analyses were identified using literature from the WU-Minn Consortium and the HCP Data Dictionary 1200 Subject Release (13–15). Table 1 provides a summary of the neurocognitive measures used in our analysis, the cognitive domains they represent and the estimated test duration. A detailed description of the tests and references to source literature are available in the Supporting Information.

Table 1:

Neurocognitive measures included in the analysis

Test	Cognitive domain	Time (min)	Outcomes measurement	Code in HCP database
Language processing	Semantic language comprehension	13	Correct responses	Language_Task_Story_Acc
Picture vocabulary (NIH)	Language ability	4	Score based on test difficulty with 50% correct choices	PicVocab_Unadj
Oral reading recognition (NIH)	Language decoding and reading	3	Score based on accuracy	ReadEng_Unadj
Picture sequence memory (NIH)	Visual episodic memory	7	Score based on correct adjacent pictures	PicSeq_Unadj
Penn word memory	Verbal episodic memory	4	Correct responses	IWRD_TOT
List sorting (NIH)	Working within memory	7	Score based on correct responses	ListSort_Unadj
2-back	Working memory	9	2-back body condition correct responses	WM_Task_2bk_Body_Acc_Target
			2-back face condition correct responses	WM_Task_2bk_Face_Acc_Target
			2-back place condition correct responses	WM_Task_2bk_Place_Acc_Target
			2-back tool condition correct responses	WM_Task_2bk_Tool_Acc_Target
Short Penn continuous performance	Sustained attention	3	Correct positive responses	SCPT_TP
	Attention speed		Reaction time for positive responses	SCPT_TPRT
Flanker task (NIH)	Inhibitory control and attention	3	Score based on accuracy and reaction time	Flanker_Unadj
Dimensional change card sorting (NIH)	Cognitive flexibility	4	Score based on accuracy and reaction time	CardSort_Unadj
Penn progressive matrices	Fluid intelligence	5	Correct responses	PMAT24_A_CR
Pattern comparison processing speed (NIH)	Processing Speed	3	Score based on accuracy	ProcSpeed_Unadj
Delay discounting	Impulsivity	15	AUC for the $200 condition	DDisc_AUC_200
			AUC for the $40000 condition	DDisc_AUC_40K
Relational task	Visual Relational Processing	6	Correct responses relational condition	Relational_Task_Rel_Acc
Variable short Penn sine orientation	Visuospatial processing	5	Correct responses	VSPLOT_TC

HCP = Human Connectome Project S1200 data release, NIH= National Institutes of Health Toolbox

Statistical analysis

Figure 1 shows a flow-diagram of subject exclusion and data-analysis. All analyses were performed in R (www.r-project.org, version 3.5.3, 2018–07-02). Demographics are expressed as mean ± sd and range or percentage where appropriate. For all analyses, cases with missing values were excluded and the number of observations used in the analysis (N) is reported. First, we tested the association between each neurocognitive measure and BMI by performing 20 separate LMMs using package LME4 (v1.1–18-1) (16). For each model, a neurocognitive measure was entered as the dependent variable while BMI and covariates were entered as independent variables. This approach allowed us to evaluate how BMI associates with performance on each test outcome independently while covarying for: age, gender, education, income, depression score, sleep quality score, and meeting criteria for alcohol abuse or not. We accounted for the presence of siblings in the sample by entering family ID as random intercepts in the LMMs. All continuous variables were standardized (centered to a zero-mean and scaled to the unit variance). P-values were calculated using the Satterthwaite’s approximation for the degrees of freedom, provided in the package lmerTest (v3.0–1) (17) and false discovery rate (FDR) corrected. Outcomes were considered significant if P(FDR)≤ 0.05. Since the association between BMI and cognition is likely bidirectional we repeated our analysis with the neurocognitive test outcomes as independent variables and BMI as dependent variables. We also ran our models without covariates to gain insight into their contribution to the association between BMI and cognition.

Following the univariate analyses, we used variable selection for generalized linear mixed models by L1-Penalized Estimation using glmmLasso (v1.5.1) (18, 19). All measures previously used in the univariate analyses were entered in the model as prediction candidates. Family ID was entered as random intercept. As the analysis required complete data, all incomplete data was removed from the data set (14.7%, remaining N=876). We determined the optimal parameter lambda in an exhaustive 10-fold cross-validation using least mean squared prediction error (MSE) as the model performance measure. We ensured that families were within one fold. For the final model, we used lambda+1SE, giving the most regularized model such that the error is within one standard error of the minimum. To ensure a stable result, the lambda selection procedure was repeated 10 times. We will present the LASSO model refitted on all data using the median lambda+1SE across 10 analysis repetitions.

We used results from the LASSO to select the final tests for the neurocognitive test battery by picking measures with non-zero LASSO model weights. We visualized the hierarchy of measures by plotting variance explained against time spent on task completion to guide test selection for the battery.

Additionally, we approximated effect size by calculating and plotting the Pearson’s correlation coefficients (r) between the scaled (mean=0 and SD=1) BMI and selected cognitive variables and visualized the absolute differences in test score between individuals with healthy weight (BMI 18.5–24.9) and obesity (BMI >30).

Results

Sample demographics

Demographic information is summarized in Table 2. Participants were young adults (average age = 28.93; 56.38% female) who ranged in BMI from 18.58–55.78. Notably, the prevalence of obesity was 25.32%, which is lower than current estimates for the Minnesota Area (31.6%, (20)). Although other races and ethnicities were represented, a majority of the sample was white and non-Hispanic. As 14.7% of the data was omitted because of missing values for the LASSO analysis, we summarized demographics for this group separately. As shown, the demographics are comparable.

Table 2:

Demographics

	Univariate analyses			LASSO
Variable	Mean (sd)	Range	%	Mean (sd)	Range	%
Age (years)	28.93 (3.66)	22–36		28.85 (3.65)	22–36
BMI	27.18 (5.68)	18.58–55.78		26.55 (4.96)	18.58–47.76
Gender (% female)			56.38			55.14%
Education (years)	15.05 (1.74)	11–17		15.12 (1.7)	11–17
Household income
<$10,000			5.84			5.59
10K-19,999			7.11			6.28
20K-29,999			11.39			11.64
3,30K-39,999			11.30			11.87
40K-49,999			11.39			10.62
50K-74,999			22.01			22.15
75K-99,999			13.53			14.73
>=100,000			16.94			17.12
Sleep quality index	4.74 (2.73)	0–19		4.65 (2.69)	0–19
Depression score	53.89 (5.68)	50–81		53.69 (5.54)	50–81
Met criteria for alcohol abuse			14.31			13.81
Self-reported race
Am. Indian/Alaskan Nat.			0.10			0.11
Asian/Nat. Hawaiian/Other Pacific Is.			5.65			5.94
Black/ African Am.			13.15			12.10
White			77.31			78.54
More than one			2.04			1.83
Unknown/ Not Reported			1.75			1.48
Self-reported ethnicity
Hispanic			8.28			8.56
Non-Hispanic			90.75			90.30
Unknown/ Not Reported			0.97			1.14

Linear models N=1027, LASSO N=876. Values are presented as mean (sd) and range or percentage (%). Am. = American, Nat.= native, Is.= Islands

Association between BMI and neurocognitive test performance

The LMMs for BMI and neurocognitive test performance revealed a significant association between BMI and 7 cognitive measures including the Dimensional Change Card Sorting test, Oral Reading Recognition test, Penn Word Memory test, Variable Short Penn Line Orientation test, both parts of the Delay Discounting task ($200 and $40.000), and the Relational task. The estimate ß, approximate Degrees of freedom (Df), T-values and FDR-corrected P-values (P(FDR)) are reported in Table 3.

Table 3:

Association between neurocognitive test performance and BMI

Measure	ß	N	N(fam)	Approximated df	T-Value	P(FDR)
Language Processing	−0.073	900	402	865.83	−1.82	0.138
Picture Vocabulary	−0.059	1016	423	969.15	−2.15	0.079
Oral Reading Recognition	−0.090	1016	423	1000.54	−3.17	0.008
Picture Sequence Memory	−0.045	1016	423	1001.67	−1.40	0.216
Penn Word Memory	−0.140	1015	423	985.64	−4.33	<0.001
List Sorting	0.007	1016	423	1005.16	0.23	0.861
2-back body	−0.043	918	408	862.33	−1.11	0.335
2-back face	−0.066	918	408	882.96	−1.72	0.143
2-back place	−0.038	918	408	858.96	−0.99	0.381
2-back tool	−0.007	918	408	900.90	−0.17	0.861
PCP Accuracy	−0.047	1015	423	1005.00	−1.44	0.216
PCP Reaction Time	0.046	1011	423	1000.99	1.40	0.216
Flanker	−0.007	1016	423	983.52	−0.22	0.861
DCCS	−0.082	1013	423	988.34	−2.52	0.034
Penn Progressive Matrices	−0.061	1014	423	998.82	−1.99	0.104
PCPS	−0.057	1016	423	979.34	−1.73	0.143
Delay Discounting AUC $200	−0.084	1015	423	977.43	−2.62	0.034
Delay Discounting AUC $40,000	−0.116	1015	423	992.35	−3.63	0.002
Relational Task	−0.098	892	401	866.79	−2.56	0.034
Penn Line Orientation	−0.129	1014	423	991.12	−4.13	<0.001

Tests that were significantly associated with BMI are highlighted in gray. The results are based on LMMs with (1) standardized BMI as independent variable, (2) standardized neurocognitive test outcome as dependent variable, (3) gender, age, education, sleep quality score, depression score, household income, and alcohol abuse status as covariates, and (4) family ID as random variable. ß=estimated ß, N=number of observations, N(fam)=number of unique families, Approximated Df=Satterthwaite’s approximation for the degrees of freedom, P(FDR)= False Discovery Rate-corrected P-value, PCP= Penn Continuous Performance, DCCS= Dimensional Change Card Sorting, PCPS=Pattern Comparison Processing Speed

As the effect of SES and nuisance variables (i.e., sleep score, depression, alcohol abuse) could have a different effect on BMI and cognition, we performed LMMs with neurocognitive measures as independent variables and BMI as the dependent variable. This provides insight into the bidirectional nature of the BMI-cognitive dysfunction relationship. The results are displayed in the Supporting Information, Table S1. Although the associations found in this analysis were largely similar to the LMMs above, only the associations of the Oral Reading Recognition test, Penn Word Memory test, Delay Discounting for $40.000, the Relational Match Control, and Variable Short Penn Line Orientation test with BMI reached significance after FDR correction. Re-running the analyses without covariates produced stronger associations (Table S2), suggesting that these covariates contribute the relationship between BMI and neurocognitive function.

Lasso regression analysis and variable selection

After ten 10-fold cross-validation repetitions, we found a median lambda+1 SE of e^4.425 (range: 4.25–4.55). The results are shown in Supporting Information Figure S1. Table 4 shows non-zero LASSO beta weights of the variables in the final model. Predictors are ordered based on their weight, which is a proxy of variable importance in the model. Notably, of the eight covariates included in the model only education level survived in the final model. This contrasts with the neurocognitive measures, where six of the seven variables that were identified in the LMMs, survived the strict LASSO regression analysis. Only Dimensional Change Card Sorting was excluded. A test battery including the 6 neurocognitive tests would require 38 minutes to administer (see Table 1 for estimate test times).

Table 4:

LASSO Results

Predictor	LASSO ß Weight
Social Economic Status
Education (years)	0.055
Cognitive Measure
Delay Discounting AUC $40,000	0.070
Relational Task	0.064
Penn Progressive Matrices	0.054
Oral Reading Recognition	0.043
Penn Line Orientation	0.034
Penn Word Memory	0.025
Delay Discounting AUC $200	0.021

Non-zero beta weights of the LASSO analysis ordered by size.

We also plotted the explained variance (R²) of the model variables against test duration ordered by variable weight to guide further test reduction if time constraints do not permit administration of the whole battery. Figure 2 depicts both the variance explained by each individual variable (2A) and the variance explained by regression models if the variables were entered in stepwise order (2B). As out-of-sample generalization was not tested, R² may be lower in new samples.

Figure 2:

Variance explained (R²) by education level (education in years) and selected neurocognitive measures plotted against cumulative test time. The neurocognitive measures are ordered by their model weights in the LASSO regression analysis. A) The explained variance for each individual measure. B) The explained variance for the combined models of tests.

Figure 3A depicts the correlations (r) among the variables selected by the LASSO analysis and BMI. Correlations between selected variables and BMI were between −0.13 and −0.17. Figure 3B shows differences among individuals with healthy weight (HW, BMI 18.5–24.9), overweight (OW, BMI 25–29.9) and obesity (OB, BMI > 30) for each variable selected by the LASSO analysis. These graphs show a difference of approximately 0.5 SD in test scores between HW and OB individuals.

Figure 3:

A) Pearson’s correlations (r) between the selected neurocognitive variables and of the selected neurocognitive variables with BMI. All variables were scaled to zero mean and unit variance (SD=1). B) Neurocognitive test results, represented as scaled units in standard deviations (SD) centered around the mean (mean = 0, SD = 1). Bars represent mean (±SEM). PMAT= Penn Progressive Matrices, Relational= Relational Task, ORRT = Oral Reading Recognition Test, VSPLOT= Variable Short Penn Line Orientation Test, PWMT= Penn Word Memory, DD 200= Delay Discounting for $200, DD 40k= Delay Discounting for $40.000

Discussion

Based on emerging evidence of an association between obesity and neurocognitive dysfunction we adopted a data-driven approach to develop a core neurocognitive test battery to promote the accumulation of data. We identified seven neurocognitive measures that were significantly associated with BMI in healthy young adults after correcting for potential confounding variables. We then selected the variables that best predicted BMI to include in a concise test battery of neurocognitive function associated with BMI.

Recommendations for the neurocognitive test battery

Using LASSO regression we found that the optimal model to predict BMI included a combination of years of education and six neurocognitive measures. In order of predictive capability and explained variance these were: Delay Discounting, the Relational Task, the Penn Progressive Matrices test, the Oral Reading Recognition test, the Variable Short Penn Line Orientation test, and the Penn Word Memory test. Administration of all six tests would require approximately 40-minutes.

While the LASSO regression provided the model best predicting variance in BMI, time constraints may require further test reduction. With this in mind, figure 2 provides a guide for further test selection by weighing test time against the variance explained by each test, as well as the extra variance explained by adding the test to the model. For example, given a 20minute time limit, the optimal test selection would include Delay Discounting for large monetary amounts, the Relational Task and the Penn Progressive Matrices Test, which in a combined model accounts for the largest portion of the variance.

Neurocognitive Impairment in Obesity

Although the primary aim of this study was to develop a core battery of neurocognitive function associated with BMI, our study also yielded several findings that provide further insights in to the link between BMI and cognition in young healthy adults. Our analysis adds to the evidence of a negative association between BMI and language ability, delay discounting, episodic memory, and cognitive flexibility (1, 2, 12). However, and surprisingly, we did not find a significant association between BMI and the executive function measures inhibition and working memory. These executive function domains have often been linked to risk of obesity and a prior meta-analysis confirmed a significant association with BMI (12, 21). Systematic reviews have however shown inconsistencies with one review identifying a consistent association between BMI and inhibition and working memory, while another indicated inconsistent evidence for inhibition and no evidence for working memory. Possibly, these discrepancies reflect an interaction of obesity with age and/ or effects of related co-morbidities. While some studies, like ours, had mainly young adult participants other studies have studied larger age ranges including youth and/ or middle aged and elderly participants (2, 21, 22). Similarly, studies have been more or less restrictive in their exclusion criteria. Notably, younger and older populations may be particularly at risk for obesity-associated cognitive dysfunction: the association between obesity and risk of dementia is well established (3, 4) and systematic review showed more consistent evidence for the association between obesity and executive function decline in youth (23). Moreover, and meta-analyses have consistently shown an association between T2D and executive dysfunction (24, 25). Another possible explanation is the inclusion of covariates in our analysis; however, even with these variables removed we failed to observe an association of BMI with inhibition and working memory.

Another notable finding was the negative association between BMI and tasks relying on visual and visuospatial processing. Visual and visuospatial processing are typically not evaluated in studies examining obesity and cognition. However, previous work has reported associations between BMI in both the copy and memory components of the Rey Osterrieth complex figure task (26, 27), as well as other tasks depending on visual spatial processing (28–30). In these studies, deficits are generally interpreted as executive dysfunction or memory impairment. Here we found that the association between BMI and performance on the Penn Line Orientation task was among the strongest associations observed. This finding is consistent with another recently published analysis of the HCP Young adult S900 data release, which observed that the negative correlation between BMI and cognitive task performance was strongest for nonexecutive tasks of visuospatial ability and verbal memory (5). The Penn Line Orientation task was designed to measure the complex reasoning domain of spatial ability (31). It requires subjects to match the orientation of a line to that of a sample and is not known to require prefrontal, hippocampal or mesolimbic circuits. Concurrently, BMI was associated with the Relational Task, designed to measure the ability to evaluate internally generated information, in this case a decision made about object features (32). Participants are presented with two pairs of objects varying in shape and fill texture and must determine if the top and bottom pair of objects differ across the same dimension. While both of these tasks have a complex reasoning component that could drive the association with BMI, an alternative and less explored hypothesis is the presence of an association between BMI and higher-order visual processing. Further support for this possibility comes from neuroimaging studies reporting associations between BMI and response to food cues in visual regions, particularly the lateral occipital cortex and fusiform cortex (33, 34), as well as increased connectivity of visual cortical areas with other brain regions including the ventral striatum/ caudate (35, 36). Further investigations of BMI and visual function therefore seem warranted.

Importantly, the recommendation of a test battery marks an initial step towards a more comprehensive understanding of the association between BMI and neurocognitive function. The use of this test battery in obesity trials remains to be validated. In addition, during neuropsychological exams, deficits are determined with reference to large normative datasets. Although current datasets do consistently include age and gender, they do not take diabetes and obesity into account when defining norms. It is therefore not possible to use a reference dataset to determine whether the deficits we observe indicate mild or moderate impairments or if they fall within a normal range. However, based on the current sample of healthy young adults we showed decreases in neurocognitive test scores of around 0.5 SD in persons with obesity and effect sizes (r) ranging from −0.12 to −0.17. While these effect sizes are usually considered small to medium, Funder and Ozer (37) argue that effect sizes indicating a small to medium effect at the level of a single event can become consequential over longer durations. Nevertheless, future work is needed to gain a clearer understanding of the magnitude of the changes with respect to the normal range of performance within a healthy weight sample. In addition, given the large body of literature showing that mid-life obesity is predictive of late-life dementia (3, 4), longitudinal studies would be of interest to determine if the associations between BMI and cognitive function identified in our young adults predict later cognitive changes and dementia onset.

Caveats

Several caveats are important to consider in interpreting the findings from our analyses. While the HCP neurocognitive test battery is extensive, some cognitive functions previously associated with BMI are not addressed in this test battery. For example, performance on the probabilistic learning task, a measurement of positive and negative outcome learning, was reported to be impaired in people with obesity (38, 39). Furthermore, tasks measuring response inhibition, such as the Stop Signal Task, Go/No Go task, and Stroop task, have been inversely associated with body weight status and eating behaviors associated with overeating (1), but were not included in the HCP battery.

Also, we used linear regression for all our analyses. While this allows for straight-forward interpretability of test results, it may not best fit potential effects of body weight status on all cognitive domains. For example, dopamine signaling, one of the major pathways considered to mediate effects of body composition on behavior (40), has been shown to act on cognitive functions such as working memory and cognitive control in an inverted-U-shape fashion (i.e. there is an optimum level of dopamine for cognitive functions and both lower and higher levels of dopamine are associated with lower performance) (41).

Another important caveat is that the tests in this battery are designed to understand the association between adiposity and performance across the available assessments of different cognitive domains. They were not selected to define pathophysiology or to predict treatment outcome. Moreover, the predicted BMI-cognition relationship does not allow us to infer causality. We mainly focused on the effects of BMI on neurocognitive performance. This is reflected in our discussion of the results and the choice neurocognitive tests that might not necessarily give a full reflection of the neurobehavioral correlates of BMI (see systematic reviews 1, 4, 35). However, the link between obesity and cognitive dysfunction is most likely bidirectional and further complicated by influences of diet and related metabolic impairment (6). For example, memory function and cognitive control mechanisms such as delay discounting have been related to eating behavior and weight gain (1, 43, 44). It was hypothesized that these bidirectional interactions could cause a ‘vicious cycle’ of weight gain and cognitive dysfunction (45). To gain some insight in the bidirectional nature of the BMI-cognitive dysfunction relationship we performed LMMs with neurocognitive measures as independent variables and BMI as dependent variable. This yielded a similar pattern test performance-BMI association. However, it is likely that our test selection does not fully reflect all the behaviors contributing to weight gain.

Finally, we purposefully chose to perform our analyses in a sample of young healthy adults to avoid potential effects of co-morbidities and age-related cognitive decline in our analysis. We believe this is important for disentangling the effects of body composition, metabolic co-morbidities and ageing on cognitive impairment. However, a limitation of this approach is that it may decrease generalizability to a more diverse population where co-morbidities such as obesity, T2D (or pre-diabetes) and hypertension are common. Further studies should be performed to investigate whether this test battery is appropriate for use in diverse and unique populations. Similarly, future studies should be aimed at providing in the clinical significance of the association between BMI and cognitive performance.

Conclusion

Based on analysis of a large, well-controlled set of neurocognitive test data we propose a ~ 40-minute test battery best predicting the relation between BMI and cognitive function in young adults including: Delay Discounting, the Relational Task, the Penn Progressive Matrices test, Oral Reading Recognition, the Variable Short Penn Line Orientation test and the Penn Word Memory test and the covariate years of education.

Future studies should seek to replicate the study’s findings and validate the proposed test battery across unique samples and evaluate generalizability of the results in younger and older populations as well as populations with common co-morbidities. Incorporation of this short battery into large-scale trials of obesity will help to shed light on the interrelationship between obesity and cognition: the influence of obesity on cognitive performance and reversibility thereof if weight loss is achieved and vice versa the contribution of cognition to obesity risk and to weight-loss intervention outcomes.

Study important questions.

What is already known about this subject?

• Obesity has been repeatedly associated with cognitive impairment but reports on the specific cognitive functions that are affected are inconsistent.

• To further understanding of the link between obesity and cognition it is imperative to include cognitive measures in large scale obesity trials.

What does your study add?

• Based on analysis of a data from a large well-controlled cohort, we propose a test battery for studies interested in measuring cognitive impairment in obesity.

• We visualize explained variance for each component in the test battery guiding further test reduction if required.