Diabetes-specific dementia risk score (DSDRS) predicts cognitive performance in patients with type 2 diabetes at high cardio-renal risk.

AIM
To investigate the relationship between the diabetes-specific dementia risk score (DSDRS) and concurrent and future cognitive impairment (CI) in type 2 diabetes (T2D).


METHODS
DSDRS were calculated for participants with T2D aged ≥60 years from the CARMELINA-cognition substudy (ClinicalTrials.gov Identifier: NCT01243424). Cognitive assessment included Mini-Mental State Examination (MMSE) and a composite attention and executive functioning score (A&E). The relation between baseline DSDRS and probability of CI (MMSE < 24) and variation in cognitive performance was assessed at baseline (n = 2241) and after 2.5 years follow-up in patients without baseline CI (n = 1312).


RESULTS
Higher DSDRS was associated with a higher probability of CI at baseline (OR = 1.17 per point, 95% CI 1.12-1.22) and follow-up (OR = 1.24 per point, 95% CI 1.14-1.35). Moreover, in patients without baseline CI, higher DSDRS was also associated with lower baseline cognitive performance (MMSE: F(1, 1930) = 47.07, p < .0001, R2 = 0.02); A&E z-score: (F(1, 1871) = 33.44 p < .0001, R2 = 0.02) and faster cognitive decline at follow-up (MMSE: F(3, 1279) = 38.41, p < .0001; A&E z-score: F(3, 1206) = 148.48, p < .0001).


CONCLUSIONS
The DSDRS identifies patients with T2D at risk of concurrent as well as future CI. The DSDRS may thus be a supportive tool in screening strategies for cognitive dysfunction in patients with T2D.


Introduction
People with type 2 diabetes (T2D) are twice as likely to develop dementia compared to people without diabetes. 1 This is of concern, since cognitive impairment, also already in pre-dementia stages, can interfere with diabetes self-management and is associated with an increased risk of severe hypoglycemic events. 2,3 For this reason, recent diabetes management guidelines recommend clinicians to screen for cognitive impairment in patients with T2D. [2][3][4][5][6] In 2013, the diabetes-specific dementia risk score (DSDRS) was introduced to help researchers and clinicians identify T2D individuals at risk of developing dementia. 7 The DSDRS predicts the 10-year dementia risk in patients with T2D and incorporates several readily available dementia-risk factors, such as diabetes-related complications, level of education, depression and cerebro-and cardiovascular disease. The DSDRS was developed based on a population-based registry, without availability of formal cognitive testing in all individuals. Hence, it is not clear yet if the DSDRS can also identify individuals with T2D with concurrent cognitive dysfunction cross-sectionally. Moreover, it is unknown if the DSDRS is able to predict future cognitive decline, even when it is less severe than frank dementia. Therefore, we studied the relationship between DSDRS and concurrent cognitive performance at the moment of DSDRS assessment as well as change in cognition over 2.5 years in a large prospective cohort of people with T2D at high cardio-renal risk.

Population
part of a multicenter, international, randomized, double blind study in patients with type 2 diabetes at high cardio-renal risk (CARMELINA®: ClinicalTrials.gov Identifier: NCT01897532) that investigated if treatment with linagliptin vs placebo resulted in a lower incidence of accelerated cognitive decline.
CARMELINA included adults with type 2 diabetes, HbA1c 6.5-10.0%, at high cardiovascular risk (history of vascular disease and urine albumin-to-creatinine ratio (UACR) of 30 mg/g (or equivalent)) or high renal risk (estimated glomerular filtration rate (eGFR) of 45-75 mL/min/1.73 m2 and UACR of 200 mg/g (or equivalent) or eGFR of 15-45 mL/min/1.73 m2 regardless of UACR). Participants with end-stage kidney disease, defined as eGFR of b15 mL/min/1.73 m2 or requiring maintenance dialysis, were excluded (more details 9 ). CARMELINA-COG only included participants from countries using the Latin alphabet with documented years of education and a valid baseline cognitive assessment. A cognitive assessment was considered invalid when documented test scores were considered implausible (i.e. unrealistic values). The CARMELINA-COG study found neutral results for the effect of linagliptin versus placebo on accelerated cognitive decline (more details 8 ). Therefore we made no distinction between both treatment arms in the present study. The present study is restricted to participants with a minimum age of 60 at baseline, since the DSDRS model is only validated in a population of 60 years and older. A valid follow-up cognitive assessment was required for the longitudinal analyses (see below).

Cognitive performance
Cognitive performance was assessed using three easy-to-administer neuropsychological tests: • The Mini-Mental State Examination (MMSE), a widely known screening test, is used to assess global cognitive performance. 10 The MMSE has a maximum score of 30 and evaluates different cognitive functions including orientation in time and place, verbal registration, short term verbal memory, attention, language and visuoconstruction. A MMSE score below 24 indicates cognitive impairment (CI). 11,12 Participating centers used country-specific validated versions. • The Trail Making Test (TMT) is a timed test, that assesses psychomotor speed, scanning, divided attention and mental flexibility. 13 Its timing aspect makes it sensitive for subtle changes in cognitive performance that are commonly seen in type 2 diabetes. 14 The TMT consists of two parts. In part A, participants are required to connect numbered circles in consecutive order as fast as possible (1 -2 -3 etc.). It measures psychomotor speed, scanning abilities and number sequencing. For part B, participants alternate between numbered and lettered circles, also in consecutive order (1 -A -2 -B etc.). Part B measures divided attention, working memory and task shifting. 13,15 It is more time consuming and error-prone than part A. The TMT ratio score ((TMT-B -TMT-A)/TMT-A) reflects executive functioning and reflects the additional time needed to complete part B, corrected for the time needed to complete part A. • The Verbal Fluency Test (VFT) is a timed test and measures someone's fluency of speech, which is dependent on vocabulary size, lexical access speed, strategy finding, updating and inhibition ability. 16 Participants are instructed to verbalize as many words from a certain category (i.e. animals) within 60 s. Participants were also asked to list words starting with the same letter (i.e. F -A -S). Word generation according to an initial letter gives the greatest scope for seeking strategies guiding the search for words. Category-driven search provides more structure in search strategy. 13 Both VFT measures are combined into one overall zscore. Since language-specific differences in word frequencies are known, all fluency scores were adjusted for each individual's native language, as described elsewhere. 8 A composite score combining both the z-scores on the Trail Making Test (TMT) and the Verbal Fluency Test (VFT) is used to assess attention and executive functioning all together in one robust score (A&E score), sensitive for capturing the subtle changes that are seen in type 2 diabetes (for more details about the derivation 8 ). A cognitive assessment at baseline is considered valid when it includes at least an available MMSE score. At follow-up at least a score on one of the cognitive tests (MMSE, TMT or/and VFT) should be available.

Diabetes-specific dementia risk scores
Individual dementia risk scores were calculated with help of the diabetes-specific dementia risk model (DSDRS). 7 This prognostic model was developed for calculating individual 10-year dementia risk in patients with T2D of 60 years and older, based on eight predictors that were most strongly predictive of clinical diagnosis of dementia in T2D; age, years of education, acute metabolic event, microvascular disease, clinical diagnosis of diabetic foot, depression, cerebro-and cardiovascular disease (Appendix Table A.1). Individual sum scores on the DSDRS, ranging from −1 (low risk) to 19 (high risk), were calculated by simply adding up each relative contribution of the predictors as defined in the original model (Appendix Fig. A.1).
For the main group analysis in the current study, we used a modified version of the model, since information about history of cardiovascular and cerebrovascular disease was only available in the CARMELINA dataset for participants with albuminuria. Hence, a maximum DSDRS of 16 rather than the original 19 could be obtained. All analyses were repeated separately in the albuminuria subgroup with available history of cardio-and cerebrovascular disease with the full 19-point model. For the predictor microvascular disease, the original DSDRS model used the definition of 'diabetic retinal disease and/or end-stage renal disease'. We used a definition of 'diabetic retinopathy and/or severe nephropathy with an eGFR b 30' instead, since the CARMELINA trial did not include patients with end-stage renal disease. For the predictor 'level of education', the original DSDRS model used the definition high school or less/ college or more. We used years of formal education as an indicator of educational attainment, since multiple countries with different educational systems are included in CARMELINA. For the prediction model this was dichotomized in years of formal education at or below the median/above the median of the study population (Appendix Table A.1).
For both the main group and subgroup analyses, sum scores on the DSDRS above 10 were taken together in one category due to small sample sizes in the high risk groups. Because the treatment effect in the CARMELINA trial on cognition was neutral, treatment allocation was not considered in the analyses. 8

Baseline
Logistic regression analysis was used to calculate the probability of CI (MMSE b 24) according to sum risk scores on the DSDRS. Next, for participants without CI (MMSE ≥ 24), the relationship between sum risk scores on the DSDRS and cognitive performance (MMSE and A&E z-score) was assessed using linear regression analysis. Demographic variables were not included as co-variates in the model, since these are already included in the DSDRS itself (i.e. age, years of education). We performed sensitivity analyses stratified by age bands in years (i.e. 60-64, 65-69, 70-74, 75-79, 80-84, 85+) to look at age independent effects.

Follow-up
In individuals that had no CI (MMSE ≥ 24) at baseline, we used logistic regression analysis for calculating the probability of developing CI (MMSE b 24) at follow-up according to the sum scores on the DSDRS. Due to relatively small numbers of incident CI, no post-hoc agestratified analyses were performed. Linear regression analysis was used to investigate if sum scores on the DSDRS predicted change from baseline in cognitive performance (MMSE and A&E z-score). Baseline cognitive performance and time from baseline till follow-up visit were used as covariates.

Subgroup analysis
The analysis steps above were repeated on a sub selection of the population with confirmed micro-or macro albuminuria (i.e. UACR ≥30 mg/g creatinine or ≥30 mg/L or ≥30 μg/min or ≥30 mg/24 h in two out of three unrelated spot urine or timed samples in the last 24 months prior to randomization) in whom data on history of previous cardioand/or cerebrovascular disease was available, allowing us to use the complete 19-point DSDRS model (Appendix Table A.1). No age-related stratifications were performed on this sub-set due to small sample sizes.
All statistical analyses were performed with SAS software, version 9.4 (SAS Institute, Cary, NC, USA).
For a supportive overview of all analyses, outcomes and populations, please see Table 2.

Baseline and follow-up analysis
Of the 2694 participants included in the CARMELINA-COG study, years of education and cognitive assessment were available in 2666, of whom 2253 were aged ≥60 years and therefore eligible for the baseline analysis in the current study. MMSE was available in 2241 and constituted baseline analysis. Of this group 37.3% was female. The mean age was 70.6 ± 6.5 and mean years of formal education 11.4 ± 4.0. The population was largely Caucasian (91%). The mean duration of diabetes was 16.2 ± 9.6 years ( Table 1).
At baseline, 309 (13.8%) had CI (MMSE b 24). The DSDRS was related with baseline CI risk ( Fig. 1a; OR for CI 1.17 per DSDRS point [95% CI 1.12-1.22]; p b .0001, R 2 = 0.02). The point estimate for prediction of CI by the DSDRS was similar in age-band stratified sensitivity analyses (Appendix Table A.2 and Fig. A.3), albeit with wider confidence intervals due to smaller sample sizes in subgroups.
Of those without CI at baseline (MMSE ≥ 24) (n = 1932), cognitive follow-up was obtained in 1312 (68%) after a median followup duration of 2.5 ± 0.8 years (Appendix Fig. A.2). A number of 620 (32%) participants dropped out before follow-up assessment because their last cognitive assessment was N7 days after end of treatment, there were missing or implausible values on the cognitive tests or participants died or discontinued trial medication (for more information 8,9 ). Of those that did have a follow-up (n = 1312), 1283 had an available MMSE and 1228 an A&E z-score. Compared to those that did have a follow-up, those that dropped out were slightly older (70.9 ± 6.7 vs 70.1 ± 6.2), their duration of diabetes was longer (17.0 ± 10.0 vs 15.8 ± 9.3) and 10-year dementia risk was higher (28.0 ± 16.4 vs 25.0 ± 15.1).  At follow-up, CI occurred in 88 participants (6.7%). The DSDRS significantly predicted incident CI (OR 1.24 per DSDRS point [95% CI 1.14-1.35]; p b .0001, R 2 = 0.02, Fig. 1b).

Subgroup analysis
A subset of 1035 participants with albuminuria (46% of total group), had available data on history of cardio-or cerebrovascular disease. .0001, R 2 = 0.02). X-axis: sum risk scores for DSDRS at baseline, ranging from −1 to 11. Y-axis: probability of CI (MMSE b24), ranging from 0 to 1 including 95% confidence interval. Results obtained using logistic regression analysis. DSDRS of 11 and higher are taken together due to small sample sizes. For overview of numbers per DSDRS sum risk score, see Appendix Table A .0001, R 2 = 0.24), after correction for baseline performance and follow-up duration (Appendix Fig. A.7).

Discussion
This study shows that higher scores on the DSDRS are also associated with concurrent CI and worse cognitive performance in a group of patients with T2D at high cardio-vascular renal risk, irrespective of age. Moreover, higher DSDRS predicted CI 2.5 years later, as well as more subtle cognitive decline over time.
Prognostic dementia models areby definitiondeveloped to predict future dementia. The question is if these models are also able to cross-sectionally identify people with a high probability of having CI, which could, for example, be supportive for screening. To our knowledge, no studies have tested this before. In our study population 13.8% of the participants has CI at baseline, defined as a MMSE b24, which is relatively high compared to previous studies, also considering the age of the populations involved. 17,18 This may reflect the fact that the CARMELINA population is at high cardiovascular risk and therefore also at higher risk of CI. 7 The DSDRS clearly separated people according to baseline CI risk: for example of those with a score of ≥8, over 20% has CI, compared to less than 10% CI in those with a score ≤ 3 (Fig. 1a,  Table A.3). For those that do not have CI at baseline, higher DSDRS scores are also associated with worse cognitive performance on the MMSE and A&E z-score. However, effect sizes are small, and although the association was statistically significant, the variance explained by the DSDRS was only 2%. Another question is if prediction models for dementia are also able to predict more subtle cognitive decline. We identified no previous studies either in people with diabetes or in the general population that explored this. Our study shows that for participants without CI at baseline, 6.7% developed CI after 2.5 years. Of those with a score of ≥8, around 14% developed CI, compared to % CI in those with a score ≤ 3 (Fig. 1b, Table A.3). In those without CI at baseline, higher DSDRS are significantly associated with a greater cognitive decline over a period of 2.5 years, with small to moderate effect sizes. Our results show that the DSDRS predicts a wide range of cognitive decline, from accelerated cognitive decline, to cognitive impairment, toas shown in former researchfrank dementia. 7 Several diabetes management guidelines recommend screening for cognitive problems in patients with T2D, but there is still uncertainty how this should be implemented. [2][3][4][5][6] Our findings on the cross-sectional analyses show that the DSDRS could support such screening strategies. The strength of the DSDRS, or comparable risk scores that primarily rely on demographic and clinical data mostly already available in the patients records 19 , 20 is that it is very easy to implement in daily practice (e.g. as part of the electronic medical record system). Because of its low-cost and time-efficient characteristics, the DSDRS has an advantage over other dementia prediction models that also require additional biomarkers, such as MRI or other advanced laboratory variables, 21 making the DSDRS a suitable tool for primary care. An implementation study would be needed to evaluate the feasibility and practical applicability of this approach.
A few limitations of our study should be considered. The CARMELINA trial cohort consisted of a selected T2D group at high cardio-vascular risk. 8 Compared to the DSDRS distribution previously observed in a population based sample of patients with T2D, 7 fewer participants had low risk scores, likely reflecting the high cardiovascular burden in our cohort. Moreover, there were also fewer participants in the highest risk scores range, probably reflecting that the oldest old are less likely to participate in a drug trial. Importantly, despite this different risk distribution, the DSDRS remained nonetheless predictive. Possibly it may have even better discriminative ability in a less selected cohort. The optimal threshold for differentiating those with and without CI based on the DSDRS should also preferably be determined in a population-based setting. Another point to consider is that treatment could potentially play a role in our results, particularly because the data were derived from a randomized controlled trial. Yet, when the DSDRS was developed, diabetes treatment was considered as a dementia predictor, but not retained in the final model. Moreover, CARMELINA found neutral results for the effect of linagliptin versus placebo on the cognitive outcome. 8 Another limitation is that data on cardio-and/or cerebrovascular disease was not available for all subjects; it was only registered in those with albuminuria. However, subgroup analyses showed similar results compared to the total group, suggesting that the DSDRS is still predictive when predictors are missing, which would be a convenient feature when it comes to clinical implementation. Further, a limited test battery was used to measure cognitive performance. The inclusion of additional tests to cover other cognitive domains would have been informative when drawing up extensive cognitive profiles, but in the current research it would in essence not have changed the results. Nevertheless, the cognitive tests that were applied prove to be sufficient to answer our question.
Strengths of our study include the relatively large number of patients with T2D. Our results show that the relationship of the DSDRS with cognition is not solely driven by age. We used two complimentary cognitive tests; we included the more conservative, but widely-used and easily interpretable MMSE, and in addition we used a more sensitive cognitive composite score that covers relevant cognitive domains in T2D. 14

Conclusion
The DSDRS effectively identifies patients with T2D at risk of concurrent and future cognitive impairment, also in those without dementia. In addition to informing clinicians on future dementia risk, the DSDRS can thus, in an individualized, time-and cost efficient way, advice clinicians on which T2D patients to screen or monitor for cognitive problems.

Declaration of competing interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: G.J.B.'s institution receives study grants from Boehringer Ingelheim. O.E.J. is an employee of Boehringer Ingelheim.
No other potential conflicts of interest relevant to this article were reported. Years of formal education at or below median/above median High school or less/college or more.

Acute metabolic event
Hyper-or hypoglycemia that required hospitalization in the 2 years prior to baseline assessment.
Hyper-and/or hypoglycemia event severe enough to be hospitalized based on medical history in the 2 years prior to baseline. Microvascular disease eGFR (MDRD) [mL/min/1.73 m2] b30 at baseline. End-stage renal disease (including dialysis and kidney transplantation) in the two years prior to baseline. And/or prior clinical diagnosis of diabetic retinopathy requiring retinal laser coagulation therapy or intravitreal injection(s) of an antivascular endothelial growth factor therapy.
And/or diabetic retinal disease in the 2 years prior to baseline.

Diabetic foot
Clinical diagnosis of diabetic foot defined as gangrene, amputation or lower limb ulcer that required hospitalization.
-Gangrene or lower limb ulcer that required hospitalization in the two years prior to baseline.   7 In the current study predictors on cerebro-and cardiovascular disease are not available for the complete population, but only for a subgroup (see Methods). As a result the maximum points for predicted 10-year risk of dementia that can be assigned to each person in the complete population is 16. Depression Clinical diagnosis of depression in the two years prior to baseline assessment. History of depression based on medical history in the 2 years prior to baseline.
Definitions in 7 are according to ICD-9 CM codes. a Documented by at least one lesion estimated to be ≥50% narrowing of the luminal diameter with imaging techniques or prior percutaneous or surgical carotid revascularization. b 50% narrowing of the luminal diameter of one major coronary artery by coronary angiography, MRI angiography in patients not revascularized and at least: a positive non-invasive stress test or patient discharged from hospital with a documented diagnosis of unstable angina pectoris between 2 and 12 months prior to screening visit. c Documented by previous limb angioplasty by stenting or by-pass surgery, previous limb or foot amputation due to macrocirculatory insufficiency, angiographic evidence of peripheral artery stenosis 50% narrowing of the luminal diameter in at least one limb (definition of peripheral artery: common iliac artery, internal iliac artery, external iliac artery, femoral artery, popliteal artery). d History on previous cerebro-and cardiovascular disease is only available in a subgroup of the current study and investigated with subgroup analyses.

Fig. A.2.
Flowchart. Note: Reasons for drop-out were because the last cognitive assessment was N7 days after end of treatment, there were only missing or implausible values on the cognitive tests, participants died or discontinued trial medication due to adverse events, non-compliance to protocol, refusal to continue taking medication, other or missing (for more information 8,9 ). MMSE: Mini-Mental State examination.