The present study is a prospective cohort study framed in the PREDIMED-Plus-Cognition sub-study, using a subset of participants (N = 487) of the PREDIMED-Plus trial. Full details of the study design and procedures of the PREDIMED-Plus trial have been published elsewhere (5). Further details on the study inclusion/exclusion criteria as well as the study protocol are available at http://predimedplus.com/. Briefly, the PREDIMED-Plus study is an ongoing multi-center randomized parallel-group primary prevention trial (N = 6,874) designed to assess and compare the long-term effectiveness of an intensive lifestyle intervention with an energy-reduced Mediterranean diet (er-MedDiet, intervention group), physical activity (PA) promotion and behavioral support of weight loss goals, with a more common intervention featuring energy-unrestricted traditional MedDiet recommendations without any recommendations on PA and weight loss strategies (control group). In order to promote the adherence to the MedDiet both groups were free provided with an allotment of extra-virgin olive oil (1 L/month) and occasionally, tree nuts (125 g/month). Participant's recruitment took place between October 2013 and December 2016 across 23 Spanish hospitals, universities and research institutes. Participants were randomly assigned, in a 1:1 ratio, to intervention and control groups. The eligibility criteria for participants were community-dwelling adults with overweight grade II (18) or obesity [body mass index (BMI) between 27 and 40 kg/m²] from Primary Care Health Centers of the Spanish National Health System aged between 55 and 75 years in the case of men and between 60 and 75 years in women who met at least three criteria for metabolic syndrome. The clinical trial is registered at the International Standard Randomized Controlled Trials database (ISRCTN; 89898870).

Four study sites participated in the PREDIMED-Plus-Cognition sub-study, with an in-depth assessment of the cognitive performance at baseline, 1 and 3 years after the initiation of the assigned PREDIMED-Plus intervention: Hospital del Mar Medical Research Institute (IMIM), Barcelona; Pere Virgili Institute for Health Research (IISPV), Reus; University of Valencia (UV), Valencia; Bellvitge University Hospital (HUB), Hospitalet de Llobregat. Exclusion criteria for the present sub-study are included in Supplementary Table 1. The data were analyzed using the PREDIMED-Plus-Cognition database dated 14th January 2021. All participants gave written informed consent. The protocol of the PREDIMED-Plus-Cognition sub-study was approved by the local Research Ethics Committees from the participating centers and adheres to the standards of the World Medical Association (WAMA) Declaration of Helsinki.

Baseline characteristics of study participants stratified by diabetes status are included in Table 1. Briefly, 50.5% were women, the mean (SD) age was 65.2 (4.7) years, 18.7% were employed and 62.1% were retired. Regarding their lifestyle, 12% were current smokers, most were underactive (66.9%) or sedentary (15.6%) and had a low or medium adherence to the er-MedDiet (45.4 and 41.5%, respectively). All participants were over-weight (27.3%) or obese (48.5% had obesity type I and 24.2% type II). Finally, 50.3% were taking medications for high cholesterol and 23.0% used tranquilizers or sedatives.

Compared to individuals without diabetes, individuals with type 2 diabetes were older (65.9 vs. 64.9 years), had less years of education (10.5 vs. 12.1 years) and took more treatments for high cholesterol (62.2 vs. 45.1%). Moreover, most individuals with diabetes were being treated with metformin (75.0%), only 6.7% were taking insulin and 34.5% were taking other oral medications for diabetes (alone or in combination with metformin or insulin). As expected, participants with diabetes had poorer glycemic profile than those without diabetes, with higher values of HbA1c (mean of 7.0 vs. 5.8%), fasting plasma glucose (mean of 146 mg/dL vs. 103 mg/dL) and HOMA-IR index (mean of 5.0 vs. 7.1).

Figure 1 includes a flow diagram of the follow-up of the present study.

First, participants with type 2 diabetes were classified into three distinct latent subgroups based on their HbA1c trajectory from baseline to 1 and 3 years of MedDiet intervention. Subgroup 1 (S1) contained most of the subjects with diabetes (83.1%, N = 123), and the remaining 11.5% (N = 17) and 5.4% (N = 8) were grouped into subgroups 2 (S2) and 3 (S3), respectively. Figure 2A shows the trajectories of HbA1c for the three different subgroups. At baseline, those in S1 presented good glycemic control (HbA1c <7%), with a mean (SD) of 6.6% (0.55), which decreased to 6.3% (0.6) after 1 year but following a return to 6.5% (0.7) after 3 years. This subgroup was termed “S1. Good-Stable glycemic control pattern.” S2 individuals presented poor glycemic control at baseline (HbA1c > 7%), with a mean (SD) of 9.0% (0.9), but it improved during the follow-up, with values of 7.9% (1.3) after 1 year and of 7.1% (1.1) after 3 years. This subgroup was termed “S2. Poor-Improved glycemic control pattern.” Finally, those in S3 also presented poor glycemic control at baseline with high HbA1c levels (mean of 8.1%, SD = 0.81). Although after 1 year their glycemic control slightly improved (HbA1c declined to a mean of 7.5%, SD = 0.70), after 3 years it worsened and increased to 9.6% (0.7). This subgroup was termed “S3. Poor-Worsen glycemic control pattern.” Figure 2B shows the HbA1c trajectory of individuals without diabetes.

As shown in Figure 2C, glucose trajectories in each subgroup mirrored HbA1c trajectories but with more variability, represented by a wider 95%CI. Figure 2D shows the mean glucose trajectory of individuals without diabetes. Regarding MedDiet adherence, at baseline, all subgroups scored a mean of 7.7–7.8 points, which is over the low adherence cut-off of 7 points and thus considered as moderate adherence. All subgroups presented mean improvements in their MedDiet adherence after 1 and 3 years. However, as shown in Figure 2E, S3 presented lower MedDiet adherence than S1 after 1 year (d = −0.87, 95% CI −1.59, −0.14) and especially after 3 years of follow-up (d = −1.16, 95% CI −1.89, −0.43). S2 also scored lower in MedDiet adherence than S1 after 1 year (d = −0.57, 95%CI −1.09, −0.04) but after 3 years they scored a high adherence for the MedDiet similar to the observed in S1 (11.5 vs. 11.8 points for S2 and S1, respectively). Despite of these moderate-to-large differences in MedDiet adherence between subgroups, results were not statistically significant, neither at baseline nor at the follow-up, probably due to the small subgroups sample size. See details in Supplementary Table 3. Figure 2F shows the MedDiet trajectory of individuals without diabetes.

Finally, metformin was used by 70.7% (N = 87) of participants from S1, and this treatment was maintained in 89.7% of them after 1 and 3 years of follow-up. From the remaining N = 36 participants from S1 that were not taking metformin at baseline, N = 26 (72.2%) continued without taking metformin throughout the 3 years of follow-up. Therefore, in S1 the metformin prescription did not change substantially throughout the follow-up (Supplementary Figure 1). Moreover, most of subjects from S1 who were not taking metformin did not take any treatment for type 2 diabetes (75%) or were taking other oral antidiabetic drugs (16.7%). The description and comparability between subgroups in terms of baseline characteristics and medications for diabetes is included in Supplementary Tables 4, 5.

Within the main subgroup of subjects with diabetes with a good-stable glycemic control (S1), those subjects using metformin (N = 87) were compared to those not using metformin (N = 36) (Supplementary Table 6). Before applying IPTW, those not using metformin presented higher total cholesterol (213 vs. 184 mg/dL) and higher LDL-cholesterol (131 vs. 108 mg/dL), but these differences vanished after matching with IPTW.

As shown in Table 2 and Figure 3, at baseline, individuals with type 2 diabetes from S1 treated with metformin scored moderately higher in memory (d = 0.38, 95% CI −0.02, 0.79; P = 0.115), executive functions (d = 0.51, 95% CI −0.06, 1.08; P = 0.086) and global cognition (d = 0.48, 95% CI −0.01, 1.04; P = 0.124) than those not treated with metformin. Nonetheless, an effect difference ranging from d = −0.06, a very small negative association, to 1.08, a large positive association, is also compatible with our data. However, no between-group differences in cognition were observed after 1 and 3 years of follow-up, given that those not treated with metformin compared to those treated with metformin experienced an improved performance from baseline to 3 years in memory (mean change of 0.38 vs. 0.10 SD from baseline mean [z-score], P = 0.036 for the between-group difference in mean change), executive functions (mean change of 0.36 vs. 0.02, P = 0.005) and global cognition (mean change of 0.29 vs. −0.02, P = 0.001).

Supplementary Table 7 includes the differences in each specific cognitive test between subjects with type 2 diabetes (S1) treated and not treated with metformin. At baseline, those treated with metformin scored moderately higher in decision-making abilities (d = 0.60, P = 0.002) and visuoconstructive praxis and attention (d = 0.53, P = 0.018), as well as slightly higher in short- and long-term visual memory (d = 0.40, P = 0.050; and d = 0.45, P = 0.032, respectively). However, after 3 years, those using metformin scored lower in short- and long-term verbal memory (d = −0.42, P = 0.019; and d = −0.31, P = 0.051, respectively) and presented lower improvements in decision-making abilities (mean change of −1.2 vs. 15.9 points, P = 0.015).

Finally, at baseline, both groups presented a moderate adherence to the MedDiet (mean score of 7.7). However, after 3 years of dietary intervention, MedDiet adherence increased to 12.6 points (95% CI 11.8, 13.4) in subjects with type 2 diabetes not treated with metformin and to 11.5 points (95% CI 10.9, 12.1) in subjects treated with metformin, a difference which was statistically significant (d = −0.44, 95% CI −0.84, −0.02; P = 0.031).

Irrespective of metformin exposure, participants with type 2 diabetes from S1 were compared to participants without diabetes. Participants with diabetes presented lower total cholesterol, LDL-cholesterol, lower HDL-cholesterol and were more physically active (Supplementary Table 8). Despite large differences in the glycemic profile (Supplementary Table 9), both groups did not differ in terms of MedDiet adherence and BMI, although subjects with diabetes scored higher in depressive symptomatology at baseline (d = 0.39, 95% CI 0.18, 0.60) and after 1 year (d = 0.63, 95% CI 0.41, 0.85).

There were no differences in baseline memory, executive functions and global cognition between subjects with and without type 2 diabetes (as shown in Figures 4A–D and Supplementary Table 10). However, from baseline to 3 years, those without diabetes exhibited a greater increase in their memory performance (mean change in memory z-score of 0.55 vs. 0.10, P < 0.001 for between group differences in mean change). This increase in the memory composite was mainly due to improvements in short- and long-term verbal memory. Subjects without diabetes also presented greater improvements in visuoconstructive praxis and attention (mean change of 2.8 vs. 0.9 points in RCFT figure copy task, P = 0.009) and did not present reductions in inhibition (mean change of −0.5 vs. −5.1 points in Stroop interference score), compared to subjects with diabetes. However, participants with diabetes presented greater reductions in the reaction time after 3 years (−48.5 ms vs. −0.5 ms in the CPT HRT, P = 0.016).

Subjects with type 2 diabetes from S1 treated with metformin were compared to subjects without diabetes (Supplementary Table 11). In the matched analysis (Figures 4E–H and Supplementary Table 12), baseline cognition did not differ between individuals with diabetes treated with metformin and individuals without diabetes. However, subjects without diabetes showed greater increases in memory after 1 and 3 years (mean change in z-score of 0.53 vs. 0.14, P < 0.001) and in global cognition after 3 years (mean change in z-score of 0.25 vs. −0.001, P = 0.003), compared to subjects with diabetes treated with metformin. These two groups did not differ in terms of executive functions and MedDiet adherence, neither at baseline nor after 1 and 3 years of follow-up.

Participants without diabetes were compared to participants with type 2 diabetes from S1 who were not treated with metformin (Supplementary Table 13). Before matching these groups only differed in HDL-cholesterol (52.4 mg/dL in subjects without diabetes vs. 48.0 mg/dL in subjects with diabetes without metformin) but after matching this difference was balanced and no additional differences were observed. No differences in memory, executive functions and global cognition were observed between subjects with type 2 diabetes not treated with metformin and subjects without diabetes, except in the mean rate of change in executive functions, in which participants with diabetes not treated with metformin presented a greater improvement than participants without diabetes (mean change in z-score of 0.33 vs. 0.08, P = 0.032, as shown in Figures 4I–L and Supplementary Table 14). Moreover, participants with type 2 diabetes not treated with metformin also presented a greater adherence to the MedDiet after 3 years of follow-up (d = 0.32, 95%CI −0.03, 0.67; P = 0.046).

This is the first study to date to examine the effect of metformin on cognition in older adults with type 2 diabetes following a MedDiet intervention. We first examined the heterogeneity in HbA1c trajectories after 1 and 3 years of dietary intervention. We identified three different subgroups of individuals with diabetes irrespective of the intervention group. The largest group exhibited good glycemic control that remained stable during the follow-up, while the remaining two subgroups showed poor baseline glycemic control that improved or worsened during the follow-up. Among the group with good glycemic control, we observed that those treated with metformin presented a better baseline performance in memory, executive functions and global cognition than those not treated with metformin. However, those not treated with metformin presented higher adherence to the MedDiet over time as well as greater improvements in memory, executive functions and global cognition, so that baseline differences between individuals with type 2 diabetes treated and not treated with metformin vanished after 1 and 3 years of MedDiet intervention. These results suggest that adherence to a MedDiet intervention could be superior to the potential neuroprotective effects of metformin among older adults with overweight/obesity and metabolic syndrome who have good glycemic control of their type 2 diabetes.

Our results suggest that metformin could have neuroprotective effects. Specifically, we observed that before starting the MedDiet intervention, individuals with type 2 diabetes from a group presenting good glycemic control (S1) treated with metformin presented a higher performance in memory, executive functions and global cognition than those not treated with metformin. These results agree with previous observational studies showing better memory performance (14) or greater maintenance of executive functions and global cognition (34) in cognitively normal subjects with diabetes type 2 treated with metformin, compared to those not treated with metformin. Metformin use has also been associated with lower dementia risk (35, 36) and better cognitive function (37), but results are still highly variable across studies (15). These inconsistencies could be explained by the fact that multiple neurocognitive pathways are affected by diabetes. Therefore, some pathways, but not all, may be improved with drug therapy (e.g., neurovascular alterations) (38). Moreover, metformin has the adverse effect of lowering serum vitamin B12 concentration, which can in turn, increase the risk of cognitive impairment (39).

Metformin mainly acts by reducing liver gluconeogenesis and inhibiting glucagon-mediated signaling in the liver, but it can also cross the blood brain barrier and thus affect the brain more directly (40). However, the potential neuroprotective effects of metformin have been mostly attributed to its anti-inflammatory and anti-coagulative properties, the prevention of metabolic syndrome (41) and the reduction of peripheral insulin levels that affect brain clearance of amyloid β-peptide (Aβ) (42). Several clinical trials have also tested the effects of metformin in subjects with mild cognitive impairment (MCI) (43) or Alzheimer's disease (AD) (40). In a pilot study in individuals with amnesic MCI (N = 80), metformin treatment for 12 months marginally improved the selective reminding score, but did not affect the global cognitive composite (ADAS-Cog) (43). However, in a cross-over RCT with 20 individuals with AD, metformin treatment for 8 weeks improved executive function and trends also suggested improved memory, learning and attention (40). These preliminary findings support the need for larger trials to evaluate the efficacy and cognitive safety of metformin in prodromal and dementia stages of AD, such as the one recently promoted by the University of Columbia (ClinicalTrials.gov Identifier: NCT04098666).

Our results also suggest that a higher adherence to the MedDiet could reverse the cognitive disadvantage of those subjects with diabetes that were not treated with metformin, since both groups with diabetes achieved similar cognitive scores along the follow-up. Subjects with diabetes from S1 not treated with metformin presented improvements in memory, executive functions and global cognition composites during the 3 years of follow-up, but cognition remained almost stable among those treated with metformin. Moreover, subjects with diabetes who were not exposed to metformin showed greater adherence to the MedDiet after 3 years of follow-up. The reason for this is unknown, but this indicates a group of subjects with a high capacity to make lifestyle changes. In fact, prior to participating in this study, most subjects from this group were able to control their blood glucose without medication and when offered a lifestyle intervention they adhered to it very faithfully, which probably translated into cognitive improvement. Another possibility is that those individuals with type 2 diabetes who take anti-diabetic drugs value lifestyle interventions less than those who do not take any drugs. Moreover, the greater compliance with the MedDiet among individuals with type 2 diabetes not taking metformin at baseline may also explain why they did not require metformin during the 3 years of follow-up. Previous studies have already reported the delayed need of medication for diabetes in patients with a newly diagnosed type 2 diabetes after a MedDiet intervention, compared to a low-fat diet (44, 45).

The favorable effect of the MedDiet intervention was likely due to the overall composition of the dietary pattern and not to a decreased caloric intake, weight loss or increased physical activity, because the allocation to the intervention or control group was balanced among subjects with diabetes either treated and not treated with metformin. In individuals with type 2 diabetes, the MedDiet has been consistently associated with better glycemic control (reduction of HbA1c by 0.32–0.53%) and a better profile of cardiovascular risk factors, compared to low-fat diets (4). These mechanisms could explain why adherence to the MedDiet might improve cognition in individuals with type 2 diabetes (46, 47). The high content in plant-based foods of the MedDiet (olive oil, legumes, vegetables, fruit, cereals, and nuts), along with fish and moderate red wine consumption during meals, make the MedDiet rich in phenolic compounds, n-3 polyunsaturated fatty acids and vitamins that, in conjunction, may contribute to a reduced oxidative stress and chronic inflammation and better neurovascular health (48, 49).

When individuals with diabetes were compared to those without diabetes, we did not find baseline differences in memory, executive functions and global cognition composites. These results differ from previous cross-sectional studies in the overall PREDIMED-Plus population (N = 6,823) showing worse executive functioning (evaluated with different neuropsychological tests) at baseline among participants with type 2 diabetes (8). However, we observed that participants with diabetes experienced fewer improvements in memory than participants without diabetes after 3 years of follow-up.

Nevertheless, in our study subjects with diabetes not treated with metformin experienced a greater increase in their executive functions than subjects without diabetes after 3 years of follow-up. Therefore, their greater adherence to the MedDiet could explain this difference in the rate of change in executive functions. In turn, MedDiet adherence did not differ between subjects with diabetes treated with metformin and subjects without diabetes. However, those using metformin experienced a lower improvement in their memory after 1 and 3 years, and in their global cognition after 3 years of follow-up, compared to subjects without diabetes. Thus, in the face of equivalent adherence to MedDiet, metformin was unable to neutralize the negative impact of type 2 diabetes on cognition.

The strengths of this study include its longitudinal design with 3 years of follow-up and the large number of cognitive tests that are administered to participants, covering 12 different cognitive abilities that are grouped in memory, executive functions and global cognitive composites. Moreover, the methodology used in the analysis of results allowed us to minimize confounding by indication which is not frequently addressed in most studies of metformin and cognitive associations. Finally, we also described the heterogeneity in the response to a MedDiet intervention among individuals with diabetes type 2, which aligns with the current recommendations of more patient-centered research and care in the field of diabetes (3).

However, this study has several limitations. First, the small sample size of the population with diabetes (N = 148) leads to a small number of participants in the subgroups with poor glycemic control (N = 17 in S2 and N = 8 in S3). This limits the capacity to detect statistically significant differences in their baseline characteristics. Moreover, when testing the associations between metformin use and cognition within the group presenting good glycemic control (S1), the small sample size of the untreated (N = 36) and treated (N = 87) groups also limited the study of gender effects, which should be addressed in future studies. Moreover, this sample size constraint also prevented the stratification of subjects exposed and not exposed to metformin according to their adherence to the MedDiet (high/medium/low). Consequently, it was not possible to study the simultaneous effect (interaction) of metformin use and high MedDiet adherence on cognition.

Second, there were losses in the evaluation of the cognitive function during the follow-up (within S1, 14% in the first year and 18% in the third year). They were not unexpected given the burden of neuropsychological visits and the fact that visits of this sub-study were performed on different days to those of the main trial. In addition, executive functions and global cognition composites excluded participants from the UV study site (representing 27% of subjects from S1) since not all the tests that made up the construct of executive functions were administered in this site. Therefore, selection bias cannot be completely excluded from this study.

Third, our methodology was not suitable for investigating causal effects since metformin administration was not randomized, and we did not collect data on the duration of metformin use, specific doses, or patients' adherence to their medication regimens. However, we noted that participants did not change their metformin treatment during the 3 years of follow-up. Moreover, we used IPTW to match treated and untreated subjects in each comparison. This approach allows to account for systematic differences in comorbidities between groups and is used to limit confounding by indication. We also had no information about the APOE genotype of participants, which could influence the association between metformin use and cognitive decline, as reported in previous studies (14).

Finally, this study does not have a control group since all subjects were exposed to a MedDiet intervention. However, without any intervention, individuals with metabolic syndrome would have probably presented a cognitive decline over time (50) and in this study their cognition improved independently of their underlying pathological condition.