We scrutinized the Web of Science, the Cochrane Central Register of Controlled Trials, the Cochrane Database of Systematic Reviews, MEDLINE, EMBASE, and ClinicalTrials.gov (National Library of Medicine NIH U. S., 2023) for clinical trials from their inception to June 25, 2023. Searches were conducted using a range of medical terminology for currently available SGLT2-I, GLP1-RA, and DPP4-I drugs; T2DM; and stroke (Supplementary Appendix 1). Each class of AHG active comparators was reviewed separately.

Randomized controlled trials (RCTs) that enrolled adult participants with T2DM, aged ≥18 years, and reported stroke events and treated with SGLT2-I, GLP1-RA, or DPP4-I compared with other AHG active comparators or placebo were selected.

The extracted data included information about the baseline characteristics of the enrolled population, the study design, the number of participants, the backgrounds of users or non-users of AHG active comparators, and the duration of follow-up.

To assess the quality of the evidence within RCTs, each trial was appraised according to the described criteria of bias in the Cochrane Tool Assessment (Higgins et al., 2003) for six domains: selection, detection, performance, attrition, reporting, and other. These domains were rated to be either “high risk,” “unclear,” or “low risk” of bias. The trial was considered to have low-quality evidence if any of the selection, detection, or performance domains were judged as a “high risk” of bias.

The uncertainty of RCTs was assessed within a meta-analysis based on the GRADE methodology (Guyatt et al., 2008), specifying a quality of “high,” “moderate,” “low,” or “very low” to each included trial. If the meta-analysis involved several RCTs with low risk, the quality of the trials was considered “high”. However, a trial's quality score might be downrated if there was a concern about the trial's indirectness, inconsistency, risk of bias, publication bias, or imprecision. The quality of each trial was assessed by H.A. and adjudicated by J.D. to resolve any disagreements.

The risk of stroke was the primary outcome. The secondary outcomes included non-fatal stroke reported by subtypes (ischemic or hemorrhagic) or fatal stroke. We also included outcomes of stroke reported as serious or non-serious events, which was defined according to the MedDRA Medical Dictionary for Regulatory Activities (2022). Together, these events were used to explore the collective impact of SGLT2-Is, GLP1-RAs, and DPP4-Is by the AHG active comparators' subclass on the risk of stroke.

If RCTs were designed to simultaneously evaluate the risk of stroke by subtypes (ischemic and hemorrhagic stroke), one of these outcomes of stroke were excluded from the intervention of interest arm, and the data for stroke were recalculated by subtypes separately and, following that, vs. our treatment of interest with other AHG active comparators or placebo. Stroke outcome data were collected for both long and short durations if an RCT included multiple follow-up periods.

We grouped each AHG treatment of interest, SGLT2-Is, GLP1-RAs, and DPP4-Is drug class, individually for RCTs with variable doses of these drugs to avoid unit-of-analysis errors. If RCTs with ≥2 intervention arms were designed to simultaneously assess SGLT2-Is or/and GLP1-RAs or/and DPP4-Is, one or two of these drugs were excluded from our intervention of interest arm, recalculated the data separately, and then compared with other AHG active comparators' subclass or placebo. We identified whether participants received an AHG active comparator from the following drug classes: sulfonylurea, metformin, thiazolidinediones, DPP4-Is, GLP1-RAs, α-glucosidase inhibitors, SGLT2-Is, and insulin.

Meta-analyses were performed for the primary and secondary stroke outcomes. Stroke events were analyzed, through subgroup sensitivity analyses, to explore the net stroke safety effect of relatively small trials (Phase II or III) vs. large trials (Phase IV), comparing the populations with T2DM, those with different baseline characteristics, those not at risk, and those at risk for chronic kidney disease (CKD) or CVD risk across trials.

Fixed and random effect models (Riley et al., 2011) were applied independently. The heterogeneity levels inter-study were quantified to assess the percentage of variability across trials by the I² statistic using the Q-statistic. If the I² value was ≥75, a considerable heterogeneity not owing to chance was indicated (Higgins et al., 2003). To determine the risk of potential reporting bias, Egger linear regression (Egger et al., 1997) and Begg rank correlation (Begg and Mazumdar, 1994) tests were used, which were visually evaluated with funnel plots. Reporting bias was considered absent if the funnel plot displayed a symmetrical inverted funnel.

For the quantitative evaluation, a summary value using pooled meta-analysis data was set. For RCTs with zero events in any treatment of interest arms, we combined risk of stroke using a Cochran Mantel–Haenszel test with a continuity correction of 0.5. For dichotomous outcomes, we calculated the incidence rate per 100 patient-years, 95% confidence intervals (CIs), risk ratios (RRs), absolute risk difference, and absolute risk. However, to avoid data distortion, trials in both arms with zero events were excluded.

To produce graphical outputs and perform statistical analyses, the GRADEpro GDT (GRADEpro GDT, 2022) and Review Manager (RevMan 5, 2022) software were used. For all outcomes and heterogeneity analyses, the statistically significant levels were set at a p-value of less than 0.05, and all tests were two-sided.

Participants were diagnosed previously with T2DM prior to study enrollment. The mean age of the participants was 79 years, and 61.5% of them were male. The control of T2DM was relatively poor, with a glycated hemoglobin A1c (HbA1c) level of approximately 10%. Among 259 RCTs, 55.2% of participants who received SGLT2-Is, 54.1% who received GLP1-RAs, and 43.2% who received DPP4-Is had preexisting CVD. Furthermore, 16.3% of people who received SGLT2-Is, 16.5% who received GLP1-RAs, and 20.3% who received DPP4-Is had CKD. Finally, 10.9% of people who received SGLT2-Is, 5.2% who received GLP1-RAs, and 3.9% who received DPP4-Is had experienced a previous stroke.

Most trials employed different comparative dosing therapies, which were administered via a subcutaneous injection or orally. Among 259 trials, participants in 59 trials were naïve to AHG active comparators, whereas those in 200 trials were exposed to diverse treatments of AHG agents. The trials' median follow-up durations ranged from 3 to 73 months for SGLT2-Is, 3 to 78 months for GLP1-RAs, and 1 to 99 months for DPP4-Is.

In total, 208, 19, and 32 trials were scored with a low-, unclear-, and high-bias risk, respectively. For the random sequence generation, there was an adequate descriptive of stroke outcomes data. In 32% of 32 trials, however, had an open-label design and were scored with a high risk of performance bias. The detection and selection biases had an unclear risk in 19% of trials (Figure 2, Supplementary Figure S1).

In total, 82 trials did not report stroke events, whereas stroke outcome data were reported in 177 trials. For most RCTs, the change in the levels of HbA1c from baseline was the primary outcome. However, 19 RCTs were specifically designed to assess the outcomes of MACEs and included events of stroke as secondary or primary outcomes. Trial results are detailed in Supplementary Tables S4–S6. For 177 RCTs with at least one stroke event, meta-analyses were performed. For all classes, GRADE scores were moderate (Supplementary Tables S7–S24).

In people with T2DM, stroke risk was 4% with the use of SGLT2-Is. Across the SGLT2-Is trials, there was no significant evidence of statistical heterogeneity (I² = 0%) or subgroup difference (p = 0.58). The overall risk of stroke with the SGLT2-Is class and the comparators class is summarized in Table 1.

In 49 SGLT-Is trials (1,573 strokes; 75,738 participants), SGLT-Is use did not decrease stroke risk (RR: 0.96, 95% CI [0.87, 1.06], p = 0.42) vs. non-SGLT-Is use (Supplementary Figures S2, S3). The risk of non-fatal hemorrhagic (RR: 0.50, 95% CI [0.14, 1.85], p = 0.30) or non-fatal ischemic (RR: 0.97, 95% CI [0.87, 1.07], p = 0.53) events did not decrease vs. placebo. The risk of non-fatal hemorrhagic (RR: 0.42, 95% CI [0.08, 2.12], p = 0.29) or non-fatal ischemic (RR: 0.94, 95% CI [0.58, 1.52], p = 0.79) events also did not decrease vs. other AHG active comparators (fixed and random effect models; Figure 3, Supplementary Figure S4). Meta-analysis results of the SGLT-Is trials for stroke risk with different baseline characteristics across studies, those diabetic populations not at a risk, or those at risk for CKD or CVD are summarized in the Supplementary Appendix (fixed and random effect models, Supplementary Figures S2, S3).

In SGLT-Is RCTs (6 trials, 124 strokes, 31,327 participants), the risk of fatal hemorrhagic and ischemic stroke did not decrease with SGLT2-Is use (RR: 0.89, 95% CI [0.62, 1.28], p = 0.54) vs. non-SGLT2-Is use (fixed and random effect models; Supplementary Figures S5, S6).

In people with T2DM, stroke risk was 16% with the use of GLP1-RAs. Across the GLP1-RAs trials, there was no significant evidence of statistical heterogeneity (I² = 0%) or subgroup difference (p = 0.94). The overall risk of stroke with the GLP1-RAs class and the comparators class is summarized in Table 2.

In 33 GLP1-RAs trials (1,637 strokes, 70,443 participants), GLP1-RAs use did decrease stroke risk significantly (RR: 0.84, 95% CI [0.77, 0.93], p = 0.0005) vs. non-GLP1-RAs use (Supplementary Figures S7, S8). GLP1-RAs use did decrease non-fatal ischemic events (RR: 0.85, 95% CI [0.77, 0.94], p = 0.002) but did not decrease non-fatal hemorrhagic events (RR: 0.50, 95% CI [0.09, 2.73], p = 0.42) vs. placebo. The risk of non-fatal hemorrhagic (RR: 1.66, 95% CI [0.41, 6.73], p = 0.48) or non-fatal ischemic (RR: 0.80, 95% CI [0.45, 1.44], p = 0.46) events did not decrease vs. other AHG active comparators (fixed and random effect models; Figure 4, Supplementary Figure S9). Meta-analysis results of GLP1-RAs trials for stroke risk with different baseline characteristics across studies, those diabetic populations not at risk, or those not at risk for CKD or CVD are summarized in the Supplementary material (fixed and random effect models; Supplementary Figures S7, S8).

In the GLP1-RAs RCTs (7 trials, 181 strokes, 46,097 participants), the risk of fatal hemorrhagic or ischemic stroke did not decrease with GLP1-RAs use (RR: 0.77, 95% CI [0.58, 1.03], p = 0.08) vs. non-GLP1-RAs use (fixed and random effect models; Supplementary Figures S10, S11).

In people with T2DM, stroke risk was 9% with the use of DPP4-Is. Across DPP4-Is trials, there was no significant evidence of statistical heterogeneity (I² = 0%) or subgroup difference (p = 0.44). The overall risk of stroke with a DPP4-Is class drug and the comparators class is summarized in Table 3.

In 95 DPP4-Is trials (1,488 strokes, 109,243 participants), DPP4-Is use did not decrease stroke risk (RR: 0.91, 95% CI [0.83, 1.01], p = 0.07) vs. non-DPP4-Is use (Supplementary Figures S12, S13). The risk of non-fatal hemorrhagic (RR: 0.68, 95% CI [0.22, 2.12], p = 0.51) or non-fatal ischemic (RR 0.95; 95% CI [0.84, 1.07]; p =0.40) events did not decrease vs. placebo. The risk of non-fatal hemorrhagic (RR: 1.29, 95% CI [0.58, 2.90], p =0.54) or non-fatal ischemic (RR: 0.82, 95% CI [0.67, 1.01], p = 0.06) events also did not decrease vs. other AHG active comparators (fixed and random effect models; Figure 5, Supplementary Figure S14). Meta-analysis results of DPP4-Is trials for stroke risk with different baseline characteristics across studies, those diabetic populations not at risk, and those at risk for CKD or CVD risk are summarized in the Supplementary material (fixed and random effect models; Supplementary Figures S12, S13).

In the DPP4-Is RCTs (10 trials, 181 strokes, 48,102 participants), the risk of fatal hemorrhagic or ischemic stroke did not decrease with DPP4-Is use (RR: 0.93, 95% CI [0.70, 1.24], p = 0.64) vs. non-DPP4-Is use or placebo (fixed and random effect models; Supplementary Figures S15, S16).

Meta-analysis results of trials for risk of stroke by clinical trial size (Supplementary Figures S17–S22), stratification of new AHG active comparators by antidiabetic treatments subclasses (Supplementary Figures S23–S28), and stratification by AHG active comparators of individual subclasses (Supplementary Figures S29–S34) are detailed in the Supplementary material.

In the management of diabetes, the chemotherapeutic advancement of new antidiabetic drugs has challenged the abilities of researchers to demonstrate their favorable neuroprotective effects. The clinical evidence remains uncertain whether the mortality rates of stroke was increased with sulfonylurea (Phung et al., 2013) or if it decreased with the use of metformin (Griffin et al., 2017). In this meta-analysis, when 100 diabetic people at risk of MACEs were treated with GLP1-RAs for 5 years, the risk of non-fatal strokes would be expected to decrease in 2 people (0 to 1 fewer) without an increase in the risk of fatal strokes vs. SGLT2-Is, DDP4-Is, or placebo. Previous meta-analyses on SGLT2-Is (Saad et al., 2017), GLP1-RAs (Darsalia et al., 2015; Mahmouda et al., 2017), and DPP4-Is (Darsalia et al., 2015; Savaresea et al., 2016; Mahmouda et al., 2017; Rehman et al., 2017; Xu et al., 2017) yielded discordant results for the risk of MACEs and stroke. However, unlike the observed neuroprotective putative effects of pioglitazone in diabetic and non-diabetic individuals with a previous history of stroke (Lee et al., 2017). No such impacts were observed with these antidiabetic drugs in stroke populations.

This review evaluates the impacts of SGLT2-I, GLP1-RA, and DDP4-I subclasses and classes in comparison to other AHG active comparators and placebo on the risk of non-fatal and fatal ischemic and hemorrhagic stroke in people with T2DM regardless of their risk for either CVD or CKD. One caveat of this review is that the generalizability of stroke outcome results may only be applicable to a diabetic population at risk of strokes, which may affect the outcomes' generalizability to a population with a previous history of strokes. Even though most of the included studies were conducted on Phase IV trials with a relatively few fatal stroke events and longer follow-ups period, fatal or non-fatal outcomes of stroke were obtained from secondary or primary endpoints of MACEs, as an exposure time to first stroke, or censoring events. None of these included trials stratified or predicted the recurrent stroke events according to etiological stroke type or subtype or by their severity across gender differences. Additionally, findings from older participants with poor metabolic control and an average age of 79 years may not extend to younger populations or those with better initial control of their diabetes. Consequently, the target population with diabetes appeared to exhibit diverse antidiabetic treatment responses, suggesting a need to tailor therapeutic strategies in diabetes management to potentially reduce stroke risk.

The review findings indicate that GLP1-RAs, but not SGLT2-Is or DPP4-Is, decrease non-fatal stroke risk in diabetic people. These findings may guide the treatment decisions of stroke clinicians when treating diabetic people at risk of stroke and MACEs, as well as inform future trial designs for primary and secondary stroke mortality prevention and functional outcomes with these new antidiabetic drugs.