This was a multicenter retrospective study conducted using the TriNetX Research Network, a federated research platform comprising over 103 participating health care organizations in the United States, including major academic medical centers and large integrated health systems. To preserve partner confidentiality, a complete list of institutions is not publicly available due to data-sharing agreements. Patient follow-up is conducted retrospectively by linking electronic health records over time using pseudonymized patient identifiers within each institution. In some cases, tokenization is used to link records across different data sources, including claims data. The platform reconstructs longitudinal clinical histories and enables time-to-event analyses, censoring patients after their last recorded medical event. All data are deidentified, and no direct patient contact occurs.11 The present analysis was performed on March 31, 2025, using the TriNetX Compare Outcomes tool.

Adult patients aged 18 years or older who underwent TAVI between January 1, 2015 and March 31, 2025, were identified based on standardized procedure codes. The validity of using International Classification of Diseases, 10th Revision (ICD-10) codes to ascertain clinical outcomes in patients undergoing TAVI has been previously demonstrated. In a study by Butala et al.,12 claims-based data showed excellent concordance with trial-adjudicated events for all-cause mortality (sensitivity> 99.6%) and moderate sensitivity for identifying acute kidney injury (70.2%) and myocardial infarction (MI) (63.6%).12 While previous studies have validated the use of ICD-10 codes to capture TAVI outcomes in claims-based datasets, our analysis leveraged electronic medical record data, which provides richer clinical detail and higher sensitivity in outcome detection.9 This is supported by findings from Rahafrooz et al.,13 who reported that electronic medical record-based algorithms achieved 100% sensitivity for mortality and up to 95% sensitivity for HF-related hospitalizations.13 Two cohorts were defined. The first cohort comprised patients with a documented diagnosis of HF who had their first recorded prescription of an SGLT2 inhibitor (empagliflozin, dapagliflozin, canagliflozin, or ertugliflozin) initiated within a 1-month window before or after the index TAVI procedure. This approach ensured that only patients who were SGLT2i-naive at the time of TAVI were included. The second cohort included patients who underwent TAVI but had no recorded exposure to SGLT2i and had a documented diagnosis of HF.

As an exploratory secondary analysis, we compared patients prescribed SGLT2i with those who received other antihyperglycemic agents within one month before or after the TAVI procedure. The non-SGLT2i cohort included patients prescribed any of the following medications: exenatide, tolazamide, tolbutamide, linagliptin, alogliptin, acarbose, chlorpropamide, glimepiride, nateglinide, miglitol, pioglitazone, glyburide, glipizide, sitagliptin, metformin, repaglinide, rosiglitazone, saxagliptin, pramlintide, lixisenatide, dulaglutide, semaglutide, tirzepatide, teplizumab, or liraglutide. Detailed lists of all diagnoses, procedures, and medication codes used to define the study population and exposure criteria are available in table S1 and table S2.

The exposure of interest was defined as prescription of an SGLT2i within 30 days before or after the TAVI procedure. The index event for both cohorts was defined as the date of the first TAVI procedure, and follow-up for all outcomes began 1 day following the index event. Clinical outcomes were assessed 6 months, 1 year, and 5 years after the index event. Demographic characteristics were defined by ICD-10 codes recorded at any time before the TAVI procedure and up to the same day of the TAVI procedure, with the system automatically using the most recent medication and laboratory results available in the patients’ records.

The primary outcome was all-cause mortality during the follow-up period after the index TAVI procedure. Secondary outcomes included acute HF admissions, hospitalization or emergency department (ED) visits, atrial fibrillation or flutter, ventricular tachycardia, MI, ischemic stroke, acute kidney injury, need for dialysis, and all HF encounters, including both acute and chronic presentations. All outcome definitions and associated coding are described in table S3.

Propensity score matching was performed using 1:1 nearest-neighbor matching without replacement and a strict default caliper of 0.01 to balance baseline characteristics between the 2 cohorts. This combination ensures high comparability between matched pairs but excludes treated patients with no adequate control match. The unmatched individuals typically had extreme clinical profiles, and their exclusion minimizes residual confounding, thereby improving internal validity. Variables included in the propensity score model encompassed demographics, comorbidities, medication use, and relevant laboratory values (table 1). Covariate balance before and after matching was assessed using standardized mean differences. Outcomes were compared between matched cohorts using risk differences, risk ratios, and odds ratios with 95% confidence intervals (95%CI). Time-to-event analyses were conducted using Kaplan-Meier survival curves and compared using the log-rank test. Hazard ratios with 95%CI were estimated using Cox proportional hazards regression. For continuous outcomes such as laboratory values, between-group comparisons were performed using Student t tests. Statistical significance was defined as a 2-sided P-value <.05.

All analyses were performed within the TriNetX Analytics platform. TriNetX retains the natural missingness inherent to real-world clinical data and does not perform statistical imputation. While the platform applies limited derivation (eg, estimating glomerular filtration rate from creatinine and assigning encounter dates) and enhances data completeness through natural language processing of clinician notes, the responsibility for handling missing data rests with the researcher. In our analysis, variables were selected based on clinical relevance and data availability. Left ventricular ejection fraction (EF) was included in the primary analysis but excluded from the subgroup comparison with other antihyperglycemics due to limited documentation in the comparator group.

This study used deidentified data and did not involve direct interaction with human participants. As such, it was exempt from Institutional Review Board review in accordance with the US Department of Health and Human Services regulations (45 CFR 46.104[d][4]).

The data used for this analysis were obtained through the TriNetX Research Network under institutional license. Although the data are not publicly available due to licensing agreements, they may be made available by the corresponding author upon reasonable request and with permission from TriNetX.

We identified 58 193 adults who underwent TAVI between January 1, 2015 and March 31, 2025. Among these, 3145 patients with a diagnosis of HF received an SGLT2i within 30 days of the index TAVI procedure. Of these, 3022 were matched 1:1 to 3022 patients with HF who did not receive an SGLT2i, yielding a final matched cohort of 6044 patients. Before matching, patients in the SGLT2i cohort were significantly younger (mean age, 75.2 vs 78.6 years), more likely to be male (65.0% vs 53.3%), had a higher prevalence of diabetes (63.1% vs 33.5%), hypertension (87.3% vs 82.9%), chronic kidney disease (44.4% vs 33.3%), ischemic heart disease (83.3% vs 74.8%), and peripheral artery disease (52.8% vs 49.7%).

In the SGLT2i cohort, patients were more likely to be on guideline-directed medical therapy, with higher use of beta-blockers (83.2% vs 62.7%), ACE inhibitors (38.3% vs 30.1%), and angiotensin receptor blockers (48.9% vs 25.5%). Laboratory data showed that the cohort on SGLT2i had higher hemoglobin A1c (6.7% vs 6.1%), higher N-terminal pro-B-type natriuretic peptide (NT-proBNP) (5990 vs 4968 pg/mL), and a higher glomerular filtration rate (61.8± 25.1 vs 59.8±24.9). Moreover, they had a significantly lower left ventricular EF (48.4% vs 56.4%) (table S4). After matching, baseline characteristics were well balanced (table 1).

The matched groups had a mean age of 75.6±8.7 vs 75.5±11.4 years (P=.915), and similar proportions of male patients (64.4% vs 63.6%; P=.593). Racial distribution was comparable with Caucasian patients comprising the majority in both cohorts, 78.8% and 78.2% (P=.556), and Black/African-American patients, 5.7% and 5.9% (P=.731) in the SGLT2i and non-SGLT2i cohorts, respectively. The comorbidity profile was evenly distributed between the SGLT2i and non-SGLT2i cohorts, including the prevalence of diabetes (62.4% vs 64.5%; P=.149), hypertension (87.4% vs 87.0%; P=.746), chronic kidney disease (44.3% vs 44.9%; P=.365), and ischemic heart disease (83.2% vs 83.4%; P=.814). Medication use at baseline was similar between the SGLT2i and non-SGLT2i cohorts, including beta-blockers (82.7% vs 83.1%; P=.537), angiotensin-converting enzyme inhibitor (ACE inhibitors) or angiotensin receptor blocker (86.5% vs 82.7%; P=.681), diuretics (85.3% vs 85.2%; P=.752), and statins (86.6% vs 86.1%; P=.655) (all P> .05). Laboratory profiles were largely comparable, although the SGLT2i cohort had higher hemoglobin A1c (6.7±1.5% vs 6.3±1.4%; P <.001), higher glomerular filtration rate (61.5±24.8 vs 57.2±28.4; P <.001), and lower NT-proBNP levels (6002.4±8823.4 vs 7883.3±13 486.6 pg/mL; P <.001). Left ventricular EF was slightly lower in the SGLT2i cohort (48.6±17.6% vs 52.1±14.8%; P <.001). Patients in the SGLT2i cohort had a mean follow-up duration of 424.53 days, with a standard deviation of 402.17 days. The median follow-up for this cohort was 334.50 days, and the interquartile range was 515 days. In the other cohort, the mean follow-up was 497.17 days, with a standard deviation of 468.46 days. The median follow-up for this cohort was 367 days, and the interquartile range was 607.50 days (figure S1).

All standardized mean differences were <0.1, indicating successful covariate balance after matching.

The use of SGLT2i was associated with significantly lower all-cause mortality among all follow-up intervals (table 2). At 12 months, all-cause mortality was 7.3% % in the SGLT2i cohort vs 10.5% in the non-SGLT2i cohort (hazard ratio [HR], 0.71; 95%CI, 0.60-0.85; P <.001). The benefit persisted at 5 years (10.7% vs 20.6%; HR, 0.59; P <.001) (figure 1). Survival curves separated early and remained distinct throughout follow-up, as illustrated by the 5-year Kaplan-Meier estimates (figure 2).

Figure 1.

Central illustration. Impact of SGLT2i on long-term outcomes in TAVI patients with heart failure: a propensity-matched analysis. ED, emergency department; HR, hazard ratio; MI, myocardial infarction; OR, odds ratio; SGLT2i, sodium-glucose cotransporter-2 inhibitor; TAVI, transcatheter aortic valve implantation.

(0.36MB).

Figure 2.

Kaplan-Meier curve for all-cause mortality up to 5 years following transcatheter aortic valve implantation in patients with heart failure, based on the matched cohort. Log-rank test: chi-square=61.25, degrees of freedom=1, P <.001. SGLT2i, sodium-glucose cotransporter-2 inhibitors.

(0.14MB).

Secondary outcomes are summarized in table 3. Acute HF events were more common in the SGLT2i cohort early after TAVI. At 3 months, the rate was 5.9% compared with 4.2% in the non-SGLT2i cohort (odds ratio [OR], 1.44; 95%CI, 1.14-1.83; P=.002). At 6 months, the difference remained significant (5.5% vs 4.3%; OR, 1.32; 95%CI, 1.03-1.68; P=.027), but not at 12 months (7.3% vs 6.2%; OR, 1.20; 95%CI, 0.98-1.46; P=.082) or 5 years (7.9% vs 7.3%; OR, 1.09; 95%CI, 0.90-1.32; P=.404).

SGLT2i use was consistently associated with fewer ED visits or hospitalizations at all time points. Event rates were 45.7% vs 56.8% at 12 months (OR, 0.64; 95%CI, 0.58-0.71; P <.001), and 51.0% vs 62.5% at 5 years (OR, 0.63; 95%CI, 0.56-0.69; P <.001).

Rates of ventricular tachycardia were lower in the SGLT2i cohort at each timepoint but did not reach statistical significance: 4.3% vs 4.8% at 12 months, and 5.6% vs 6.4% at 5 years (all P> .05).

Atrial fibrillation or flutter occurred at similar rates between cohorts. No significant differences were observed at 12 months (9.5% vs 9.6%) or 5 years (13.1% vs 13.0%).

Acute MI rates were comparable through 12 months but differed at 5 years. At 12 months, acute MI rates were comparable between groups (9.6% vs 10.6%; P=.200). However, by 5 years, the SGLT2i cohort experienced a significantly lower incidence of acute MI compared with the non-SGLT2i group (12.0% vs 14.4%; OR, 0.81; 95%CI, 0.70-0.94; P=.007). Stroke rates were similar at 3 months (4.2% vs 4.1%) and 12 months (6.4% vs 6.3%), but a reduction was observed at 6 months (4.8% vs 6.1%; OR, 0.79; 95%CI, 0.62-0.99; P=.041). This difference was not sustained at 5 years (8.2% vs 9.1%; OR, 0.89; P=.212).

Renal outcomes, including acute kidney injury and the need for dialysis, did not differ significantly at 6, 12, or 60 months. At 3 months, acute kidney injury occurred in 12.4% of patients in the SGLT2i cohort vs 11.7% of those in the non-SGLT2i cohort (OR, 1.07; P=.400), with the difference remaining negligible over time, reaching 22.6% vs 22.5% at 5 years (OR, 1.00; P=.975). Similarly, dialysis events were numerically lower in the SGLT2i cohort throughout follow-up (0.7% vs 1.0% at 3 months; 2.0% vs 2.5% at 5 years), but these differences were not statistically significant (P=.223 and P=.222, respectively).

In a propensity-matched cohort of 1475 patients per cohort, SGLT2i use was associated with significantly lower all-cause mortality at 12 months (7.6% vs 10.4%; HR, 0.78; 95%CI, 0.61-0.99; P=.044) and at 5 years (11.3% vs 20.2%; HR, 0.69; 95%CI, 0.57-0.83; P <.001) (table S5, figure S2). Rates of ED visits and hospitalizations were consistently reduced in the SGLT2i cohort at all time points, including 5 years (46.2% vs 60.3%; OR, 0.56; P <.001).

The SGLT2i cohort also showed a lower 5-year incidence of atrial fibrillation or flutter (11.7% vs 16.2%; OR, 0.69; P=.011), acute MI (11.2% vs 15.1%; OR, 0.71; P=.002), and stroke (7.7% vs 10.0%; OR, 0.75; P=.028). Renal outcomes favored SGLT2i, with significantly lower rates of acute kidney injury at 6 months (13.2% vs 15.9%; OR, 0.80; P=.037), 12 months (16.2% vs 20.3%; OR, 0.76; P=.004), and 5 years (20.5% vs 25.9%; OR, 0.74; P=.001), as well as reduced dialysis requirements at 5 years (2.0% vs 3.6%; OR, 0.54; P=.010) (table S6).

This large, multicenter retrospective cohort study uses deidentified electronic health record data from 103 healthcare organizations, offering a comprehensive and generalizable assessment of SGLT2i use following TAVI in patients with HF. Major strengths of the study include its substantial sample size, extended follow-up of up to 5 years, and inclusion of clinically relevant variables such as EF, laboratory data, and medication use. The application of propensity score matching further strengthens internal validity by mitigating confounding from measured covariates. However, as with all observational studies, the potential for residual confounding from unmeasured or unknown variables cannot be fully excluded. In particular, while EF was included in the primary analysis, it was excluded from the subgroup comparison with other antihyperglycemics due to limited availability, introducing potential bias related to HF severity. Since TriNetX does not perform statistical imputation, missing data must be addressed through analytic design. Although we used rigorous matching based on available covariates, the potential for residual confounding may have influenced findings. Additionally, outcomes were defined using ICD-10 diagnostic codes and were not independently adjudicated, raising the possibility of misclassification bias. Furthermore, procedural outcomes assessed by echocardiography, such as residual aortic regurgitation, could not be evaluated due to the absence of standardized imaging data among participating institutions within the TriNetX network.

We included patients who had documented exposure to SGLT2i within 30 days before or after the index TAVI procedure to better reflect real-world prescribing patterns and optimize cohort retention. Restricting exposure classification to the post-TAVI period alone would have markedly reduced the size of the exposed cohort and introduced selection bias toward patients with more structured medication regimens, such as clinical trial participants. While this broader definition enhances generalizability, we acknowledge that it does not capture long-term adherence or treatment duration. Additionally, although we could not assess treatment discontinuation in the SGLT2i group, all patients in the non-SGLT2i cohort remained unexposed to SGLT2i throughout the follow-up period, minimizing the risk of exposure crossover. Future studies examining longitudinal medication use and adherence patterns may help clarify the temporal relationship between SGLT2i exposure and sustained clinical benefit.

Finally, due to platform limitations, direct assessment of loss to follow-up within the matched cohort was not feasible, potentially restricting the ability to accurately account for missing longitudinal data. Despite these limitations, the study provides meaningful real-world evidence supporting the association between SGLT2i use and improved outcomes in a high-risk post-TAVI population.