I am analyzing a surgical cohort where about 10% of patients were on SGLT2 inhibitors before surgery. Among these, roughly half continued therapy perioperatively and half had it paused. The remaining 90% of patients were not on SGLT2 inhibitors at all.
My study question: Is pausing SGLT2-inh. before surgery associated with acute heart failure?
My primary outcome is an ordinal composite endpoint over 3 months with increasing severity:
No event
Initiation of i.v. diuretics without hospitalization for acute heart failure (AHF)
Hospitalization for AHF
My main clinical question is whether pausing SGLT2 inhibitors (vs continuing) increases the odds of worse outcomes.
I am considering modeling exposure as a 3-level categorical variable:
No SGLT2
SGLT2 continued (reference)
SGLT2 paused
and then fitting a proportional odds logistic regression (adjusting for age, sex, surgical risk, CAD, CKD, diabetes, CHF, urgency of surgery).
My question:
Is this the optimal way to structure the exposure variable, given that for 90% of the cohort (no SGLT2) the “pause vs continue” concept is not applicable? Or would you recommend a different approach (e.g., separate model restricted to SGLT2 users, interaction coding, etc.)?
The most important issues are doing a rigorous landmark analysis where the follow-up clock is started at the same time (completion of surgery), and that the design is very brittle with respect to confounders. It is important to poll 10 disinterested clinical experts to get their list of factors used in the decision to withhold the drug, and to make sure the data collection included all of their reasons so you can adjust for them. You can add to understanding the practice patterns by modeling the drug continuation choices as a function of several baseline variables and seeing what the big predictors are.
To you original question you can keep all the categories and just expect wide confidence intervals when small subsets are involved.
As you suggested, we sent an online survey to anesthesiologists who are not involved in the study to gather more possible reasons to pause the drug.
Regarding the second answer, I tested and noticed that filtering patients to SGLT2i users to have a binary exposure gives less wide confidence intervals. But i thought we loose a lot of data and have less degrees of freedom to adjust for.
Your research question seems like a very important one. Don’t take anything I’m about to say too seriously, since I’m not a statistician, but rather a physician Assuming that you have access to an experienced statistician who could help you, wouldn’t this scenario be ripe for a subjective Bayesian approach (see the other thread on interpreting “confidence intervals” in observational studies) (?)
As far as I can tell, the main reason why regulatory agencies advised that SGLT-2 inhibitors be held in the preoperative period was because of an increased risk for perioperative euglycemic ketoacidosis. Product labels advise holding SGLT-2 inhibitors for a few days before surgery to reduce the risk of this rare complication.
While it seems like the vast majority of the reported cases of euglycemic ketoacidosis have involved patients with diabetes, I did find a publication that mentioned cases involving non-diabetic patients:
If the overwhelming majority of reported cases of ketoacidosis involve diabetic patients, then it would seem important for regulatory agencies to explain why they decided to implement an “across-the-board” recommendation to hold SGLT-2 inhibitors in the preoperative period, rather than restricting their advice to the diabetic subset of SGLT-2 users. Intuitively, it would seem much less risky for patients using their SGLT-2 inhibitor primarily for sugar control to hold their medication perioperatively, as compared with patients using it primarily for CHF control. Indeed, it seems very plausible that the risk/benefit calculus for holding the medication could be net favourable for patients with diabetes and no CHF, but net harmful for patients with CHF (+/- diabetes) (?)
This very recent publication from Circulation (Aug/25) speaks to the issue. Unfortunately, it’s paywalled, so I could only read the first page- maybe you’ll have access.
The best way to answer your question would obviously be to randomize surgical patients with CHF to either continue their SGLT-2 inhibitors preoperatively, or to hold them preoperatively, and then to compare outcomes between arms. The question of clinical equipoise would be tricky here, though, given the current indication-agnostic labelling advice to hold the medication perioperatively.
Anesthesiologists would be able to tell you whether they routinely hold SGLT-2 inhibitors perioperatively for all their patients or only for those who are diabetic (i.e., whether they interpret the product label with more nuance than it conveys).
Assuming that there’s no RCT in the works to address this question, then the next best option would be a prospective observational study. Anesthesiologists could be surveyed in the design phase to help you determine what type of information/patient-level variables you’d need to collect going forward. You’d then watch what happens to patients with CHF whose medications are held versus continued perioperatively (depending on the best judgement of the patient’s anesthesiologist or heart failure specialist).
However, it sounds like your study will be retrospective, in the sense that the data has already been collected (and you had no control over what was collected)(?) I guess the first step would be for you to draw a DAG, with input from anesthesiologists, before you look at information in the database. If anesthesiologists tell you that there are several important considerations that would factor into their decision to hold or not hold an SGLT-2 inhibitor in a given patient preoperatively, and, if you subsequently discover that your surgical database does NOT provide information about these variables, then you might decide NOT to do the study (?) If you find that you DO have sufficiently-detailed information in the database to proceed, then a Bayesian approach would allow you to elicit (from anesthesiologists and heart failure specialists), their “prior” as to whether they would expect CHF patients to decompensate (or not) post-operatively if their medication were held. I suspect many would say “yes,” meaning that you might, in this situation, have a fairly informative prior (?)
thank you a lot for the answer and important points.
Exactly, you summarized the key reasons why we perform this study. It is a retrospective analysis, however, in a large prospective cohort with a wide range of documented variables.
The Bayesian approach seems very interesting to me, we will definitely discuss it.
Hi Koray,
Interesting research question. I always think it’s useful to think about these things as an RCT you’d want to conduct if you had the resources.
I don’t understand why you’d include people not on SGLT2, because it doesn’t seem relevant to your estimand.
My main clinical question is whether pausing SGLT2 inhibitors (vs continuing) increases the odds of worse outcomes.
These patients don’t appear in your question and wouldn’t be part of your hypothetical RCT. For the patients who continued and stopped you might be able to find an adjustment set to make the two populations exchangeable and target a sensible causal estimand but to me it seems that people without SGLT2 are a very different population and you’d really end up comparing apples with oranges.
Reflecting on why this therapeutic question seems better-suited than most to an observational study, it occurred to me that the clinical reasoning in this scenario would be similar to the reasoning that physicians employ when deciding whether to prescribe a medication “off-label.”
Many conditions in medicine (e.g., chronic pain) have no, or only a few, therapies with well-established efficacy. But all physicians know that patients don’t stop coming to us just because we don’t have any evidence-based therapies to offer for their condition. Many keep coming back with the same problem repeatedly, begging for help. Enter “off-label” prescribing: applying a therapy for an indication different from the one for which it has demonstrated efficacy in RCTs. Here, we’re effectively grasping at straws. We’re in a position where we don’t have a solid reason to believe that anything we prescribe is going to help - we’re making an educated extrapolation. Occasionally, we might get lucky. For example, a patient might report that his arthritis pain improves after we prescribe a medication approved only for pain emanating from other body parts (side note- we could check that the medication is actually the “cause” of his improvement by performing a treatment dechallenge/rechallenge). But, in the event that we don’t get lucky (i.e., the patient doesn’t feel better), we at least hope not to make him worse with our off-label prescribing. Specifically, we really want to avoid applying a therapy that we have reason to believe might harm him. The lower our expectation for efficacy of the therapy we’re applying, the easier it will be for the risk/benefit calculus to become unfavourable. In this context, we are MUCH more likely to allow a potential observational study’s harm signal to influence our clinical decision-making than we are when we apply therapies with well-established efficacy.
Circling back to your research question: At this point, it seems like we might not have any compelling reason to believe that holding CHF patients’ SGLT-2 inhibitors preoperatively will be more beneficial than harmful (?) Indeed, there’s already evidence (from the EMPEROR trial) that CHF patients’ outcomes can be adversely affected within only 30 days of randomized SGLT-2 inhibitor withdrawal in non-surgical settings (see reference below). The outstanding question is whether they’d do worse even with very short periods of cessation in the perioperative period.
Current SGLT-2 inhibitor product labelling has effectively created a situation that’s analogous to applying a therapy (the “therapy” being withholding the drug) that hasn’t been shown to be efficacious in an RCT. So, if your observational study were to suggest that holding CHF patients’ SGLT-2 inhibitors even for a short time preoperatively could be harmful, then this finding could encourage: 1) regulators to refine their labelling advice; and/or 2) someone to design a perioperative-focused RCT to settle the question more definitively.
Hermanides et al. Risk of preoperative discontinuation of SGLT2 inhibitors. BJA June, 2024.
Assuming that you have access to an experienced statistician who could help you, wouldn’t this scenario be ripe for a subjective Bayesian approach (see the other thread on interpreting “confidence intervals” in observational studies) (?)