This is a single-center retrospective review of patients aged > 18 years initiating alpelisib treatment at a tertiary cancer center, Memorial Sloan Kettering (MSK). The research protocol was approved by MSK’s Institutional Review Board. Patients were potentially eligible if they initiated alpelisib on any date prior to May 26, 2022. To ensure that findings of the current work were not duplicative of a previous publication, patients were excluded if any of their data were included in that publication [9]. No exclusion criteria by cancer type, or stage were imposed, although it was expected that nearly all patients received alpelisib according to its clinical indication for metastatic breast cancer in patients with activating PI3K mutations.
Alpelisib users were identified by a database query for electronic prescriptions for alpelisib. Alpelisib use was confirmed by manual chart review by two reviewers (any of MAW, DL, RD, JF). Baseline comorbidities, height, weight, body mass index (BMI, in kg/m2), random serum and point-of-care glucoses, serum creatinine, serum albumin, serum bicarbonate, calculated anion gap, and HbA1c levels were extracted from the electronic medical record (EMR). Estimated glomerular filtration rate (eGFR) was calculated from serum creatinine using the CKD-EPI creatinine-based 2021 equation [17].
Start and stop dates and indications for alpelisib, all antidiabetic therapies (metformin, SGLT2 inhibitors, dipeptidyl-peptidase 4 [DPP4] inhibitors, glucagon-like peptide receptor agonists [GLP1-RA], sulfonylurea, thiazolidinedione, meglitinide, and insulin), and corticosteroids (prednisone, hydrocortisone, methylprednisolone, and dexamethasone) were captured by manual chart review. Alpelisib interruptions, dose reductions, and discontinuations, along with the reason for these events, were also captured by manual chart review, as were dates and reasons for hospitalizations. All chart reviews were conducted by two reviewers and discrepancies in medication exposure dates were resolved by consensus among three members of the study team. Days spent inpatient (excluding the day on which hospitalization occurred) were excluded from the analysis because outpatient antidiabetic drugs are typically held and replaced with insulin during hospital admission.
Follow-up ended with permanent alpelisib discontinuation, death, loss to follow-up (defined as three or more months with no encounters in the MSK system), or on 6/30/2022.
The co-primary outcomes were change in random glucose levels (mg/dL) as measured from serum or point-of-care testing at MSK facilities, and incidence of DKA. Home glucose monitoring, including continuous glucose monitoring, was not captured. Rates of hospitalization due to hyperglycemia as well as alpelisib interruptions, dose reductions, and discontinuation due at least in part to hyperglycemia were reported along with a composite of all these events as ‘hyperglycemia-related treatment disruption.’ Rates of death and progression of disease were also reported. DKA was defined as events satisfying all of the following diagnostic criteria upon presentation: (1) serum bicarbonate ≤ 18 mmol/L and/or blood pH ≤ 7.3, (2) anion gap > 10 mEq/L, and (3) presence of serum and/or urine ketones.[18] These criteria were applied by two chart reviewers.
Rates of DKA were calculated, both overall and stratified into three exposure categories for patients on alpelisib: (1) no antidiabetic drugs, (2) on antidiabetic drugs excluding SGLT2 inhibitors, and (3) on SGLT2 inhibitors (with or without other antidiabetic drugs). Each case was also described with respect to initial symptoms, vital signs, laboratory data (eg, serum and point-of-care glucose levels), medications, clinical management, and outcome.
Change in blood glucose levels associated with each time varying exposure were described and analyzed using a mixed linear model, hierarchical at the patient level. The analysis of change in blood glucose level was restricted to time periods when patients were taking alpelisib. Patients on antidiabetic drugs at baseline were excluded from this specific analysis, as were patients who did not have random glucoses measured both at baseline and during follow-up. Antidiabetic drugs and corticosteroid exposure were included as time-varying variables (e.g., a patient who took metformin for only a portion of their follow-up time would be classified as metformin-exposed during that period, but not before or afterwards). Unadjusted changes in glucose level from baseline associated with exposure to each antidiabetic drug class and to corticosteroids are reported. In the primary adjusted analysis, antidiabetic drugs and steroid exposure were all included together in the mixed linear model as time-varying covariates. Results were further adjusted for baseline age, sex, date of alpelisib initiation, eGFR, glucose levels, and BMI.
In sensitivity analysis, to address concerns about adjustment for mediators, the mixed linear model was repeated as a marginal structural model (MSM), with stabilized inverse probability weights (IPW) applied to each period of follow-up. For example, if patients with no reduction of serum or point-of-care glucose levels after starting an SGLT2 inhibitor were likely to then stop SGLT2 inhibitor, this could create a bias in favor of SGLT2 inhibitor because non-responders to SGLT2 inhibitors would contribute less follow-up time. Including the initial change in glucose levels as a covariate would potentially adjust for this bias, but such adjustment could itself produce bias by ‘adjusting away’ the initial effects of SGLT2 inhibitor use on glucose levels, even though such an effect would be a mediator of SGLT2 inhibitor benefit, not a confounder. By using weights instead of covariates, an MSM is able to adjust for such potential mediators without producing bias.[19] MSM was conducted separately for metformin, SGLT2 inhibitors, and for corticosteroids. For each MSM, the IPW was calculated using exposure history to metformin, SGLT2 inhibitor, and corticosteroids as well as prior random glucose readings. Other antidiabetic drug classes were not included in the MSM because there were insufficient observations for the models to calculate IPW to converge. Models were also adjusted for non-time varying covariates measured at baseline: age, sex, date of alpelisib initiation, eGFR, glucose levels, and BMI.
Work in preclinical models has indicated that choice of antidiabetic drug may have implications not only for glucose levels on alpelisib, but on levels of insulin (or C-peptide). Differences in insulin level could be clinically significant in this context because high insulin levels could in theory override PI3K inhibition and undermine the anti-cancer effect of alpelisib and similar drugs. In an exploratory analysis, random C-peptide levels were also extracted from the electronic medical record and levels of C-peptide summarized in four exposure categories for patients: (1) not on alpelisib (ie, before or after exposure, since all study patients were on alpelisib at some point), (2) on alpelisib but no antidiabetic drugs, (3) on alpelisib with antidiabetic drugs excluding SGLT2 inhibitors, and (4) on alpelisib and SGLT2 inhibitor (with or without other antidiabetic drugs).