Sodium-Glucose Cotransporter-2 Inhibitors for Hyperglycemia in Phosphoinositide 3-kinase Pathway Inhibition

Purpose Phosphoinositide 3-kinase (PI3K) inhibition is used for the treatment of certain cancers, but can cause profound hyperglycemia and insulin resistance, for which sodium-glucose cotransporter-2 (SGLT2) inhibitors have been proposed as a preferred therapy. The objective of this research is to assess the effectiveness and safety of SGLT2 inhibitors for hyperglycemia in PI3K inhibition. Methods We conducted a single-center retrospective review of adults initiating the PI3k inhibitor alpelisib. Exposure to different antidiabetic drugs and adverse events including diabetic ketoacidosis (DKA) were assessed through chart review. Plasma and point-of-care blood glucoses were extracted from the electronic medical record. Change in serum glucose and the rate of DKA on SGLT2 inhibitor versus other antidiabetic drugs were examined as co-primary outcomes. Results We identified 103 patients meeting eligibility criteria with median follow-up of 85 days after starting alpelisib. When SGLT2 inhibitors were used to treat hyperglycemia, they were associated with a decrease in mean random glucose by −54 mg/dL (95% CI −99 to −8) in adjusted linear modeling. Five cases of DKA were identified, two occurring in patients on alpelisib plus SGLT2 inhibitor. Estimated incidence of DKA was: alpelisib plus SGLT2 inhibitor, 24 DKA cases per 100 patient-years (95% CI 6, 80); alpelisib with non-SGLT2 inhibitor antidiabetic drugs, 7 (95% CI 0.1, 34); alpelisib only, 4 (95% CI 0.1, 21). Conclusions SGLT2 inhibitors are effective treatments for hyperglycemia in the setting of PI3K inhibition, but given possible adverse events, SGLT2 inhibitors should be used with caution.


Introduction
The phosphoinositide 3-kinase (PI3K) pathway is critically important for tumor progression-promoting cell growth, motility, survival, and metabolism. In breast cancer, the PI3K pathway can be hyperactivated by somatic mutations in PIK3CA, which encodes the catalytic subunit of PI3K (P110α). These mutations are found in 40% of breast tumors, and associate with resistance to conventional anticancer agents [1]. Tumor PI3K signaling can be therapeutically targeted by small molecules that speci cally bind and inhibit P110α-like taselisib, serabelisib, inavolisib, and alpelisib. Alpelisib was shown to improve clinical outcomes in postmenopausal women and men with hormone receptor-positive, HER2-negative, PIK3CA-mutant advanced breast cancer in the SOLAR-1 trial (Study Assessing the E cacy and Safety of Alpelisib Plus Fulvestrant in Men and Postmenopausal Women With Advanced Breast Cancer Which Progressed on or After Aromatase Inhibitor Treatment [NCT02437318]) [2].
In addition to blocking the action of mutant PI3K in the tumor, PI3K inhibitors systemically inhibit PI3K in host tissues, which impairs the intracellular action of insulin and induces insulin resistance. PI3K inhibitor driven insulin resistance causes hyperglycemia, which can be severe [3,4]. Of 284 patients who received alpelisib in SOLAR-1, hyperglycemia occurred in 63.7% of patients, with fasting plasma glucose levels exceeding 250 mg/dL in 36.6% [2]. Severe hyperglycemia caused by alpelisib has also led to diabetic ketoacidosis (DKA) in some case reports [5][6][7]. These on-target adverse hyperglycemic events were reported in other trials of P110α inhibitors, as well as inhibitors of AKT, another targetable kinase in the PI3K signaling cascade [8].
Hyperglycemia related to systemic inhibition of PI3K and AKT can result in emergency room visits, inpatient admissions, treatment interruptions, dose reductions, and treatment discontinuation. In a retrospective study of 251 cancer patients new to PI3Kα inhibitor use, we found hyperglycemia caused dose interruption in 13%, dose reduction in 11%, and hospitalization in 3% of patients [9]. Of note, this population largely consisted of clinical trial participants for whom baseline type 1 or 2 diabetes was an exclusion criterion. In 'real world' settings, where patients with diabetes are not excluded, hyperglycemia may be a major challenge to using PI3K or AKT inhibitors. Approaches to hyperglycemia management in patients who receive alpelisib and other PI3K/AKT inhibitors are not clearly established. Blocking the PI3K/AKT pathway can induce a highly insulin resistant phenotype in which some antidiabetic drug classes, particularly insulin and insulin secretagogues, might be ineffective. In SOLAR-1, the study protocol was amended about half-way through enrollment to encourage the use of metformin as a rstline anti-hyperglycemic agent which resulted in its use in 87.1% of patients with hyperglycemia [10]. However, the effectiveness of metformin in the setting of PI3K inhibitor use has still not been systematically studied and the effects of other antidiabetic drug classes in this setting are even less wellunderstood. Choice of an antidiabetic agent in the setting of PI3K/AKT inhibitor use may have implications beyond the quality of glycemic control. For example, high levels of endogenous or exogenous insulin might overcome the therapeutic blockade of PI3K and limit the anti-cancer e cacy, as has been shown in animal models [3].
Metformin is recommended as a rst-line agent for the treatment of alpelisib-induced hyperglycemia, even though its effectiveness in the setting of PI3K blockade is not established [10]. Of alternative antidiabetic drug classes, sodium-glucose cotransporter-2 (SGLT2) inhibitors have received special attention. In animal models with cancer exposed to PI3K inhibitors, SGLT2 inhibitors reduced glucose and insulin levels more than metformin and intriguingly were also associated with reduced cancer progression [11][12][13]. In humans, several clinical case reports describe impressive glycemic responses to SGLT2 inhibitors in patients on alpelisib [3,14,15]. In our prior exploratory analysis we found SGLT2 inhibitor use was associated with larger glucose reductions than other classes of antidiabetic drugs (e.g., metformin) [9]. However, SGLT2 inhibitors are not without risk. We and others have identi ed cases of DKA in patients receiving alpelisib and SGLT2 inhibitors [5,9,16]. SGLT2 inhibitor therapy increases the risk for euglycemic DKA, which could be exacerbated by alpelisib-induced insulin resistance. Therefore, we conducted this follow-up retrospective cohort study with two goals: to test the hypothesis that SGLT2 inhibitors reduce glucose more effectively than other classes of antidiabetic drugs in patients on alpelisib, and to describe the rate of DKA with and without SGLT2 inhibitor use in alpelisib-treated patients.

Research Design And Methods
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 ndings 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 identi ed by a database query for electronic prescriptions for alpelisib. Alpelisib use was con rmed by manual chart review by two reviewers (any of MAW, DL, RD, JF). Baseline comorbidities, height, weight, body mass index (BMI, in kg/m 2 ), 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 ltration 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 (de ned 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 de ned 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 strati ed 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 speci c 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 classi ed 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 bene t, 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 insu cient 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 signi cant 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 Cpeptide 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).
In the analysis of the effect of antihyperglycemics on mean random glucose levels, 77 patients (75%) were included (after exclusion of six patients (6%) on antidiabetic drugs at baseline and 20 (20%) without glucose measurements during the follow-up period). An average of seven glucose readings over mean follow-up of 166 days were analyzed. In unadjusted mixed linear modeling, eight of 77 patients (10%) had time with SGLT2 inhibitor exposure, which was associated with lower mean glucose levels compared to SGLT2 inhibitor unexposed time in adjusted linear analysis (-54 mg/dL [95% CI -99 to -8]). Time with metformin exposure, as contributed by 31 of 77 patients (40%) who received metformin, was also associated with lower mean glucose levels in adjusted linear analysis (-34 mg/dL [95% CI -60 to -7]) ( Fig. 1). Con dence intervals (CIs) for other antidiabetic drug classes were wide and crossed null in the setting of limited sample size: four patients (5%) received insulin, which had adjusted glucose reduction of -43 mg/dL (95% CI -111 to 28); two patients received a sulfonylurea (-4 mg/dL; 95% CI -94 to 92); two patients received a thiazolidinedione (-45 mg/dL; 95% CI -144 to 55), and two patients received a DPP4 inhibitor (-73 mg/dL; 95% CI -157 to 10). Corticosteroids were used in the treatment of 11 of 77 patients (14%) and were associated with signi cantly increased mean glucose values (+ 71 mg/dL; 95% CI 42 to 98). MSM showed an association between SGLT2 inhibitor exposure and glucose reduction but not for metformin (Fig. 1). MSM could not be carried out on other antidiabetic drugs because insu cient data were available to run the statistical models needed.
We identi ed ve cases of DKA (Table 2). Of these, two occurred in patients on concomitant SGLT2 inhibitor and alpelisib (with one patient also on metformin and one patient also on a sulfonylurea), two occurred in patients receiving concomitant metformin and alpelisib, and one occurred in a patient on no antidiabetic drugs.  In an exploratory analysis, a total of 49 random (i.e., non-fasting) C-peptide levels were measured in 23 patients (22%). Serum glucose levels were measured simultaneously for 48 of the 49 C-peptides ( Supplementary Fig. 1A). Twelve samples were collected when patients were not receiving alpelisib (median serum C-peptide level = 4.0 ng/mL; IQR 2.3,.1), 13 were measured during periods when patients received alpelisib without any antidiabetic drugs (median 9.9 ng/mL; IQR 4.1, 13.8), 14 were measured during periods when patients received concomitant alpelisib and antidiabetic drugs other than an SGLT2 inhibitor (median 10.6 ng/mL; IQR 7.7, 14.5), and 10 were measured during periods when patients received concomitant alpelisib and an SGLT2 inhibitor (4.6 ng/mL; IQR 4.0, 8.1). Unadjusted mixed linear modeling showed increased C-peptide levels in patients on alpelisib (P < 0.01) and alpelisib plus other antidiabetic drugs (P < 0.01), but not on alpelisib plus SGLT2 inhibitor (Supplementary Fig. 1B).

Conclusions
Findings from this retrospective review show that SGLT2 inhibitor use is consistently associated with substantial reductions in glucose levels in patients on alpelisib. However, the occurrence of two cases of DKA in 11 patients (18%) on an SGLT2 inhibitor is concerning. SGLT2 inhibitors may be a uniquely effective antidiabetic drug class in patients on alpelisib, but the risk of DKA must be considered and mitigated. These ndings may also be relevant to managing hyperglycemia in patients taking other drugs that inhibit the PI3K/AKT pathway.
The glucose-lowering effectiveness for SGLT2 inhibition in this study is consistent with previous animal data, clinical case reports, and exploratory observational analyses [9,12,14,15]. This study's strengths relative to previous work are that it is hypothesis testing rather than exploratory, controlled for important confounders (particularly steroid use), and includes sensitivity analysis using an MSM. This study was too small to establish whether SGLT2 inhibitors are more effective than other agents, such as metformin, in the setting of alpelisib use. However, SGLT2 inhibitors are now the only drug class with consistent evidence of glucose-lowering effectiveness in the context of patients receiving alpelisib.
Alpelisib increases C-peptide levels in patients, a marker of insulin production [20,21]. Our data suggest that SGLT2 inhibitors may abrogate this effect. This nding is consistent with preclinical data showing that SGLT2 inhibitors, but not metformin, could prevent high endogenous insulin levels in animal models on PI3K inhibitors [3]. Researchers have speculated that SGLT2 inhibitors might improve clinical response to PI3K inhibition through this mechanism, since high insulin levels might be reactive to PI3K inhibition and allow tumors to escape the anticancer effect. The exploratory ndings here support further investigation of that hypothesis.
These potential bene ts from SGLT2 inhibitor use with alpelisib need to be weighed against the potential for harm. The two cases of DKA observed in patients treated with a combination of alpelisib and an SGLT2 inhibitor imply a rate of approximately one case per four person-years of exposure, far higher than the DKA rates of approximately one case per 1,000 patient-years seen in patients with type 2 diabetes who take an SGLT2 inhibitor [22]. While two cases comprise a small sample, it is supported further by our previous study, which identi ed one unambiguous case of DKA in 15 patients treated with an SGLT2 inhibitor while using PI3K or AKT inhibitors [9]. Prescribers who choose to use an SGLT2 inhibitor with alpelisib should be aware of this potential risk.
Limitations of this analysis include that the results regarding effectiveness are limited by small sample size and potential for time varying confounding. For example, the decision to start any antidiabetic drug might be accompanied by lifestyle modi cations that are not captured in EMR data. Balancing the risks and bene ts of SGLT2 inhibitors in patients on alpelisib is challenging. Due to the risk of DKA, clinical practice at our institution is not to use SGLT2 inhibitors as rst-line agents, as many patients can achieve adequate glycemic control with metformin and lifestyle modi cation. However, SGLT2 inhibitors are our preferred second-line choice either in addition to or in place of metformin. We advise patients of the risk of DKA and that poor dietary intake is likely to increase the risk, so that they should hold the SGLT2 inhibitor on any day they are fasting or consuming less than 50% of their usual caloric intake. The value of home urine ketone monitoring is unclear because asymptomatic ketosis may be common on SGLT2 inhibitors and may not warrant intervention [23].
In summary, these study ndings can aid in treatment decisions when managing PI3K inhibitor-induced hyperglycemia. While these ndings apply to alpelisib speci cally, they likely generalize to other PI3Kα inhibitors, pan-PI3K inhibitors, and potentially to other inhibitors on the PI3K/AKT pathway. We recommend that metformin and lifestyle modi cation should be the rst-line treatment for alpelisib-

Data availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.