Baseline characteristics from both the X2101 (221 patients, data as of March 22, 2017) and SOLAR-1 trials (284 alpelisib plus fulvestrant patients, data as of June 12, 2018) were previously reported [12, 16, 17]. The median time to onset of grade ≥ 3 hyperglycemia (SOLAR-1 patients treated with alpelisib) was 15 days. Among the patients in the analysis from the SOLAR-1 trial, all-grade and grade 3/4 hyperglycemia (grouped terms) occurred in 187/284 (65.8%) and 108/284 (38.0%) patients, respectively (Fig. 1A and B). Of the patients from SOLAR-1 who developed hyperglycemia, 163/187 (87.2%) required a concomitant medication. One antihyperglycemic medication was required in 67/187 (35.8%) patients, while 49/187 (26.2%) and 47/187 (25.1%) required 2 and ≥ 3, respectively; metformin was the most common medication (142/163 [87.1%]), followed by insulin (52/163 [31.9%]). Plotting time to first event against time to first antihyperglycemic medication indicates that there were delays in treatment of hyperglycemia in some patients, even when hyperglycemia was identified early (Fig. 1C). A greater proportion of patients who received early intervention (first quantile for time from hyperglycemia to antihyperglycemic medication) for grade 1/2 hyperglycemia had improvement to a lower grade vs patients who received later intervention (fourth quantile) (Fig. 1D). Discontinuations due to hyperglycemia occurred in 19/284 patients (6.7%). Among patients in the BYLieve trial, metformin was given to 56/58 (96.6%) of patients who received an antihyperglycemic medication in Cohort A, 65/68 (95.6%) in Cohort B, and 61/66 (92.4%) in Cohort C. Insulin was given to 21/58 (36.2%) in Cohort A, 17/68 (25.0%) in Cohort B, and 18/66 (27.3%) in Cohort C.

Cumulative incidence of all-grade hyperglycemic events among patients in SOLAR-1 treated with alpelisib + fulvestrant (A) or grade 3/4 hyperglycemic events among patients in SOLAR-1 treated with alpelisib + fulvestrant (B). Time to first medication vs first hyperglycemic event in patients treated with alpelisib from SOLAR-1 who had a hyperglycemia adverse event and was treated by antihyperglycemic medication (C) and subsequent grade of hyperglycemia among patients in SOLAR-1 with early (first quantile of time from hyperglycemia to treatment) and late (fourth quantile) treatment with antihyperglycemic medication (D). Cumulative incidence curves using Kaplan–Meier method. Hyperglycemic events by Standardized MedDRA query. AE, adverse event; G, grade

Model development using the training set was conducted to better identify patients at high risk for grade 3/4 hyperglycemic events. From a set of 36 variables, univariate analysis identified 13 variables for model building, defined by an adjusted P < 0.1 (Table 1). According to β coefficient, the most influential, significant parameters for predicting grade 3/4 hyperglycemia in the univariate analysis were FPG and body mass index (BMI); these were also the most important parameters identified by the elastic net model and linear model with stepwise variable selection. Based on the variables identified in the univariate analysis, 6 models were investigated for classification of risk of a grade 3/4 hyperglycemia event (Additional file 1: Figure S2). The area under the curve (AUC) over time in the training and test sets was used to compare the accuracy of the model scores at distinguishing patients who will develop from those who are not likely to develop a grade 3/4 hyperglycemia event. Model 4, the random forest model, performed notably better than the other models.

For developing a clinically relevant simplified random forest model, a conditional permutation importance analysis was performed to identify the relative importance of the variables in model 4 (Fig. 2A). Model 7, the simplified random forest, was generated using only the 5 most influential variables from model 4: FPG, BMI, HbA1c, monocytes, and age. The AUC over time was similar for models 4 and 7 in the training and test sets (Additional file 1: Figure S2). At 2 months, the AUCs for models 4 and 7 were similar in the training set (0.994 and 0.991, respectively) and the test set (0.729 and 0.767, respectively). Patients were classified into high- or low-risk groups for grade 3/4 hyperglycemia based on the predictive risk score using model 7. In the training set, there was clear classification between the low- and high-risk groups for the probability of grade 3/4 hyperglycemia events (Fig. 2B). In the high-risk group, there was an 86.2% risk of a grade 3/4 event by 2 months’ treatment; there was a < 5% probability of an event in the low-risk patients over 30 months (Fig. 2B, Additional file 1: Table S1). In the test set, the high-risk group had approximately 57.6% chance of a grade 3/4 hyperglycemia event by month 2, while the low-risk patients had < 20% probability of developing grade 3/4 hyperglycemia over 30 months (Fig. 2C, Additional file 1: Table S1). The median (95% CI) time to grade 3/4 hyperglycemia events in the high-risk group was 15 (13–15) days in the training set and 20 (15-NE) days in the test set (Additional file 1: Table S1).

Variable importance of random forest (model 4) in the X2101 + SOLAR-1 training set (A) and prognostic association of risk status based on a random forest model with fasting plasma glucose, BMI, HbA1c, monocytes, and age (model 7) with time to grade 3/4 hyperglycemia event in the X2101 + SOLAR-1 training set (B) and the X2101 + SOLAR-1 test set (C). A was calculated by conditional permutation importance, B was calculated by cumulative incidence curves, and C was calculated using the Kaplan–Meier method. bid, twice daily dosing; BMI, body mass index; DBP, diastolic blood pressure; FPG, fasting plasma glucose; HDL, high-density lipoprotein cholesterol; RBC, red blood cells; SBP, systolic blood pressure.

Using model 7, individual scores could be derived for a patient at baseline in order to determine the corresponding risk of a hyperglycemia event at a given time (Additional file 1: Figure S3). The density of individual scores for patients in the training set showed a bimodal distribution of high- and low-risk groups. In the pooled studies (n = 505), compared with low-risk patients, high-risk patients at baseline were more likely to be obese (BMI ≥ 30 kg/m2; 37.2% high risk vs. 15.0% low risk, respectively), have higher FPG (≥ 5.6 mmol/L; 71.3% vs. 22.6%), and have higher HbA1c (≥ 5.7%; 56.7% vs. 27.6%) (Additional file 1: Table S2).

Applying model 7 to the alpelisib arm of SOLAR-1 resulted in 106 patients in the high- and 178 in the low-risk groups. All-grade hyperglycemia occurred in 101/106 (95.3%) high- and 86/178 (48.3%) low-risk patients; grade 3/4 events occurred in 96/106 (90.6%) and 12/178 (6.7%), respectively. Antihyperglycemic medication was used in 94/106 (88.7%) high- and 70/178 (39.3%) low-risk patients. The majority of high-risk patients who received an antihyperglycemic medication required ≥ 3 (39/94 [41.5%]), followed by 2 (31/94 [33.0%]), and 1 medication (24/94 [25.5%]). Conversely, the majority of low-risk patients only required 1 antihyperglycemic medication (43/70 [61.4%]), followed by 2 (18/70 [25.7%]) and ≥ 3 medications (9/70 [12.9%]).

Alpelisib discontinuations were similar overall (90/106 [84.9%] vs. 154/178 [86.5%] for high vs low); however, discontinuation due to AE was more common in the high-risk (32/90 [35.6%]) vs. low-risk (39/154 [25.3%]) group, while discontinuation due to progressive disease was more common in the low-risk group (46/90 [51.1%] vs. 91/154 [59.1%] for high vs. low). Among patients who discontinued alpelisib, there were more discontinuations due to hyperglycemia in the high- vs. low-risk group (15/90 [16.7%]) vs. 4/154 [2.6%]). Dose reductions (85/106 [80.2%] vs. 83/178 [46.6%]) and interruptions (89/106 [84.0%] vs. 116/178 [65.2%]) were more common in the high- vs low-risk groups, respectively. Median relative dose intensity for high- vs. low-risk group was 75% vs. 93%, with median dose intensities of 223.9 vs. 278.8 mg/day, respectively.

For external validation, model 7 was applied to all patients in BYLieve, resulting in 103 vs. 237 patients in the high- vs low-risk categories. Consistent with the results observed in SOLAR-1, a significant difference was observed in time to grade 3/4 hyperglycemia favoring low-risk patients (HR, 0.367; 95% CI, 0.241–0.558; P < 0.0001) (Fig. 3).

Time to grade 3/4 hyperglycemia in patients classified as high or low risk by model 7 in BYLieve

Efficacy analysis among patients in SOLAR-1 with a PIK3CA mutation who were treated with alpelisib + fulvestrant showed Kaplan–Meier PFS curves with a slight separation prior to 6 months, after which the curves appeared to converge for the remaining time points. The median PFS was similar between high- (11.0 months) and low-risk patients (10.9 months; HR, 1.097; 95% CI, 0.733–1.641) (Fig. 4).

Progression-free survival for patients with PIK3CA mutations with high and low risk of hyperglycemia (model 7) treated with alpelisib in SOLAR-1