Risk Factors of Hyperglycemia After Treatment With the AKT Inhibitor Ipatasertib in the Prostate Cancer Setting: A Machine Learning–Based Investigation

PURPOSE

Hyperglycemia is a major adverse event of phosphatidylinositol 3-kinase/AKT inhibitor class of cancer therapeutics. Machine learning (ML) methodologies can identify and highlight how explanatory variables affect hyperglycemia risk.

METHODS

Using data from clinical trials of the AKT inhibitor ipatasertib (IPAT) in the metastatic castrate-resistant prostate cancer setting, we trained an XGBoost ML model to predict the incidence of grade ≥2 hyperglycemia (HGLY ≥ 2). Of the 1,364 patients included in our analysis, 19.4% (n = 265) of patients had HGLY ≥2 events with a median time of first onset of 28 days (range, 0-753 days), and 30.0% (n = 221) of patients on an IPAT regimen had at least one HGLY ≥2 event compared with 7.0% (n = 44) of patients on placebo.

RESULTS

An 11-variable XGBoost model predicted HGLY ≥2 events well with an AUROC of 0.83 ± 0.02 (mean ± standard deviation). Using SHapley Additive exPlanations analysis, we found IPAT exposure and baseline HbA1c levels to be the strongest predictors of HGLY ≥2, with additional predictivity of baseline measurements of fasting glucose, magnesium, and high-density lipoproteins.

CONCLUSION

The findings support using patients' prediabetic status as a key factor for hyperglycemia monitoring and/or trial exclusion criteria. Additionally, the model and relationships between explanatory variables and HGLY ≥2 described herein can help identify patients at high risk for hyperglycemia and develop rational risk mitigation strategies.

INTRODUCTION

Ipatasertib (IPAT) is an investigational small molecule cancer therapeutic that potently inhibits all three isoforms of the serine/threonine kinase AKT. AKT is the central node of the phosphatidylinositol 3-kinase (PI3K)/AKT/mammalian target of rapamycin (mTOR) pathway, whose normal activation is associated with cell cycle progression, proliferation, and survival. The aberrant activation of this pathway is involved in tumorigenesis, which is prevalent in many malignancies by numerous genetic and nongenetic mechanisms.¹ As such, many cancer therapeutics are directed to inhibit this pathway. Although PI3k/AKT/mTOR inhibitors can have antitumorigenic effects, these classes of therapies are associated with various adverse effects,² most notably hyperglycemia.^3-5 This is because the PI3K/AKT pathway also plays a key role in glucose homeostasis and mediates the metabolic effects of insulin downstream of the insulin receptor (a receptor tyrosine kinase).^6-9 Because hyperglycemia is an on-target toxicity of these drugs, it is often managed through dose reductions that may limit therapeutic efficacy.²

CONTEXT

• Key Objective

• To characterize the risk factors of hyperglycemia after treatment with the AKT inhibitor ipatasertib (IPAT) in a population with prostate cancer.

• Knowledge Generated

• Hyperglycemia incidences could be well predicted with IPAT exposure and baseline HbA1c being key predictive factors. We additionally identified that baseline fasting glucose, magnesium, and high-density lipoproteins can further stratify hyperglycemia risk.

• Relevance

• As hyperglycemia is a major toxicity associated with phosphatidylinositol 3-kinase/AKT inhibitors that are utilized and being evaluated in oncology settings, the identified risk factors described in this work can be utilized to identify patients at high risk for hyperglycemia as well as develop rational risk mitigation strategies. The findings support using patients' diabetic status as a key factor for hyperglycemia monitoring and/or trial exclusion criteria.

Increased incidences of hyperglycemia have been observed for IPAT in combination with prednisone/prednisolone and abiraterone compared with placebo in the phase III IPATential150 study.¹⁰ A clinical case-control study found that higher incidences of grade 3-4 hyperglycemia were observed in the PI3K/AKT/mTOR inhibitor group compared with the control group in phase I clinical trials.⁴ Another review of phase I clinical trials of PI3K/AKT/mTOR showed that of 341 patients, 81 patients (23.8%) developed grade ≥2 hyperglycemia and 20 (6.7%) patients developed grade ≥3 hyperglycemia.⁵ Furthermore, analysis of the clinical data using statistical testing showed that the patient age and history of diabetes were associated with the occurrence of grade 3 hyperglycemia.⁵ However, to our knowledge, to date, there has not been a comprehensive analysis of factors contributing to hyperglycemia in this context using advanced machine learning (ML) algorithms. Compared with linear statistical models, the use of ML algorithms has the potential to identify nonlinear relationships between explanatory variables and patients' outcomes.¹¹ In this work, we applied the ML algorithm XGBoost to build a predictive model of hyperglycemia and identified the associated factors using the model explainability framework of SHapley Additive exPlanations (SHAP) analysis.¹²

METHODS

Design, Setting, and Participants

In this post hoc analysis, we sought to characterize the relationship between explanatory variables and hyperglycemia risk in the IPAT studies. This characterization was done by generating XGBoost models to predict the incidence of grade ≥2 hyperglycemia, which is defined as fasting glucose levels >160 mg/dL. SHAP analysis was used to explain how the ML model used explanatory variables to predict HGLY ≥2 probability as detailed below (see SHAP analysis). This analysis was performed using clinical trial data from the phase Ib/II/III A. MARTIN and IPATential150 studies^10,13 evaluating the safety and efficacy of IPAT with abiraterone + prednisone/prednisolone versus placebo with abiraterone + prednisone/prednisolone in the metastatic castrate–resistant prostate cancer (mCRPC) setting. The A. MARTIN (ClinicalTrials.gov identifier: NCT01485861) and IPATential150 (ClinicalTrials.gov identifier: NCT03072238) studies were conducted according to Good Clinical Practice guidelines of the International Conference on Harmonisation and the principles of the Declaration of Helsinki and were approved by the independent review boards or ethics committees at all study sites. All participants provided written informed consent before enrollment in the trial where all patients consented for their data to be used for research purposes to advance science and public health. The phase II trial was conducted on patients previously treated with docetaxel-based chemotherapy, whereas the phase III trial was conducted on asymptomatic or mildly symptomatic, previously untreated patients. Patients with controlled diabetes were eligible for the trials and included in this analysis, but because of the risk of hyperglycemia, patients requiring insulin or had fasting glucose ≥150 mg/dL at baseline were excluded from the trials. Patients on stable diabetes medications for ≥4 weeks before initiation were eligible for the phase III study. The initial data set consisted of data from 1,364 male patients age ≥18 years, with 42 features related to dose/exposure, demographics, physiology, biomarkers, and disease history. Data points that are >10 standard deviations from the mean were removed from the data set. Overall, 5.1% of values were missing from the full data set (Appendix Table A1; Appendix Fig A1).

Selection of Drug Exposure Metric

Cmin, Cmax, and AUC after a single dose (sd) of IPAT were derived for patients on the basis of a previously developed population pharmacokinetic model.¹⁴ Inclusion of placebo patients in this analysis helps to provide a precise estimate of the impact of 0 exposure and thereby facilitates the IPAT exposure-response analysis on HGLY ≥2 events. For our analyses, we only considered the exposure-related variable (ie, IPAT dose [IPADOSE], Cminsd, Cmaxsd, and AUCsd) that had the greatest predictive utility in terms of SHAP-based feature importance. This was done by training XGBoost models on all patient data and all explanatory variables to predict HGLY ≥2 (Appendix Fig A2). The models were tuned using 25 iterations of a Bayesian hyperparameter search using the hyperopt package¹⁷ in the search space defined in Appendix Table A2. SHAP values were calculated on out-of-sample data using 20-fold cross-validation (see SHAP analysis). This preliminary analysis suggested that Cminsd was the exposure-related metric that had the greatest utility in predicting HGLY ≥2 incidents; thus, Cmaxsd, AUCsd, and IPADOSE were excluded from subsequent analyses. The 39 remaining explanatory variables yielded a robust HGLY ≥2 predictive performance of 0.82 ± 0.02 in terms of AUROC (±standard deviation [SD]) on the basis of five-fold cross-validation repeated 5 times (see Model performance and reliability).

Feature Selection

Although ML algorithms such as XGBoost generally scale well with a large number of explanatory variables, eliminating noninformative variables can improve model performance¹⁵ and interpretability. We used a custom implementation of a SHAP-based recursive feature elimination (SHAP RFECV) to identify a parsimonious set of explanatory variables to predict HGLY ≥2. In this method, hyperparameters were tuned only once initially using 25 iterations of a Bayesian hyperparameter search in the search space defined in Appendix Table A2. SHAP values were calculated on out-of-sample data using 20-fold cross-validation (see SHAP analysis), and the same data splits were used to calculate SHAP values in the remaining feature elimination steps to eliminate variation in feature importance on the basis of data splitting. In each step, the least important explanatory variable in terms of mean absolute SHAP value was removed, model performance in terms of AUROC was evaluated using 5-fold cross-validation, and SHAP values were recalculated using models trained on the remaining explanatory variables. This process was repeated until no explanatory variables remained. Each feature elimination run yielded a selected set of explanatory variables that yielded maximal performance in terms of mean AUROC. The process was repeated 10 times with splits of the data varying between runs to assess model performance (Appendix Fig A3). The 11 explanatory variables that represent the union of selected features across runs was used for the final model, which were Cminsd, BHBA1C, BGLUC, BMG, BHDL, BCREAT, BAMYLASE, BALP, BPLAT, BAST, and BTRIG (see Appendix Table A1 for definitions).

Hyperparameter Tuning

Hyperparameter tuning usually improves model performance, but it can lead to overfitting. To prevent overfitting, we used a nested cross-validation strategy to ensure that selected hyperparameters generalized well onto unseen test data.¹⁶ In brief, we generated 25 random train-test splits into 80:20 proportions and used Bayesian hyperparameter search to optimize AUROC using 5-fold cross-validation on training sets¹⁷ in the search space defined in Appendix Table A2. We found that hyperparameters tuned on the training sets generalized well onto test sets (Appendix Fig A4). This suggests that overfitting was not a concern for our hyperparameter tuning procedure. We used the mean XGBoost hyperparameters associated with max_depth = 2 for the final model which were max_depth: 2, eta: 0.10, min_child_weight: 0.28, reg_alpha: 2.55, reg_lambda: 2.52, and subsample: 0.68.

Model Performance and Reliability

To estimate model performance, we reported the mean ± SD of AUROC measurements on validation sets from 5-fold cross-validation repeated five times.

SHAP Analysis

SHAP is a game theoretic framework that deconvolves the impact of explanatory variables (SHAP values) on ML predictions. We extracted SHAP values in a probability scale using the SHAP package.¹² In the SHAP formulation, the sum of variable impacts (SHAP values) plus an expected value equals the model prediction (Eq 1):

$$\widehat{f}\left(x\right)={\mathrm{\varphi }}_0+{\displaystyle \sum _{p\in S}{\mathrm{\varphi }}_p}$$ (1)

Here, fˆ (x) represents a model's prediction for a single instance x. ϕ₀ is the expected value, which can be thought of as an a priori best guess of HGLY ≥2 probability. S represents the set of model explanatory variables, and ϕ_p is the SHAP value for the explanatory variable p (ie, how much p changes the probability of HGLY ≥2). To obtain unbiased SHAP values, SHAP values were calculated for out-of-sample data when models were trained using a leave-one-out approach. The leave-one-out approach is an extreme form of cross-validation, where models are trained for each patient using all of the other patient data to maximize the use of available data to draw SHAP-based inferences.

Bootstrapped SHAP Analysis

The uncertainty in SHAP values was estimated using bootstrap analysis. Models were trained on data using sample-with-replacement with the bootstrap samples being the same size as the original data set. SHAP values were generated for out-of-sample data. Because the expected probability of an instance being out-of-sample is 1/e, we ran 1,000 × e (ie, 2,783) bootstrap iterations to generate approximately 1,000 models to calculate 95% CIs for SHAP value estimates.

RESULTS

A total of 1,364 patients with mCRPC from the A. MARTIN and IPATential150 clinical trials were included in the analyses in this report. All arms consisted of 1,000 mg once daily abiraterone + 5 mg twice a day prednisone/prednisolone treatment regimen, in conjunction with daily orally administered placebo (n = 628), 200 mg IPAT (n = 87), or 400 mg IPAT (n = 649). Overall, there was a 19.4% rate of HGLY ≥2 in this population (95% CI, 17.4 to 21.6).

Model Predictive Performance and Reliability

After feature selection and hyperparameter tuning (see the respective section under Methods), we developed an 11-variable XGBoost model using all placebo and treated patients' data to predict the incidence of HGLY ≥2. Model performance in terms of AUROC was 0.83 ± 0.02 (mean ± SD) using 5-fold cross-validation repeated five times.

In addition to model performance, it is beneficial to know if model predictions are reliable: that is, do the empirical HGLY ≥2 rates corroborate with the model predicted probabilities? To address this question, we estimated each patient's probability of HGLY ≥2 (Fig 1A) using a Leave One Out Cross-Validation (LOOCV) approach to maximize the use of available data, whereby a model was trained for each patient using data from all other patients to generate predictions.¹⁸ We found that the mean of the binned predictions correspond well to the empirical HGLY ≥2 rates on the vertical axis in Figure 1A, with the empirical HGLY ≥2 rates falling within the 95% binomial CI associated with the binned predictions.

FIG 1.

Model reliability and SHAP summary diagram. (A) Distribution of model predictions. Predictions were binned into 10 evenly spaced bins, except the first and last bin, each contained 5% of the data. (B) Prediction reliability diagram showing empirical HGLY ≥2 rates on the vertical axis versus predicted HGLY ≥2 probability on the horizontal axis. The vertical bars represent the intervals encompassing the middle 95% binomial distribution associated with the binned predicted probabilities, which serve as a range one would expect empirical rates if predictions were generated from a reliable model. (C) SHAP-based feature importance ranking from highest to lowest. (D) SHAP summary diagram demonstrating how explanatory variable values (color) contribute to ∆ predicted probability of HGLY ≥2 (position along the horizontal axis). For each explanatory variable, an individual point represents an individual patient. See Appendix Table A1 for variable definitions. SHAP, SHapley Additive exPlanations.

Summary of SHAP Analysis

Having demonstrated that HGLY ≥2 predictions are reliable, we next generated model-based estimates of explanatory variable impacts on predicted HGLY ≥2 probability using SHAP analysis. SHAP analysis deconvolves how explanatory variables contribute to model predictions. In the SHAP formulation, SHAP values represent explanatory variable impacts on model predictions relative to an expected value, which can be thought of as an a priori best guess of HGLY ≥2 probability on the basis of all (placebo and treated patients') data. We calculated SHAP values on a probability scale using LOOCV, which are summarized in Figures 1C and 1D.

The summary plots in Figure 1 allow for general characterizations of relative importance (Fig 1C) and the directionality of explanatory variable impacts (Fig 1D). The explanatory variables are shown in the order of importance (high to low) on the y-axis, where importance is defined as the mean absolute SHAP value associated with each explanatory variable. For each explanatory variable in Figure 1D, the variable impact (ie, SHAP value) is represented by the position along the x-axis, and the variable value is represented by the color. The combination allows one to see how variable values are associated with their impact. Overlapping points are jittered vertically, so the distribution of SHAP values is seen clearly. The most important explanatory variable for predicting HGLY ≥2 was Cminsd. As Cminsd value increased, indicated by the blue-to-red gradient, there was an increasingly positive impact on the predicted probability of HGLY ≥2 indicated by the position along the horizontal axis. Similarly, high HbA1c (BHBA1C) and fasting glucose (BGLUC) were predictive of greater HGYL ≥2 probability. In contrast, low baseline magnesium (BMG) and high-density lipoproteins (BHDLs) were predictive of greater HGYL ≥2 probability.

Functional Relationships Between Explanatory Variables and HGLY ≥2 Probability

To examine the estimated functional relationship between explanatory variables and HGLY ≥2 probability, one can plot SHAP values versus the corresponding variable values in what are known as feature-dependence plots.¹² Standard feature-dependence plots do not capture the uncertainty in the estimated explanatory variable impacts that may arise from variability in the sampled data. We used a sample-with-replacement bootstrap approach to estimate the uncertainty in the SHAP values (see methods for more details). Using this approach, we examined the IPAT exposure-response relationship for HGLY ≥2 in Figure 2. In these plots, each point corresponds to a single patient, with the error bars representing the 95% CIs on the basis of bootstrap analysis.

FIG 2.

Exposure-response analysis for HGLY ≥2. (A) Each point represents the mean marginal contribution (SHAP value) of Cminsd on predicted HGLY ≥2 probability for a single patient across bootstrap iterations. The colors represent the patients' actual exposure (placebo or different tertiles of exposure). (B) Grouped patient data. Error bars represent 95% CIs across bootstrap iterations. See Appendix Table A1 for variable definitions. SHAP, SHapley Additive exPlanations.

Bootstrapped feature-dependence plots for four key baseline explanatory variables are demonstrated in Figure 3. BHBA1C was an important predictor of HGLY ≥2, consistent with previous investigations on the relationship between HbA1c and hyperglycemia in the context of PI3K-AKT-mTOR inhibitors in oncology populations.⁵ We examined the bootstrapped feature dependence plots for BHBA1C binned into standardly defined normal, prediabetic, and diabetic ranges (≤5.7%, 5.7%-6.4%, and >6.4%, respectively; Fig 3A).¹⁹ BHBA1C exhibits a nonlinear functional relationship with predicted HGLY ≥2 probability, where BHBA1C starts increasing HGLY ≥2 probability in the prediabetic range. BHBA1C in the normal range (≤5.7%) decreased predicted HGLY ≥2 probability by 4.4% (95% CI, –5.8 to –3.1), whereas BHBA1C values in the diabetic range (>6.4%) increased predicted HGLY ≥2 probability by 18.4% (95% CI, 12.72 to 24.34) relative to expected value.

FIG 3.

Impact of key baseline explanatory variables on predicted HGLY ≥2 probability. Bootstrapped feature-dependence plots for (A) BHBA1C, (B) BGLUC, (C) BMG, and (D) BHDL. The colors represent how data points were binned; BHBA1C was binned into normal, prediabetic, and diabetic ranges (≤5.7%, 5.7%-6.4%, and >6.4%, respectively), and other explanatory variables were binned into quartiles. Error bars represent 95% CI estimates across bootstrap iterations for individuals. See Appendix Table A1 for variable definitions.`

We found that BGLUC, BMG, and BHDLs were also important predictors of HGLY ≥2. Increased BGLUC was predictive of increased HGLY ≥2 probability (Fig 3B). In contrast, levels of BMG and BHDLs were inversely predictive of HGLY ≥2 probability (Figs 3C and 3D, respectively). The functional relationship of all 11 explanatory variables with HGLY ≥2 probability and empirical relationships with HGLY ≥2 rate were evaluated (Appendix Fig A5 and A6), but only explanatory variables that had significant impacts on HGLY ≥2 probability are demonstrated in Figure 3 and summarized in Figure 4. Here, significance was defined as 95% CIs for estimated impacts that did not overlap across subgroups defined by the lowest and highest quartiles of explanatory variable levels.

FIG 4.

Summary of model-estimated impacts on predicted HGLY ≥2 probability. The mean estimated impacts on HGLY ≥2 probability are shown for important explanatory variables. Impacts are on a probability scale (horizontal axis) relative to an expected probability. Error bars represent the 95% CI across bootstrap iterations for the patient subsets delineated on the vertical axis. See Appendix Table A1 for variable definitions

DISCUSSION

Hyperglycemia is an on-target toxicity associated with PI3K/AKT inhibitors. Characterization of the risk factors associated with hyperglycemia in this context can be helpful for identification of high-risk patients and for the development of risk mitigation strategies. In this work, we used data from the phase Ib/II A. MARTIN and phase III IPATential150 studies evaluating the safety and efficacy of IPAT in combination with abiraterone + prednisone/prednisolone in the mCRPC population.¹⁰ As all patients in these studies had a background treatment of abiraterone + prednisone/prednisolone, we were unable to examine the potential impacts of these background treatments in this study. Although we note that high doses of prednisone (eg, 25 mg once daily) are known to promote gluconeogenesis and increase hyperglycemia risk,²⁰ a recent investigation demonstrated no change in glucose dynamics when IPAT was administered as a monotherapy vs. when administered in combination with a 5-mg prednisone dosing regimen.²¹ In contrast, abiraterone does not affect glucose metabolism appreciably.²² However, concomitant administration of abiraterone is known to increase IPAT exposure by decreasing clearance,¹⁴ and IPAT in combination with abiraterone + prednisone was shown to increase peak glucose and average glucose dynamics compared with IPAT monotherapy.²¹ Thus, although prednisone/prednisolone may have had limited impact on overall HGLY ≥2 rates in these studies, the drug-drug interaction with abiraterone may have contributed to increased IPAT exposures and consequently increased hyperglycemia rates.

Using data from these studies, we developed a predictive model of HGLY ≥2 using XGBoost. The SHAP explainability framework was then used to characterize the relationships between explanatory variables and HGLY ≥2, which may be relevant for other PI3K/AKT inhibitors. After feature elimination, 11 variables were identified as being potentially important for predicting HGLY ≥2. All the identified variables may not necessarily have had a strong impact on HGLY ≥2 because we wanted to be inclusive of variables that had even minor impacts on predicting HGLY ≥2. This aligned with our goal of performing a comprehensive characterization of explanatory variable relationships with HGLY ≥2 as well as ensuring confounding factors were accounted for in these characterizations.²³ For instance, the explanatory variables BPLAT and BAMYLASE did not have significant impacts on predicting HGLY ≥2, but they were selected after feature elimination perhaps because of their potential associations with type 2 diabetes^24,25 or interactions with glucose metabolism.²⁶ Because of their minor and statistically insignificant importance, we urge caution in interpreting these findings. However, we identified that Cminsd, BHBA1C, BGLUC, BMG, and BHDLs were key explanatory variables for predicting HGLY ≥2. The importance of BHBA1C and BGLUC in predicting HGLY ≥2 was to be expected as these are well-established risk factors for hyperglycemia.⁵ In contrast, the importance of BMG and BHDLs was less expected. To assess the biological plausibility that true relationships exist between these explanatory variables and hyperglycemia, we examined the totality of evidence surrounding the associations between MG²⁺ and HDL and hyperglycemia with the aid of the Causaly platform.²⁷

Our ML analysis suggests that low Mg²⁺ and HDL levels are associated with increased risk for HGLY ≥2 which is consistent with the literature. Many studies have demonstrated that magnesium (Mg²⁺) levels are generally reduced in individuals with type 2 diabetes^28,29 and reduced Mg²⁺ levels have been shown to be predictive of the onset of type 2 diabetes.³⁰ Although magnesium serves as a cofactor and activator for numerous enzymatic reactions, it has critical roles in insulin secretion and signaling which may predispose individuals with magnesium deficiency to hyperglycemia.³¹ Various studies have demonstrated that magnesium supplementation improves insulin sensitivity and glycemic control.³¹ Similar to Mg²⁺, several studies have found that low HDL levels are predictive of the onset of type 2 diabetes.^32-34

Interventions that elevate HDL levels have been shown to improve glycemic control in patients with type 2 diabetes.^35,36 A potential mechanism by which HDL may promote euglycemia is by increasing insulin-dependent and insulin-independent cellular uptake of glucose.^35,37

In this report, we assessed the risk factors of HGLY ≥2 in the context of IPAT treatment in the context of mCRPC using baseline explanatory variables and early IPAT exposure data from all three phases of international clinical trials. This allowed us to generate a robust model to have a general understanding of risk factors for IPAT-induced hyperglycemia. The utility and relative impacts of the identified risk factors remain to be further evaluated in the context of other conditions, other PI3K/AKT inhibitor-based interventions, and on a mixed-gender population.

Nevertheless, this work highlights that advances in ML and explainability are promising methodologies to identify risk factors of adverse events and further refine and personalize treatments.

In conclusion, hyperglycemia risk is a known, on-target adverse event for PI3K/AKT inhibitor therapies in the oncology setting. We assessed the factors associated with the incidence of grade ≥2 hyperglycemia using ML on a clinical trial data set examining the safety/efficacy of the small molecule AKT inhibitor IPAT in the mCRPC population. HGLY ≥2 incidences could be predicted well with an AUROC of 0.83 ± 0.02 using an 11-variable XGBoost model. We identified IPAT exposure, BHBA1C, BGLUC, BMG, and BHDLs as the key explanatory variables for predicting HGLY ≥2 incidents. The highlighted risk factors in this work may be helpful for identifying patients at high risk for HGLY ≥2 in the context of PI3K/AKT inhibitor therapies and the development of rational hyperglycemia risk mitigation strategies.

APPENDIX 1

TABLE A1.

Variable Definitions

TABLE A2.

Hyperparameter Search Space for XGBoost Models

FIG A1.

Pattern of missingness in the initial data set. Each row represents a patient (n = 1,364) and columns represent the variables explored in this work (see Appendix Table A1 for definitions). The dark bands represent the presence of available data; conversely, the light bands indicate missing data. The patient data are sorted by the clinical trial phase in which they were enrolled (phase 1: blue, phase 2: orange, and phase 3: green shadings).

FIG A2.

SHAP analysis using the full initial data set (ie, 42 features including Cminsd, Cmaxsd, AUCsd, and IPADOSE). Cminsd was the variable related to drug exposure that was most predictive of HGLY ≥2 in this population; thus, Cminsd was used in subsequent analyses. SHAP values were calculated on out-of-sample data using 20-fold cross-validation. See Appendix Table A1 for variable definitions. CV, cross-validation; IPADOSE, IPAT dose; SHAP, SHapley Additive exPlanations.

FIG A3.

Feature elimination. SHAP RFECV was used for backward feature elimination with a step size of 1. In each step, the order of variable importance on the basis of SHAP analysis was recalculated. The explanatory variables that maximized model performance in terms of AUROC were selected in each run. The union of the selected features across 10 runs was used for the final model. CV, cross-validation; SHAP, SHapley Additive exPlanations; SHAP RFECV, SHAP-based recursive feature elimination.

FIG A4.

Assessment of the generalizability of hyperparameters for the 11-variable model. Twenty-five random train-test splits in an 80:20 proportion of the data were used to tune hyperparameters on the training data and evaluate model performance on the test data. The boxplot demonstrates the performance on validation data (blue), which represents 125 AUROC calculations resulting from 5-fold cross-validation on the 25 training sets. Performance was also assessed on the 25 holdout sets (red) that were not incorporated in the hyperparameter tuning. The similar AUROC distributions suggest that our hyperparameter tuning method does not result in overfitting.

FIG A5.

Associations between explanatory variables and empirical HGLY ≥2 rates. Empirical HGY ≥2 rate plots are stratified by 0 or 400‐mg IPAT dose, with the exception of the plot for Cminsd. Because of the limited data on patients on 200 mg once daily (n = 87), rates are not shown for patients on 200‐mg once-daily IPAT. IPAT, ipatasertib.

FIG A6.

Model-estimated explanatory variable impacts on HGLY ≥2 probability. Blue-to-red gradients for mean ∆ prediction correspond to the magnitude of mean explanatory variable impacts relative to the expected value. We found several explanatory variables contributed to increased HGLY ≥2 probability as highlighted in Figure 3. Baseline glucose (BGLUC) above approximately 115 mg/dL contributed to increased predicted HGLY ≥2 probability by 7.3% (95% CI, 4.0 to 11.0). Magnesium (BMG) levels below approximately 2 mg/dL contributed to increased predicted HGLY ≥2 probability by 3.8% (95% CI, 1.3 to 6.7). High-density lipoproteins (BHDLs) below approximately 41.9 mg/dL increased predicted HGLY ≥2 probability by 2.9% (95% CI, 0.7 to 5.7).

AUTHOR CONTRIBUTIONS

Conception and design: Rashed Harun, Rucha Sane, Jin Y. Jin, James Lu

Financial support: Jin Y. Jin

Administrative support: Jin Y. Jin

Collection and assembly of data: Rashed Harun, Dhruvitkumar S. Sutaria

Data analysis and interpretation: Rashed Harun, Kenta Yoshida, James Lu

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/cci/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).