Model Development of CDK4/6 Predicted Efficacy in Patients With Hormone Receptor–Positive, Human Epidermal Growth Factor Receptor 2–Negative Advanced or Metastatic Breast Cancer

Abstract

PURPOSE

Three cyclin-dependent kinase 4/6 inhibitors (CDKIs) are approved by the US Food and Drug Administration for the treatment of patients with hormone receptor–positive, human epidermal growth factor receptor 2–negative advanced or metastatic breast cancer in combination with hormonal therapy (HT). We hypothesized that on an individual basis, efficacy outcomes and adverse event (AE) development can be predicted using baseline patient and tumor characteristics.

METHODS

Individual-level data from seven randomized controlled trials submitted to the US Food and Drug Administration for new or supplemental marketing applications of CDKIs were pooled. Progression-free survival (PFS), overall survival (OS), and AE prediction models were developed for specific treatment regimens (HT v HT plus CDKI). An individual's characteristics were used in all models simultaneously to create a group of predicted outcomes that are comparable across treatment settings.

RESULTS

Accuracy of the PFS and OS prediction models for HT were 66% and 64%, respectively, with the strongest predictors being menopausal status and therapy line. The corresponding AE prediction models resulted in an average area under the curve of 0.613. Accuracy of the PFS and OS prediction models for HT plus CDKI were 62% and 63%, respectively, with the strongest predictors being histologic grade for both. The corresponding AE prediction models resulted in an average area under the curve of 0.639.

CONCLUSION

This exploratory analysis demonstrated that models of efficacy outcomes and AE development can be developed using baseline patient and tumor characteristics. Comparison of paired models can inform treatment selection for individuals on the basis of the patient's personalized goals and concerns. Although use of CDKIs is standard of care in the first- or second-line setting, this model provides prognostic information that may inform individual treatment decisions.

INTRODUCTION

There are currently three cyclin-dependent kinase 4/6 inhibitors (CDKIs) approved by the US Food and Drug Administration (FDA) for the treatment of patients with hormone receptor–positive, human epidermal growth factor receptor 2 (HER2)–negative, advanced or metastatic breast cancer (MBC) in combination with hormonal therapy (HT) in both the first- and second-line setting. Palbociclib received accelerated approval on February 3, 2015, followed by regular approval on March 31, 2017.^1-5 Ribociclib was first approved on March 13, 2017.^1,6-8 Abemaciclib was first approved on September 28, 2017, and is also approved as monotherapy in a later-line setting.^1,9-11 Previous studies have identified certain patient characteristics and tumor features, such as histology and metastatic locations, that are prognostic or predictive of HT response.^12-17

CONTEXT

• Key Objective

• Three cyclin-dependent kinase 4/6 inhibitors (CDKIs) are approved by the US Food and Drug Administration for the treatment of patients with hormone receptor–positive, human epidermal growth factor receptor 2–negative, advanced or metastatic breast cancer in combination with hormonal therapy (HT). This research study examined whether demographic and baseline clinicopathologic data can be used to predict efficacy outcomes (progression-free survival, overall survival, and adverse event development) on an individual patient basis. Patient-level data from seven randomized controlled trials submitted to the US Food and Drug Administration were pooled.

• Knowledge Generated

• The pooled analysis demonstrates the capabilities of predictive modeling in the context of efficacy outcomes for patients. It also highlights the insufficiency of using only clinicopathologic data, prompting the incorporation of higher-resolution data.

• Relevance

• Utilization of comparative models can inform the conversation between patient and clinician, leading to a treatment plan that is aligned with the best interests of the patient. This would allow for maximizing benefit and quality of life while decreasing risks and toxicities.

To investigate the efficacy of CDKIs in combination with HT in less common clinicopathologic subgroups of MBC, Gao et al performed a pooled analysis of seven pivotal registration trials submitted to the FDA in support of marketing applications for palbociclib, ribociclib, and abemaciclib.^18,19 They found that the addition of CDKI to HT benefitted all analyzed subgroups and determined that further research is needed to characterize benefit within the first- or second-line setting. Although this study provided greater characterization of CDKI benefit in specific subgroups of participants, development of a model incorporating additional patient parameters may enhance prediction of benefit from CDK inhibition for individual patients. We hypothesized that on an individual patient basis, efficacy outcomes (measured as progression-free survival [PFS] and overall survival [OS], separately) and adverse event (AE) development can be predicted using demographic and baseline clinicopathologic variables.

METHODS

The set of models described in this article are intended for use in patients with hormone receptor–positive, HER2-negative MBC who are considering the use of a CDKI in the first- or second-line setting. To construct these models, we used randomized clinical trial data submitted to the FDA that compared the efficacy of CDKI in combination with HT with HT alone.

Prediction Problem

We hypothesized that on an individual patient basis, efficacy outcomes and AE development can be predicted using a model that incorporates baseline demographics and disease characteristics variables. We retrospectively analyzed data to construct prognostic models for survival (PFS and OS via random survival forests [RSF]) and classification (AE development via random forests [RF]). For the survival models, accuracy was based off Harrell's concordance index.²⁰ For the classification models, accuracy was measured via overall misclassification rate and visualized with the receiver operating characteristic curve.

Data

Individual-level data from seven randomized controlled trials (Fig 1) submitted to the FDA for new or supplemental marketing applications that investigated the efficacy of three CDKIs (palbociclib, ribociclib, and abemaciclib) in combination with standard HT (letrozole, anastrozole, and fulvestrant) were pooled. These trials included participants with hormone receptor–positive and HER2-negative advanced breast cancer or MBC. The intent-to-treat populations from these trials totaled 4,415 participants. However, 177 participants from the MONALEESA-7 trial received tamoxifen as their endocrine therapy, which has not been approved for use in combination with ribociclib, and were therefore excluded from analysis. The remaining 4,238 participants consisted of 1,605 who received HT alone and 2,633 who received a CDKI with HT.

FIG 1.

CONSORT diagram of patients included in study. CDKI, cyclin-dependent kinase 4/6 inhibitor; HT, hormonal therapy.

For each participant, demographic and clinical variables collected at initial trial screening visits were extracted and used as predictors. These variables included age, race, ethnicity, country and region of treatment, height, weight, body mass index, menopausal status, Eastern Cooperative Oncology Group (ECOG) performance status, estrogen receptor status, progesterone receptor status, histologic subtype, histologic grade, initial disease stage, sites of metastatic disease, number of metastatic sites involved, and line of therapy. To describe the study population, we used medians with minimum and maximum values and total values with proportions (Table 1). For response variables of the survival models, time to progression, time to death, time to last follow-up, and censorship status were extracted. For AE models, response variables were recorded as binary development (yes or no) irrespective of grade, frequency, or temporal development. Distributions of all outcome variables can be found in Table 2. Specific agents received while on the clinical trial were captured as a means of group stratification.

TABLE 1.

List of Demographic and Baseline Clinical Variables Extracted From Clinical Trial Databases for Use in Prediction Models

TABLE 2.

List of Outcome Variables to be Predicted With Models

After extraction, data values were cleaned to ensure consistent naming conventions (eg, a tumor of grade 1 v G1) and equivalent units (eg, weight in pounds v weight in kilograms). Descriptive statistics of each variable were analyzed for outlier detection and subsequent removal. Across all variables, there was a missingness of 8.9% that was imputed via missForest, a nonparametric algorithm for mixed-type data.²¹ For model building, the participants were split into two groups on the basis of the type of therapy they received (HT v HT plus CDKI). For each of these groups, an 80%-20% split for model training (HT: 1,284; HT plus CDKI: 2,107) and model testing (HT: 321; HT plus CDKI: 526) was used. All final models were internally validated using an ensemble bootstrap method that samples from the training set with replacement.

Survival Models

For prediction of PFS and OS in both the HT and HT plus CDKI participant groups, four separate RSFs consisting of 1,000 decision trees were constructed. In these trials, PFS was defined as time from date of random assignment to the initial date of documented cancer progression or death, whichever occurred first, and OS was defined as time from date of random assignment to death because of any cause. We specifically used an RSF because of the incorporation of censored survival data as a prediction variable.

AE Models

Binary RF classification models consisting of 1,000 decision trees each were constructed for prediction of AE development in both the HT and HT plus CDKI participant groups across the 18 most reported AEs that occurred in at least 10% of all participants. Grade and timing of AEs were ignored. Because of the infrequency of AEs in general, we undersampled the majority class (negative for AE development) to create a balanced set of participants for model building. RF models were also tested with oversampling the majority class (positive for AE development) and without sampling. Additionally, we tested k-nearest neighbors (KNN) and support vector machines (SVM) as potential models. Final model architectures were chosen on the basis of average area under the curve (AUC) values in combination with sensitivity and specificity measures.

Individual Patient Predictions

Using the set of participants reserved for model testing, the demographic and clinical variables of a single participant were used in each of the four survival models and 36 AE models to make predictions. Regardless of the participant's actual treatment pathway, two sets of predictions were made, one assuming administration of HT monotherapy and one assuming HT coadministration with CDKI. The resulting risk profiles for 3.5 years of treatment in both the PFS and OS setting were plotted on the same graph for visual analysis and compared across therapy choice via hazard ratios (HRs). For both PFS and OS, HR < 0.9 was defined as favoring HT plus CDKI, HR > 1.1 was defined as favoring HT, and 0.9 ≤ HR ≤ 1.1 was defined as no difference. Associated AE development probabilities were plotted as grouped bar graphs for visual comparison, whereas average probabilities across all AEs (P^HT and P^{HT plus CDKI}) were computed for general risk stratification (high [P > .5] v low [P < .5]). Both plots are paired together alongside demographic and clinical characteristics to produce an individual participant's complete assessment.

RESULTS

Participants

Participants' baseline demographics and clinicopathologic characteristics were similar across both HT and HT plus CDKI groups. Age ranged from 23 to 91 years (median 60) with the majority identifying as White (76%), followed by Asian (20%) and Black (2%). Most participants identified as non-Hispanic (91%). Clinically, most of the participants were postmenopausal (80%) and had an ECOG performance status of 0 (62%) or 1 (38%). Most of them had estrogen receptor–positive (100%) and/or progesterone receptor–positive (82%) disease. Ductal carcinoma was the most common histologic subtype (84%). All participants were female. Additional baseline characteristics for the pooled group and the separation into the HT and HT plus CDKI groups can be found in Table 1. Full details can be found in the Data Supplement.

Survival Models

The prediction models for PFS and OS in the HT group had overall accuracies of 66% and 64%, respectively, whereas prediction models in the HT plus CDKI group had overall accuracies of 62% and 63%, respectively. These values are based off Harrell's concordance index, which measures the concordance between orderings of observed and predicted survival times. For the participants reserved for the training set, the actual PFS and OS are compared against their predicted values in Figure 2. In all four cases, the models most strongly predicted survival in the first three years after start of treatment. For the HT PFS model, the strongest predictor, which contributed the most to the final prediction of the model, was the participant's menopausal status (Fig 3A), followed by stage, histologic grade, region, and number of metastatic sites. For the HT plus CDKI PFS model, the strongest predictor was histologic grade (Fig 3C), followed by menopausal status, country, region, age, and metastatic site (ie, visceral v bone only). For the HT OS model, the strongest predictor was the participant's line of therapy (Fig 3B), followed by number of metastases, metastatic site, ECOG, and region. For the HT plus CDKI OS model, the strongest predictor was histologic grade (Fig 3D), followed by number of metastases, line of therapy, ECOG, and metastatic site.

FIG 2.

Kaplan-Meier curves of PFS and OS for (A) HT patient group and (B) HT plus CDKI patient group. Actual patient values are shown in blue, whereas the predicted curves are shown in red (HT) and orange (HT plus CDKI). Solid lines represent OS and dashed lines represent PFS. CDKI, cyclin-dependent kinase 4/6 inhibitor; HT, hormonal therapy; OS, overall survival; PFS, progression-free survival.

FIG 3.

(A-D) Variable importance scores for the HT and HT plus CDKI survival models. BMI, body mass index; CDKI, cyclin-dependent kinase 4/6 inhibitor; ECOG, Eastern Cooperative Oncology Group; ER, estrogen receptor; HT, hormonal therapy; OS, overall survival; PFS, progression-free survival; PR, progesterone receptor.

AE Models

For AE development prediction, the RF models with undersampling associated with the HT group had an average AUC value of 0.613 (range 0.543-0.711), whereas the HT plus CDKI group had an average AUC value of 0.639 (range 0.546-0.714; Fig 4). For comparison, using unadjusted data in the HT setting yielded comparable average AUC values for RF (0.615), RF with oversampling (0.617), KNN (0.603), and SVM (0.528). For the HT plus CDKI group, average AUC values for RF (0.627), RF with oversampling (0.629), KNN (0.628), and SVM (0.541) were also comparable. Despite similar performance across all models, many naively made predictions completely aligned with the majority class (ie, predicting all participants did not develop the specific AE), resulting in a sensitivity of 1 and specificity of 0, and thus were not chosen as final models.

FIG 4.

(A and B) AUCs for the HT and HT plus CDKI AE models. AE, adverse event; AUC, area under the curve; CDKI, cyclin-dependent kinase 4/6 inhibitor; HT, hormonal therapy.

Individual Patient Predictions

For both individual PFS and OS predictions, we observed three specific scenarios as defined by favoring HT plus CDKI, favoring HT, and no difference, yielding nine different results (Data Supplement). The left column shows individual patients who favor HT plus CDKI in terms of PFS (HR < 0.9), whereas the right column shows those who favor HT in terms of PFS (HR > 1.1). Similarly, the top row shows individual patients who favor HT plus CDKI in terms of OS (HR < 0.9), whereas the bottom row shows those who favor HT in terms of OS (HR > 1.1).

For AE development, we observed four specific scenarios (Data Supplement). The top two graphs show individual patients who have a predicted low average probability of AEs on HT alone (P^HT < 0.5), whereas the bottom graphs show patients who have a predicted high average probability of AEs on HT alone (P^HT > 0.5). Similarly, the left two graphs show individual patients who have a predicted low average probability of AEs on HT plus CDKI (P^{HT plus CDKI} < 0.5), whereas the right graphs show patients who have a predicted high average probability of AEs on HT plus CDKI (P^{HT plus CDKI} > 0.5).

The paired survival benefit and AE risk stratifications yield distinct predictive phenotypes. For an individual, with specific demographic and clinical features, the results of the combined 40 models yield a predicted risk profile that is unique to that patient, as shown in the Data Supplement. The patient described in the Data Supplement has a predicted PFS and OS benefit from HT plus CDKI (HR^PFS = 0.684; HR^OS = 0.360) and a low risk of AE development with both HT plus CDKI (P^{HT plus CDKI} = 0.382) and HT alone (P^HT = 0.360). Of note, there is a singularly high prediction of diarrhea and abdominal pain with HT plus CDKI and a cough on either therapy. The patient described in the Data Supplement has no predicted PFS or OS benefit on either therapy (HR^PFS = 1.020; HR^OS = 0.960) and a low risk of AE development on HT alone (P^HT = 0.416). They have a borderline high risk of AE development with HT plus CDKI (P^{HT plus CDKI} = 0.505) in which nine categories have individually high prediction risks (neutropenia, diarrhea, nausea, leukopenia, headache, constipation, cough, infections and infestations, and rash). The patient described in the Data Supplement has a predicted PFS and OS benefit from HT alone (HR^PFS = 1.635; HR^OS = 1.847) and a low risk of AE development with both HT plus CDKI (P^{HT plus CDKI} = 0.450) and HT alone (P^HT = 0.444). Despite having low AE risk, six (nausea, arthralgia, hot flush, constipation, back pain, and abdominal pain) and seven (neutropenia, nausea, arthralgia, headache, hot flush, back pain, and rash) individual AEs have high risk on HT alone and HT plus CDKI, respectively.

DISCUSSION

We developed predictive models for PFS, OS, and AE in patients with hormone receptor–positive, HER2-negative MBC receiving HT with and without a CDKI. We found that menopausal status was the most important predictor of PFS on HT alone, whereas line of therapy was the most important predictor of OS. For the HT plus CDKI models, we found that the tumor histologic grade was the most important predictor of both PFS and OS. We also showed that although some AEs are infrequent, they can be predicted via robust sampling techniques.

Using baseline characteristics for an individual patient, comparison of predictions from paired models can provide prognostic information. Combining results of survival models and AE models can provide a more complete picture for both patients and physicians. CDKI-associated toxicities are generally manageable, prompting some patients to opt for treatment regardless of the predictive modeling. For other patients, the relationship between outcome predictions and AE predictions can inform the conversation between patient and clinician, ideally leading to a treatment plan that is aligned with the best interests of the patient.

For the survival models, the variable importance scores provide a deeper view on the risk assessments for an individual patient with respect to both PFS and OS. For any given model, a clinician can use the relative values of each variable to make a judgment on treatment benefit with regard to the other variables and models. For example, if the histologic grade of a patient could not be assessed (or the value is borderline between two groups), the treating physician could take this into account when analyzing the predicted results of a model given histologic grade ranks high in three of the four survival models. Additionally, the importance scores can provide a guideline for which variables are critical for capturing within the clinic for the purpose of prediction and which may benefit from enhanced resolution (eg, numerical staging v TNM staging) to add predictive power.

As with all model-building efforts, there are inherent biases that affect the results. Pooling data from seven randomized controlled trials allowed us to minimize any potential selection bias that may arise from any one trial by itself and increased the statistical power of the study. The variables that were chosen as predictors were intended to be a set of common variables that could easily be captured at a baseline clinical visit via routine methods. Regardless, some of these items are limited by their subjectivity (eg, ECOG performance status) and oversimplification (eg, race and stage at diagnosis [numerical v TNM]). Finally, the data and variables used for these models do not capture all patient variables (eg, diet, activity levels, financial toxicity, psychologic well-being, etc), thus reducing the prediction accuracy and increasing the chance for inaccurate results. Because of this, all predictions should be used within the context of the patient's current clinical status and with the expert guidance of a treating physician.

The accuracies of the survival models and AUCs of the AE models indicate modest performance highlighting the insufficiency of using only clinicopathologic data. As such, the investigation and incorporation of higher-resolution data, such as imaging, genomics, proteomics, and activity data, is the next logical step. In fact, Prat et al²² have recently published on the correlation between molecular subtype and PFS for patients receiving ribociclib in the MONALEESA trial. Similarly, O'Leary et al have explored genetic mutations observed in circulating tumor DNA in patients enrolled in PALOMA-3 as predictors of response to palbociclib.^23,24 These molecular and genomic analyses have shown great promise as predictive biomarkers. Adding these biomarkers to the existing models will improve prediction accuracy and, when combined with variable importance scores, can provide more refined predictions for therapeutic performance. Eventually, this same framework of models and techniques could be extended to additional drug types and patient groups. This would allow for enhanced mapping of therapy options and more personalized treatment recommendations and aid in generation of synthetic control groups for future clinical trials.

It is important to remember that the survival predictions represent risks over time as opposed to a single value. A potential next step for the survival models would be to incorporate real-time, temporal patient data to have survival risks update as new information becomes available. Similarly, the AE development predictions represent a binary occurrence variable, leaving room for the incorporation of grade, frequency, and timing of development. These improvements could enhance the overall clinical utility of the models, not just by empowering patient-physician conversations around treatment options but also by creating the potential to maximize benefit and quality of life while decreasing risks and toxicities.

To maximize the utility of these models, a web-based user interface that can be accessed by the public for general use could be a meaningful next step. This would provide a straightforward method of disseminating knowledge to patients, clinicians, and researchers, allowing for increased visibility while maintaining transparency. Increased usage of these models should lead not only to better outcomes but also to the incorporation of additional models in similar therapy settings. Although a single prediction model is limited in its capabilities, multiple models have the potential to enhance their utility as decision aids for patients and their oncologists.

AUTHOR CONTRIBUTIONS

Conception and design: Jeremy Mason, Richard Pazdur, Peter Kuhn, Gideon M. Blumenthal, Julia A. Beaver

Administrative support: Richard Pazdur

Collection and assembly of data: Jeremy Mason, Yutao Gong, Richard Pazdur, Gideon M. Blumenthal

Data analysis and interpretation: All authors

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 the 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).