Predicting Adjuvant Endocrine Therapy Initiation and Adherence Among Older Women with Early Stage Breast Cancer

Abstract

Purpose:

The CALGB 9343 trial demonstrated that women age 70 or older with early-stage, ER+ breast cancer (BC) may safely forgo radiation therapy (RT) and be treated with breast conserving surgery (BCS) followed by adjuvant endocrine therapy (AET) alone. However, most patients in this population still undergo RT in part because AET adherence is low. We sought to develop a predictive model for AET initiation and adherence in order to improve decision-making with respect to RT omission.

Methods:

Women ages 70 and older with early stage, ER+ BC were identified using the Surveillance, Epidemiology, and End Results (SEER)-Medicare database. Comorbidities, socioeconomic measures, prescription medications, and demographics were collected as potential predictors. Bivariate analysis was performed to identify factors associated with AET initiation and adherence. Stepwise selection of significant predictors was used to develop logistic-regression classifiers for initiation and adherence. Model performance was evaluated using the c-statistic and other measures.

Results:

11,037 patients met inclusion criteria. Within the cohort, 8,703 (78.9%) patients initiated AET and 6,685 (60.6%) were adherent to AET over one year. Bivariate predictors of AET initiation were similar to predictors of adherence. The best AET initiation and adherence classifiers were poorly predictive with c-statistics of 0.65 and 0.60, respectively.

Conclusions:

The best models in the present study were poorly predictive, demonstrating that the reasons for initiation and adherence to AET are complex and individual to the patient, and therefore difficult to predict. Initiation and adherence to AET are important factors in decision-making regarding whether or not to forgo adjuvant RT. In order to better formulate treatment plans for this population, future work should focus on improving individual prediction of AET initiation and adherence.

Background

Adjuvant radiation therapy (RT) lowers the risk of breast cancer recurrence in patients with invasive breast cancer who are treated with breast conserving surgery (BCS) but does not change survival[1]. Breast cancer recurrence is further reduced by the addition of adjuvant endocrine therapy (AET) in patients with estrogen receptor (ER) positive tumors[2,3]. The overall benefit of this regimen— BCS followed by RT and AET— has been shown to be heterogeneous, with some patient groups benefiting more than others from the addition of RT[3,4]. It has also been shown that age at diagnosis is a significant predictor of invasive breast cancer recurrence, with older age women having a decreased risk of recurrence than younger women[5,6]. The Cancer and Leukemia Group B (CALGB) 9343 trial randomized women 70 years and older with early-stage ER+ tumors who were treated with BCS to RT or no RT. The short and long-term results of this study demonstrated that RT did not improve overall or disease-free survival in this patient population [7,8]. There was a small improvement in recurrence rates for the RT cohort (1% in the group versus 4% in the no RT group). In response to the results of CALGB 9343, the 2005 National Comprehensive Cancer Network (NCCN) clinical practice guidelines and 2010 European Society for Medical Oncology (ESMO) practice guidelines stated that CALGB 9343 eligible patients could safely be treated with BCS and AET alone[9,10].

Despite the results of CALGB-9343 and subsequent change in NCCN and ESMO guidelines, the vast majority of older women with invasive breast cancer continue to be treated with adjuvant RT [11,12]. A recent study by Reyes et al which analyzed national Cancer Database data on elderly women with invasive breast cancer in the United States between 2004 and 2014 showed that omission of RT only occurred in 0.1% of CALGB-9343 eligible women[13]. The reasons why so many elderly women still undergo adjuvant RT despite the apparent safety of its omission likely includes patient preferences, physician opinion, and known lack of adherence to AET overall, especially in older patients [11]. AET adherence outside of the clinical trial setting is low, and leaves AET non-adherent patients who also omit RT at an increased risk for breast cancer recurrence[3]. There have been numerous studies examining potential predictors of AET non-adherence, though a significant limitation is that these have largely focused on factors present once patients had undergone RT and were already receiving AET[14–16]. Despite these efforts, accurately predicting AET adherence at the time a decision is made regarding RT is not currently possible.

Considering the results of CALGB-9343, there is a significant proportion of older women who will be adherent to AET and could therefore safely omit RT. In order to confidently recommend omission of RT and formulate and individual treatment plan that is neither under- nor over-treatment, it would be beneficial for the treating physician(s) to accurately predict patient initiation and level of adherence to AET in the immediate post-operative period. We therefore sought to develop an algorithmic classifier to predict AET initiation and adherence for the population of women over 70 with early stage, ER+ breast cancer.

Methods

Patient identification and data collection

The Surveillance, Epidemiology, and End Results (SEER)-Medicare database was queried to identify women diagnosed with breast cancer from 2007–2015 whose clinical characteristics matched the inclusion criteria of the C9343 trial. These included women with a first diagnosis of breast cancer at age 70 or older with ER+ tumors <2cm in size, lymph node negative, with stage 1 disease, who underwent BCS (Figure 1).

Fig. 1.

Consort diagram demonstrating the number of patients included in the study and the reasons for patient exclusion.

We only included women without AET initiation or that newly started AET whose BCS surgery claims were identified within 9 months of diagnosis. In order to capture measures without bias due to missing claims, we restricted all beneficiaries to be fee-for-service/ non-health maintenance organization (HMO) and part D enrolled during the year after their diagnosis date and during the year prior to their BCS surgery. Furthermore, we ensured beneficiaries with RT therapy, had 6-month part D enrollment after RT initiation (Figure 1). Patient comorbidities, socioeconomic measures, and demographics were collected as potential predictors of initiation and adherence to AET. The Charlson comorbidity index was utilized as an overall measure of comorbid conditions[17]. AET adherence was measured using the Medication Possession Ratio (MPR) which was calculated as the ratio of days’ supply of drug dispensed to the total days where the medication was prescribed. MPR is a validated indicator of adherence. In this study, the conventional cutoff of 0.8 was used with high adherence defined as MPR ≥.80 in accordance with prior literature on medication adherence in patients with breast cancer [18,19]. Patients for whom AET was indicated but not initiated were classified as having zero adherence.

Adherence to prior medications is a possible predictor of adherence to AET. As such, we chose to include adherence to commonly prescribed medications (statins and antihypertensives) as possible predictors. In order to capture adherence to prior medications, statin and antihypertensive national drug codes (NDC) were used to identify patients prescribed these medications. The Proportion of Days Coverage (PDC) for a medication type was assessed for the sub-cohort of patients prescribed these medications. The PDC was calculated for the year following first detection of statin or antihypertensive medication use. “HI PDC Statin/hypertension” was defined as PDC by statin/hypertension medications (> 80% vs <=80%) during the first year after prescription of these medications.

Descriptive statistical methods

Bivariate relative risks for each candidate predictor were examined. A log-Poisson model with robust standard errors [20–22] was used to estimate the risk ratio of AET adherence (MPR ≥.80) and AET initiation by levels of the risk factor. Risk was considered statistically significant if the false discovery rate (FDR) adjusted p-values were <0.05.

Classifier development and selection of predictors

To establish a predictive model, the dataset was divided into training (50%), validation (25%), and test sets (25%). These sets were created using the “partition” statement in the SAS HPLOGISTIC procedure which randomly assigns observations to one of the 3 roles. All observations were randomly assigned proportionally to 50% training, 25% test, and 25% validation. The HPLOGISTIC procedure in the SAS statistical software was used to develop an optimal classifier using the training dataset and regression with stepwise selection. This method begins by selecting the most predictive variable and sequentially adds predictive variables according to entry p-value significance level while removing variables according to their stay p-values. The selection proceeds until all candidates for removal are significant at stay level and no candidate for entry is significant at the entry level. The final sequence of variables in the model was then chosen by the sequence minimizing the Akaike information criterion (AIC). Additionally, varying best models were compared by varying the entry/stay p-value thresholds from 0.01 to 0.50 and comparing their performance, using their validation set c-statistics (or area underneath a ROC curve) to choose the model with best discrimination. The final model was the one with the highest c-statistic.

We then examined whether prediction could be improved by adding two-way interactions into the final model. This was performed using the HPLOGISTIC procedure, and entry/stay procedures were varied from 0.01 to 0.50 and final performance was compared using c-statistics in the validation data set. The main effects model was retained based on parsimony and best performance on the validation set. Finally, actual performance of the model was evaluated in the test data set, calculating test data c-statistics to examine discrimination and a test data Hosmer-Ledeshow test to examine calibration. Model parameters were fit using all the data.

Treatment of missing data

In order to treat missing data in race and marital status, a multiple imputation FCS procedure was conducted using Proc MI in SAS. For purposes of classifier development, only the imputed values form the first imputed data set for marital status and race were selected. For purposes of estimation, analyses were separately conducted on the imputed data sets and combined according to Rubin’s multiple imputation rules using proc MIANALYZE.

Results

Patient demographics and candidate predictors of AET

We identified 11,037 patients who met the C9343 inclusion criteria. Of the eligible patients, 8,703 (78.9%) initiated AET within one year of breast cancer diagnosis, and 8,523 (77.2%) underwent RT. Overall, a majority of the patients were Caucasian (89%) with a mean age of 76.5 years (± 5.0). A significant proportion of patients (39.4%) had poor adherence to AET as measured by a low medication possession ratio (<0.80) at 1 year of follow up. Patient demographics, comorbidities, and candidate predictors of AET initiation and adherence are shown in Table 1.

Table 1.

Characteristics and Predictors of Adjuvant Endocrine Therapy Initiation and Adherence (N = 11,037)

Overall	N (%)		N(%)
Age Quartiles		Charlson Category
Q1 [70–72]	2907 (26.3)	0	4011 (36.3)
Q2 [72<−76]	3326 (30.1)	1	2981 (27.0)
Q3 [76<−80]	2451(22.2)	2,3	2891 (26.2)
Q4 [81+]	2353 (21.3)	4+	1154(10.5)
Mean (Std) [IQR]	76.5 (5.0) [8]	Mean(Std)[IQR]	1.4 (1.6)[2.0]
Race		Back/Neck Pain	4559 (41.31)
White	9890 (89.7)	Migraine	1053 (9.5)
Other	1136(10.3)	Muscular Pain	3727 (33.8)
Missing	11	Arthritis/Joint Pain	5777(52.34)
Marital Status		Neuralgia	661 (6.0)
Single	5467(51.8)	Chronic Pain	335 (3.0)
Partner	5079 (48.2)	Depression	1550(14.0)
Missing	491	Substance Abuse	140. (1.3)
Yost SES Index		Anxiety	1586(14.4)
Q1[≤10.5K]	2760 (25.0)	Congestive Heart Failure (CHF)	1007 (9.1)
Q2[10.5K<−11.2K]	2760 (25.0)	Diabetes	3235 (29.3)
Q3[11.2K<−11.6K]	2779 (25.2)	Acute Myocardial Infarction (MI)	123 (1.1)
Q4[11.6K<]	2738 (24.8)	Dementia	241 (2.2)
Mean(Std) [IQR]	11K(0.8K)[1K]	Peripheral Vascular Disease (PVD)	1881 (17.0)
Rad. Facilities /100K 2010,50mi radius		Cerebrovascular Disease (CVD)	1627(14.7)
Q1[≤0.4]	3157 (28.6)	History of MI	380 (3.4)
Q2[0.4<−0.5]	2410(21.8)	Hemiplegia/Paralysis	49 (0.4)
Q3[0.5<−0.7]	2713 (24.6)	Rheumatologic Diseases	532(4.8)
Q4[0.7<]	2757 (25.0)	COPD	2758 (25.0)
Mean(Std) [IQR]	0.54(0.27)[0.27]	Moderate-Severe Renal Disease	1072 (9.7)
Radiation Oncs/lOOk 2010,50 mi radius		Liver Disease	102 (0.9)
Q1[≤1.2]	2794 (25.3)	Peptic Ulcer Disease	186(1.7)
Q2[1.2<−1.5]	2728 (24.7)	Radiation Therapy
Q3[1.5<−1.9]	2820 (25.6)	No RT	2514(22.8)
Q4[≤1.9]	2695 (24.4)	RT	8523 (77.2)
Mean(Std) [IQR]	1.6 (0.6) [0.7]	AET Therapy
PCP / 100k 2010–15, 50 mi radius		No	2334 (21.1)
Q1[≤67]	2784(25.2)	Yes	8703 (78.9)
Q2[67<−78]	2840(25.7)	High AET MPR
Q3[78<−89]	2979(25.3)	≤.80	4352(39.4)
Q4[89<]	2616(23.7)	>.80	6685 (60.6)
Mean(Std) [IQR]	77.6(14.8)[21.9]
Metropolian Status		Anti-Hypertensive Adherence
Non-Metropolitan	1631 (14.8)	<.80	1002(15.0)
Metropolitan	9406 (85.2)	>.80	5667 (85.0)
Dual Status lyr post diagnosis		Statin Adherence
Non-Dual	9537 (86.4)	≤.80	1147 (27.6)
Dual	1500(13.6)	>.80	3008 (72.4)
Tumor Size
<1 cm	4523 (41.0)
1–2 cm	6514(59.0)
Tumor Sequence
1st Tumor	9261 (83.9)
2nd+Tumor	1776(16.1)
Polypharmacy (1 yr <BCS)
Q1[≤6]	2801 (25.4)
Q2[6<−10]	2887 (26.2)
Q3[10<−15]	2738 (24.8)
Q4[15<]	2611 (23.7)
Mean(Std)[IQR]	11.5 (7.2) [9.0]

SES = Socio-economic Status; Rad. = Radiation; Radiation Oncs = Radiation Oncologists; PCP = Primary Care Provider; COPD = Chronic Obstructive Pulmonary Disease; AET = Adjuvant Endocrine Therapy; MPR = Medication Possession Ratio

Bivariate predictors of AET initiation

In order to create a classifier tool to predict which patients would initiate and adhere to AET, we first performed a bivariate analysis to identify significant individual predictive factors. There were 22 factors significantly associated with AET initiation. Factors associated with lower initiation of AET included increasing age with a risk-ratio (RR) of 0.84 (95% CI 0.83–0.86), single marital status (RR 0.95, 95% CI 0.93–0.97), white race (RR 0.96, 95%CI 0.93–0.99), lower primary care practitioner density (RR 0.96, 95%CI 0.93–0.98), lower radiation oncologist practitioner density (RR 0.95, 95%CI 0.92–0.97), second tumor diagnosis (RR 0.94, 95%CI 0.91–0.97), and a number of comorbid conditions. Figure 2 demonstrates the results of the bivariate analysis of potential AET initiation predictors. Interestingly, patients that had undergone radiation were 20% more likely to initiate AET than those who omitted it (RR 1.20, 95% 1.17–1.23).

Fig. 2.

Bivariate predictors of adjuvant endocrine therapy initiation. Ref = Reference; RR = Risk Ratio; LCL, UCL = 95% Confidence Intervals; P = Probability; P0 = Reference Probability; fdr pval = FDR adjusted p-value; YostSES = Yost Socioeconomic Status; PCP = Primary Care Provider; Rad Fac = Radiation Facility; Rad Onc = Radiation Oncologist; dx = diagnosis; PVD = Peripheral Vascular Disease; CVD = Cardiovascular Disease; MI = Myocardial infarction; COPD = Chronic Obstructive Pulmonary Disease; CHF = Congestive Heart Failure

Bivariate predictors of AET adherence

We then examined the potential predictors of AET adherence by analyzing bivariate predictors of having an MPR greater or less than 0.80. The factors predicting a high MPR largely followed a similar pattern to the factors significantly associated with AET initiation. Notably, however, lower radiation facility density (RR 0.95, 95%CI 0.91–0.99), substance abuse history (RR 0.81, 95% CI 0.69–0.96), and COPD history (RR 0.95, 95%CI 0.91–0.98) were significantly associated with lower chance of high MPR but not associated with AET initiation (Figure 3). Additionally, adherence (Proportion of Days Coverage, or PDC >0.80) to antihypertensives (RR1.13, 95%CI 1.07 – 1.20) and cholesterol-lowering drugs (RR 1.14, 95%CI 1.08 – 1.21) was significantly associated with MPR > 0.80 for AET but not for AET initiation. Undergoing radiation was predictive of AET adherence just as it was for initiation (RR 1.23, 95%CI 1.18–1.28).

Fig. 3.

Bivariate predictors of adjuvant endocrine therapy adherence. Adherence was considered low if the medication possession ratio (MPR) was <0.80. “Hi” MPR defined as MPR > 0.80. Ref = Reference; RR = Risk Ratio; LCL, UCL = 95% Confidence Intervals; P = Probability; P0 = Reference Probability; fdr pval = FDR adjusted p-value; YostSES = Yost Socioeconomic Status; PCP = Primary Care Provider; Rad Fac = Radiation Facility; Rad Onc = Radiation Oncologist; dx = diagnosis; PVD = Peripheral Vascular Disease; CVD = Cardiovascular Disease; MI = Myocardial infarction; COPD = Chronic Obstructive Pulmonary Disease; CHF = Congestive Heart Failure; HT = Anti-hypertensive drug; ST = Statin drug

Logistic regression classifiers of AET initiation and adherence

The training dataset was used to develop predictive models of AET initiation and adherence. After best model selection, 23 variables were included in the logistic regression classifier for AET initiation (Figure 4), and 22 variables were included in the adherence classifier (Figure 5).

Fig. 4.

Logistic Regression Classifier for AET Initiation. Ref = Reference; OR = Odds Ratio; LCL, UCL = 95% Confidence Intervals; Pr = p-value; COPD = Chronic Obstructive Pulmonary Disease; CHF = Congestive Heart Failure

Fig. 5.

Logistic Regression Classifier for Adherence (MPR > .80). Ref = Reference; OR = Odds Ratio; LCL, UCL = 95% Confidence Intervals; Pr = p-value; COPD = Chronic Obstructive Pulmonary Disease.

In the testing dataset, the best AET initiation classifier was poorly predictive (c-statistic = 0.65), with a specificity of 53%, sensitivity of 69%, and positive predictive value (PPV) of 84%. Similarly, the AET adherence classifier was unreliable (c-statistic = 0.60), with a specificity of 49%, sensitivity of 64%, and PPV of 66% (Table 2). Addition of antihypertension or statin adherence prior to surgery did not meaningfully improve the model performances.

Table 2.

Performance of AET initiation and adherence classifiers.

Measure	Initiation classifier	Adherence classifier
Area under ROCC*	0.65	0.60
Sensitivity	69%	64%
Specificity	53%	49%
Positive Predictive Value	84%	66%
Negative Predictive Value	32%	48%

^*ROCC = Receiver Operating Characteristic Curve. Area under ROCC of a random classifier is 0.50. Area under ROCC >0.85 is typically considered a good classifier.

Discussion

Utilizing the SEER-Medicare database, we sought to establish a set of predictive classifiers to determine which CALGB 9343 eligible patients were likely to initiate and adhere to AET after BCS. We identified a number of demographic and patient characteristics that were associated with AET initiation and MPR on bivariate analysis and used these to build our predictive models. Despite rigorous selection among candidate predictors, the best classifiers remained poorly predictive of both endpoints. To our knowledge, this is the first report of an attempt to develop an AET initiation and adherence classifier in this population of breast cancer patients.

The benefit of AET has been well established but a significant proportion of patients do not initiate or adhere to therapy. As such, there have been several efforts to identify patient features associated with AET initiation and adherence [23,16,24,25,15,26,27]. In the present study, adherence at one year after diagnosis was just 60.4%. This is consistent with other studies which have shown a non-adherence rate ranging between 20–60% [25,28–31,26,27]. Additionally, the findings from our bivariate analysis largely reflect those from these prior studies, with features such as high Charlson Comorbidity Index, lack of healthcare access, and demographic factors correlating with lower initiation and adherence.

Various studies have also shown that increasing age is correlated with lower AET utilization [32,33,11]. The most common reason for poor adherence is medication side-effects, with joint pain and weight gain being common reasons for AET cessation [34,16,25,29]. Additionally, non-adherence may also reflect physician and patient perceptions of lack of benefit from AET at older age. In one study of adherence among breast cancer patients, older age correlated significantly with lack of belief in AET benefit [23]. This suggests that shaping patient beliefs about therapy is essential in encouraging adherence, especially if radiation is omitted. The results of CALGB 9343 suggest that RT is over-treatment in older women with early stage breast cancer. Importantly, AET is a critical component of therapy when radiation is omitted in order to prevent unacceptable recurrence and survival rates. The lower adherence in older patients is therefore particularly concerning when treating physicians consider recommending treatment plans based on the findings of CALGB 9343. RT omission may in fact represent under-treatment in many cases, as prior studies have found AET adherence as low as 49% in patients for whom RT is omitted [11]. Ultimately, in order to prevent either over or under-treatment, patients need to be correctly assigned a treatment based on their ability to adhere to AET or not.

Predicting AET initiation and adherence in older women with early-stage breast cancer remains an unsolved challenge. The best models developed here were poorly predictive of initiation and adherence with c-statistics of 0.65 and 0.60, respectively. The models ultimately incorrectly classified 22% of patients with respect to initiation and 38% for adherence. Attempts to develop predictive medication adherence classifiers have been made in other fields as well, with a low rate of success overall. One such study, of patients for whom bisphosphonates were prescribed for osteoporosis, reported a similar success rate as ours for predicting adherence [35]. This study utilized logistic regression to predict high or low MPR (>=0.80) and further attempted to improve prediction by comparing the model results in parallel using recursive partitioning trees. Despite also finding significant individual predictors of bisphosphonate adherence, the resulting models in the above study had a best c-statistic of 0.62. Others have attempted to develop predictive models by selecting predictors using the least absolute shrinkage and selection operator (LASSO) and machine learning, though best prediction accuracy did not surpass 70% [36,37].

The relatively high proportion of patients misclassified in our models and in other studies suggests that medication adherence is affected by a variety of factors that are inadequately captured in retrospective review, and are likely most dependent on factors such as side-effects which develop well after a decision is made regarding RT. While certain factors such as demographics, comorbidities, and disease specifics may individually associate with AET adherence or initiation, most models do not consider individual patient psychology. Personality and health attitudes, which may vary significantly among individuals despite similar clinical and demographic profiles, affect initiation and adherence to AET and other medications [38,39]. Though these factors have been studied in patients taking AET, these data are not easily captured retrospectively and therefore not included in studies developing predictive models of initiation and adherence. Given the high frequency of misclassification, future efforts could establish psychological screenings to be included in predictive models when making decisions regarding whether or not to omit RT.

There are several limitations to the present study. While the SEER-Medicare database captures a significant proportion of women over 70 undergoing care for early stage breast cancer in the United States, it is not all-encompassing. The MPR was captured using Medicare part D claims data which is based on whether a prescription was filled or not, and may not be representative of whether AET was actually taken. Furthermore, the study is limited by its retrospective nature and the fact that candidate predictors were chosen from the available variables captured in SEER. Patient specifics such as healthcare attitudes and personality traits could not be included in the predictive models which may have limited their accuracy.

Conclusions

While not widely adopted, the CALGB-9343 trial supports omission of RT in older women with early stage breast cancer. Initiation and adherence to AET are important factors in decision-making regarding whether or not to forgo adjuvant RT. We sought to develop a predictive model of AET initiation and adherence. Despite including a wide range of socioeconomic factors, social determinants of health, comorbidities, and cancer-specifics, the best models remained poorly predictive. The decision to initiate and adhere to AET is individual and complex and therefore difficult to predict from factors available in SEER. Future work should focus on improving individual prediction of AET initiation and adherence in order to avoid over-treatment or unnecessary treatment. This information will allow patients and clinicians to make a patient-centered decision regarding RT omission after BCS.