The Tyrer-Cuzick Model Inaccurately Predicts Invasive Breast Cancer Risk in Women With LCIS

Abstract

Background

The Tyrer-Cuzick model has been shown to overestimate risk in women with atypical hyperplasia, although its accuracy among women with lobular carcinoma in situ (LCIS) is unknown. Here we evaluate the accuracy of the Tyrer-Cuzick model for predicting invasive breast cancer (IBC) development among women with LCIS.

Methods

Women with LCIS participating in surveillance from 1987-2017 were identified from a prospectively maintained database. Tyrer-Cuzick score (version 7) was calculated near the time of LCIS diagnosis. Patients with prior or concurrent breast cancer, a BRCA mutation, receiving chemoprevention, or with pleomorphic LCIS were excluded. Invasive cancer-free probability was estimated using the Kaplan-Meier method.

Results

1192 women with a median follow-up of 6 years (interquartile range [IQR] 2.5-9.9) were included. Median age at LCIS diagnosis was 49 years (IQR 45-55), 88% were White; 37% were postmenopausal, 28% had ≥1 first-degree family member with breast cancer, and 13% had ≥2 second-degree family members with breast cancer. 128 patients developed an IBC; median age at diagnosis was 54 years (IQR 49-61). 5- and 10-year cumulative incidences of invasive cancer were 8% (95% confidence interval [CI] 6-9%) and 14% (95% CI 12-17%), respectively. The median Tyrer-Cuzick 10-year risk score was 20.1 (IQR 17.4-24.3). Discrimination measured by the C-index was 0.493, confirming that the Tyrer-Cuzick model is not well calibrated in this patient population.

Conclusion

The Tyrer-Cuzick model is not accurate and may overpredict IBC risk for women with LCIS and therefore should not be used for breast cancer risk assessment in this high-risk population.

INTRODUCTION

Lobular carcinoma in situ (LCIS) is seen in approximately 0.5-4.0% of benign breast biopsies, and there has been a reported increased incidence,¹ likely attributed to the improvement in screening imaging techniques. Women diagnosed with LCIS have an elevated risk of subsequent ductal carcinoma in situ (DCIS) or invasive breast cancer development, with a risk of approximately 2% per year after diagnosis, and a cumulative risk of 26% at 15 years.² There is ongoing effort to better personalize risk prediction for individual women with LCIS to allow more personalized counseling on appropriate breast screening and risk-reducing strategies.

Multiple models have been developed to assist with breast cancer risk prediction efforts. One frequently used model which incorporates both personal and family history risk factors with personal history of benign breast disease, including LCIS, is the Tyrer-Cuzick model. The Tyrer-Cuzick model, also known as the IBIS (International Breast Cancer Intervention Study) is a risk assessment tool originally developed using a population-based cohort of unaffected offspring of mothers with breast or ovarian cancer.³ This model incorporates personal risk factors, such as endogenous hormonal factors, age, body mass index, and benign breast disease, with a comprehensive family history to provide an overall predicted 10-year and lifetime invasive breast cancer risk compared to the general population.⁴ In a study comparing 5 different breast cancer prediction models, the Tyrer-Cuzick model was found to be the most accurate in predicting breast cancer risk among a population of women screened in a program for women with a family history of breast or ovarian cancer.⁵ However, in a study from the Mayo Clinic Benign Breast Disease cohort which assessed risk in women with atypia, the Tyrer-Cuzick model was found to significantly overestimate risk and had poor concordance between predicted risk and invasive breast cancer development among women with atypical hyperplasia⁶.

The Tyrer-Cuzick model incorporates benign breast disease, including LCIS; however, there are minimal data assessing the accuracy of this model among women with this high-risk breast lesion. The aim of this study was to evaluate the accuracy of the Tyrer-Cuzick model for predicting invasive breast cancer development among women with LCIS.

METHODS

Following institutional review board approval at Memorial Sloan Kettering Cancer Center (MSK), women participating in surveillance following a diagnosis of LCIS were identified from a prospectively maintained database. Women entering surveillance between 1987 and 2017 were chosen for this analysis. The data required to calculate a Tyrer-Cuzick score were retrospectively collected for each patient near the time of LCIS diagnosis. Using the Tyrer-Cuzick version 7, a 10-year score was generated for each participant. Women missing a Tyrer-Cuzick score, with prior or concurrent DCIS or invasive breast cancer (defined as breast cancer diagnosed within 6 months of LCIS diagnosis), those receiving chemoprevention, and those with pleomorphic LCIS and BRCA mutation carriers were excluded. Outside pathology slides were reviewed by MSK breast pathologists to confirm the LCIS diagnosis using current World Health Organization diagnostic criteria.

Women participating in surveillance for LCIS at MSK were offered annual or biannual clinical breast examination and annual mammography. The frequency of clinical breast examination and the use of ultrasound or MRI screening was at the discretion of the treating physician and patient. MRI for high-risk screening was available after April 1999, and our experience with this screening modality in this cohort has been published previously.⁷

Patient and disease characteristics were summarized using the median and interquartile range (IQR) for continuous variables, and the frequency and percentage for categorical variables. Cancer-free probability was estimated using the Kaplan-Meier method. Cancer-free survival (CFS) was defined as time from LCIS diagnosis to development of first breast cancer. Patients who underwent prophylactic bilateral mastectomy or those who developed DCIS were censored at the date of the event. Those women who did not develop breast cancer during the study period were censored at their date of last follow-up. Cox regression models were used for multivariable analysis, with predictors of interest determined a priori.

Performance of the Tyrer-Cuzick model in this patient population was assessed via calibration and discrimination. The calibration plot is a visual assessment of the agreement between predicted risk and observed risk. To create the calibration plot, the Tyrer-Cuzick 10-year risk predictions were divided into deciles. In each decile, the median-predicted risk was estimated, and this was plotted on the x-axis as the value of predicted 10-year risk. The Kaplan-Meier estimate of cancer-free survival at 10 years was estimated among patients in each decile of predicted risk. One minus this value was plotted on the y-axis as the value of observed 10-year risk. Discrimination, a measure of how accurately patients were classified as cancer-free or not, was assessed using a concordance index (C-index) for right censored survival time data. A p-value < 0.05 was considered statistically significant. All statistical analyses were conducted in R software version 3.5.0 (R Core Development Team, Vienna, Austria).

RESULTS

1192 women diagnosed with LCIS and meeting study eligibility criteria between 1987 and 2017 were included in analyses. Median age at LCIS diagnosis was 49 years (IQR 45-55). 74 (6.2%) underwent bilateral prophylactic mastectomies during the surveillance period. Table 1 lists patient characteristics: 87.9% of patients were White, 36.6% were postmenopausal, 28.2% had ≥ 1 first-degree family member with breast cancer, and 13.3% had ≥ 2 second-degree family members with breast cancer.

TABLE 1.

Patient Characteristics

LCIS lobular carcinoma in situ, IQR interquartile range

Variable	Overall (n = 1192)
Age LCIS diagnosis, years, median, (IQR)	49 (45-55)
Missing	177
Race:
White	87.9% (1048)
Black	4.2% (50)
Asian	4.7% (56)
Unknown	3.2% (38)
Menopausal Status:
Pre/Peri	61.0% (727)
Post	36.6% (436)
Unknown	2.4% (29)
First-degree family member with breast cancer
Yes	28.2% (336)
No	70.1% (836)
Unknown	1.7% (20)
Second-degree family member with breast cancer
Yes	13.3% (159)
No	85.0% (1013)
Unknown	1.7% (20)

Development of Breast Cancer

Median follow-up among those who did not develop cancer was 6 years (IQR 2.5-9.9). During that time, 128 patients developed an invasive breast cancer. Median age at development of invasive breast cancer was 54 years (IQR 49-61). 107 patients had a breast cancer event before 10 years of follow-up, and 262 were followed for at least 10 years and did not have a breast cancer event. 802 patients were censored before 10 years, and therefore their 10-year breast cancer status is unknown. Patients who developed DCIS were censored at the time of diagnosis.

The 5- and 10-year cumulative incidences of cancer were 8% (95% confidence interval [CI] 6-9%) and 14% (95% CI 12-17%), respectively (Fig. 1). The median Tyrer-Cuzick predicted 10-year risk score was 20.1 (IQR 17.4-24.3)(Fig. 2).

Fig 1.

Cumulative incidence of invasive breast cancer. LCIS lobular carcinoma in situ

Fig. 2.

Distribution of Tyrer-Cuzick model 10-year risk score.

Evaluation of Risk Model

The Tyrer-Cuzick risk score was poorly calibrated in this patient population (Fig. 3.). For most of the range of predicted risk, the Tyrer-Cuzick score overestimated the risk of cancer, as evidenced by points falling below the 45-degree line, while in the lowest range of predicted risk, the Tyrer-Cuzick score underestimated the risk of cancer, as evidenced by one point falling above the 45-degree line. Discrimination measured by the C-index was 0.493, indicating that this model was no better than chance at predicting whether or not a patient would develop invasive breast cancer at 10 years.

Fig 3.

Calibration of the Tyrer-Cuzick risk score for predicting 10-year cancer risk. Error bars represent the 95% confidence interval of observed 10-year cancer probability.

DISCUSSION

Efforts are ongoing to improve risk stratification among women with high-risk breast lesions to allow personalization of recommended breast surveillance and risk-reducing strategies. While the Tyrer-Cuzick model is frequently used to estimate risk among women with LCIS, here we find that this tool may overestimate invasive breast cancer risk and has poor concordance between predicated and observed risk among this population.

The Tyrer-Cuzick model was originally developed using a Swedish population-based registry, which included daughters (n = 158,041) of women with breast or ovarian cancer, and their cancer incidence was followed between 1961 and 1993.³ This model differs from other risk assessment tools because it integrates a more comprehensive set of personal variables including benign breast disease, including both atypia and LCIS, as well as an extensive family history. In a study by Amir et al., 5 different breast cancer risk prediction models were compared among a cohort of 4536 women in surveillance for a family history of breast or ovarian cancer, and found that the Tyrer-Cuzick model was the most accurate and had a c-statistic of 0.762 among this population.⁵ The women included in this comparison were predominately followed for a family history of breast cancer, and there are no available data on the number of women with LCIS or other high-risk breast lesions.

Conversely, among a cohort of over 1100 women with LCIS, we found that the Tyrer-Cuzick model was not accurate and that it may overestimate risk of invasive breast cancer development in this cohort. Our findings are consistent with the results from an Australian population-based study of women with LCIS by Lo et al..⁸ The authors calculated the 10-year Tyrer-Cuzick score (IBIS-RET version 7), and assessed the calibration and discriminatory accuracy of this model among 732 women with LCIS from a population-based cancer registry. They reported a mean observed 10-year risk of invasive breast cancer of 14.1% compared to a mean Tyrer-Cuzick 10-year calculated risk of 20.9%. Similarly, the model showed poor discrimination, with a c-statistic of 0.54. Because this cohort was identified from a population-based cancer registry, it lacked information on chemoprevention use or prophylactic mastectomy for risk reduction, which may impact the outcomes of a prediction model. Our study, which includes detailed information on not only personal and familial risk factors, but also on the use of risk-reducing strategies, found nearly identical results, with a 10-year cumulative incidence of invasive breast cancer of 14% compared to a Tyrer-Cuzick predicted 10-year risk of 20.1% and a c-statistic of 0.493.

The accuracy of the Tyrer-Cuzick model has also been evaluated in a large population with atypical hyperplasia. In a cohort of 9376 women from the Mayo Benign Breast Disease cohort who underwent open benign breast biopsy between 1967 and 1991, 331 women were identified with atypical hyperplasia and comprised this study population. Boughey et al.⁶ calculated the 10-year Tyrer-Cuzick score and reported an observed-to-predicted ratio of invasive breast cancer of 0.53 (95% CI, 0.37-0.75), indicating that the model overpredicts breast cancer development among women with atypia. Furthermore, the c-statistic was 0.540, revealing that the Tyrer-Cuzick model did not accurately discriminate on an individual level between women who developed invasive breast cancer and those who did not.

Interestingly, studies assessing additional risk factors among women with atypical hyperplasia and LCIS have reported that family history of breast cancer is not an additive risk factor among women with high-risk breast lesions.^6,9,10 The Tyrer-Cuzick model incorporates an extensive family history and therefore may inherently overpredict risk when taking both of these independent risk factors into account.

Because the Tyrer-Cuzick model was found to have poor concordance and may overestimate risk among women with LCIS, this tool should not be used for risk assessment among this population. Two factors that have been reported to discriminate risk among women with LCIS include volume of LCIS and mammographic breast density.^11,12 In the era of personalized medicine, ongoing efforts are needed to improve breast cancer risk prediction, which drives recommendations for risk reduction among high-risk cohorts.

This study is a single-institution series that represents, to our knowledge, one of the largest cohorts of women with LCIS observed longitudinally over time. One limitation of this study is that the factors needed to generate a Tyrer-Cuzick score were retrospectively obtained from clinical documentation near the time of LCIS diagnosis, and, therefore, data points may have been omitted from the risk calculator if not documented in the chart. It is acknowledged that a documented family history included affected family members, but infrequently included a detailed history of unaffected family members. Because the Tyrer-Cuzick model includes both affected and unaffected family members, there is under-representation of unaffected family history. For a small number (n = 11) of women with detailed family trees available for review from a genetic consultation, the Tyrer-Cuzick model was calculated with both the affected-only and the affected-and-unaffected family tree, and we found no difference in reported scores, suggesting that this limitation would not significantly impact our results.

Conclusions

The Tyrer-Cuzick model may overpredict invasive breast cancer risk in women with LCIS. Individual risk estimates are not concordant between predicted risk and invasive breast cancer development, suggesting that this model should not be used for breast cancer risk assessment in women with LCIS.

Synopsis: Here we evaluate the Tyrer-Cuzick model’s accuracy for predicting invasive breast cancer development among women with LCIS. We find it overpredicts invasive breast cancer risk in this population, and that it should not be used for breast cancer risk assessment in LCIS patients