Risk Model Tutorial

1. To identify women who may benefit from risk-reducing medications
The Gail model is used to determine risk for purposes of advising on use of medications to reduce risk. In the National Surgical Adjuvant Breast and Bowel Project (NSABP) P1 study [2], women at increased risk for breast cancer were defined as follows: 1) age 35 to 59 years with at least a 1.66% five-year risk for developing breast cancer by the Gail model; or 2) personal history of lobular carcinoma in situ (LCIS); or 3) age over 60 years. 13,388 such women were randomized to receive tamoxifen or placebo daily for five years. Tamoxifen reduced the risk of invasive breast cancer by 49% and reduced the risk of noninvasive cancer by 50%.

The reduced risk of breast cancer was only seen for estrogen-receptor expressing tumors. There was a 2.5-fold increase in risk of endometrial cancer in women taking tamoxifen and a decrease in hip and spine fracture risk. Blood clots causing stroke and deep vein thrombosis are increased in women taking tamoxifen [3, 4].

2. To identify women who may carry a pathogenic (disease-causing) variant (mutation) in BRCA1or BRCA2
Some models (e.g. Tyrer-Cuzick [IBIS], Penn II, BOADICEA, BRCAPRO) estimate the probability of a pathogenic/likely pathogenic BRCA1/2 variant; however, most genetic testing guidelines are now criterion based (e.g. NCCN) as opposed to probability based. In practical terms, clinical decision-making around genetic testing is rarely based on a priori probabilities.

3. To identify women who meet criteria for elevated-risk screening MRI
Breast density increases the risk for developing breast cancer and also increases the risk of cancer going unseen on a mammogram.  Compared to women with scattered fibroglandular density, which is not considered dense and is most common over the age of 50, individuals with heterogeneously dense breasts have about a 1.5-fold greater risk of developing breast cancer while those with extremely dense breasts have about a 2-fold greater risk [5].  Visual (BI-RADS) breast density assessed by the radiologist, or quantified by software is included in many of the models of breast cancer risk.

Women with an estimated lifetime risk (LTR) of breast cancer of 20% or greater are considered “high risk” and are recommended to have annual screening MRI. Starting age varies according to different guidelines [1, 6-8] but may be as early as age 25-30 [1]. To estimate lifetime risk for purposes of supplemental MRI screening, one can use any model that predicts risk of a disease-causing mutation, such as: the Tyrer-Cuzick (IBIS), BOADICEA/CanRisk, or BRCAPRO models; or the Claus model can be used. Lifetime risk decreases with increasing age, and it is unusual for women over age 70 to be considered “high risk”. The Gail model should NOT be used, as it only collects limited family history and does not include breast density. The BCSC Invasive Breast Cancer Risk Calculator provides 5- and 10-year risk, for which there is currently no MRI screening guideline. Validated artificial intelligence (AI) based risk assessment such as Mia, Mirai, iCAD, or Transpara may also be used.

Other women at elevated risk are also recommended for screening MRI, including those with known pathogenic variants.

The Risk Models Table that follows features details and live links to several commonly utilized breast cancer risk assessment models.

Models that do include breast density in risk calculations:

• Tyrer-Cuzick Model (IBIS) version 8 update was based in part on input from Dr. Jennifer Harvey and Dr. Martin Yaffe and includes breast density. In this model, breast density is one of the top five factors determining breast cancer risk. This model is the most comprehensive and tends to be the most accurate at predicting risk at the population level.
• Breast Cancer Surveillance Consortium (BCSC) model [1] is a modification of the Gail model and was developed and validated in a large, ethnically diverse, prospective cohort of women undergoing screening mammography. It includes the risk factors with the greatest population attributable risks for breast cancer, including age, breast density, family history, history of a breast biopsy, and a polygenic risk score (PRS) based on common genetic variations [2].
• Artificial Intelligence (AI) is being used to identify textural and other findings beyond breast density on mammograms that predict increased risk; such information is complementary to the Tyrer-Cuzick model (v.8) [3].

• • A recent study of a mammography-based deep learning model developed at Massachusetts General Hospital called Mirai more accurately identified high-risk patients than the Tyrer-Cuzick v8 risk model and prior deep learning models. Mirai identified 41.5% of women developing breast cancer in the next five years as high-risk, compared to only 22.9% with the Tyrer-Cuzick model v8. Mirai performed similarly across all density categories [4].  Addition of family history or mutation status did not further improve on performance of AI alone.
  • In a study from the Karolinska Institutet [5], use of AI to identify mammographic microcalcifications and masses, even if not the site of actual malignancy, and differences between right and left breasts,  successfully predicted women who would develop interval or advanced cancer in the two years after a normal mammogram and improved short-term (2-to-3-year) risk assessment over Tyrer-Cuzick (v.7) or Gail models [6]. This model proved more accurate than traditional risk models and can augment genetic/family history to help identify women who should and, importantly, who should not, have supplemental screening after 2D mammography.

Models that do not include breast density in risk calculations:

• Gail [7]
• Penn II
• Claus [8]

Risk models and diagnostic considerations:

• Any particular risk model will not include all known risk factors
• Estimated risk can vary substantially between models
• Age: As a woman gets older, her 5- and 10-year risk of developing breast cancer increase, but her lifetime risk decreases
• Risks change every year and should be periodically reassessed. For example, risk changes with age (see above), family history may change if family members have been diagnosed with breast or ovarian cancer, or a breast biopsy may have been performed.

Risk model limitations:

• Adoption (or otherwise unknown family history)
• Small family size
• All models underestimate rates of breast cancer. At best they identify about 70% of women who will develop cancer at the population level.
• All models are low in accuracy at the individual level (“discrimination”)
• Risk models have been largely developed and validated in populations of white European ancestry and may not perform equally in other ethnic populations

Polygenic risk score for improvement in model performance:
Single nucleotide polymorphisms (SNPs, informally, “snips”) are variations that occur normally throughout the DNA. Individually, their impact on breast cancer risk is small; however, when combined to create a “polygenic risk score” (PRS) or “SNP score”, some of these variations increase the risk of breast cancer. Including PRS with classical risk factors used in current risk models may improve accuracy of risk assessment [1-4]. Measurement of PRS is not yet widely available.

Artificial Intelligence (AI)-based risk assessment:
Mammography-based features beyond breast density identified from deep learning models [5, 6] and iCAD software [7, 8] have been used in combination with traditional risk factors to improve risk prediction. More recently a model was developed and validated in a U.S. multiethnic cohort to estimate risks using tomosynthesis imaging features and age to help identify women who would be diagnosed with an interval cancer or a cancer at the next screen after negative or benign tomosynthesis screen [9]. These approaches can help better identify women who should and, importantly, who should not, have supplemental screening after 2D mammography or tomosynthesis. AI can also help triage mammograms more likely to be normal, allowing radiologists to focus greater attention on those more likely to have cancer.

AI is also being used with breast ultrasound.  AI-based decision support performs similar to breast imaging specialists [10] and has been shown to improve accuracy of sonographic breast lesion assessment among nonspecialists [11]. AI can reduce false positives on ultrasound [12].

See NCCN Clinical Practice Guidelines entitled “Genetic/Familial High-Risk Assessment: Breast, Ovarian, Pancreatic, and Prostate” for genetic testing indications [13].

Women who have an estimated lifetime risk (LTR) of 20% or more by models that predict mutation carrier status (e.g., Tyrer-Cuzick (IBIS), Penn II, BOADICEA, BRACAPRO), or the Claus model, are recommended to have annual MRI screening (to include mammography after age 30). The Gail model should not be used for MRI screening purposes, as it includes very limited family history.

In women recommended for MRI but unable to access or tolerate it, consider contrast-enhanced mammography or supplemental ultrasound screening [14].

^*10-year risk by Tyrer-Cuzick model can be divided by two to estimate 5-year risk (personal communication, Jack Cuzick, PhD, 6/10/20)