 |
|
|||||||||||
| Mammography-Based AI (Image-only) Risk Prediction Models | |
|
|
|
| Description: |
The purpose of mammography-based AI risk prediction software in individuals undergoing bilateral screening mammography is to predict the 5-year risk of individuals who will receive a diagnosis of breast cancer or develop breast cancer within the 5-year timeframe.
Breast cancer is the most commonly diagnosed non-skin cancer in U.S. women and remains a leading cause of cancer death. The American Cancer Society (ACS) estimates the average lifetime risk for a U.S. female is approximately 13.1 percent (about 1 in 8) with a lifetime probability of breast-cancer death of approximately 2.3 percent (about 1 in 43) (ACS, 2024). For 2024 the ACS projected approximately 310,720 new invasive female breast cancers, 56,500 cases of female ductal carcinoma in situ, and 42,250 female breast-cancer deaths in the United States, plus 2,790 invasive cases and 530 deaths in men. Screening confers the greatest absolute benefit in individuals most likely to develop breast cancer, since the number of deaths averted per woman screened scales with baseline risk. Mammography in average-risk women has demonstrated a reduction in breast cancer mortality of approximately 20 percent, with smaller absolute benefit in younger women whose cancer incidence is lower (Myers, 2015; Marmot 2012). Across the most recent decade of available data (2012–2021), invasive breast cancer incidence increased by approximately 1 percent per year, with a steeper trend in women younger than 50 (1.4 percent per year) than in women 50 or older (0.7 percent per year). As the incidence in younger populations continues to grow, conventional mammography will offer limited value for individual-level risk estimation in this group, as mammographic sensitivity is lower in younger women whose, typically, denser breast tissue obscures 2D imaging. Furthermore, established risk prediction models, calibrated principally on older cohorts and weighted toward lifetime-risk horizons, provide little discriminating signal at younger ages. Mammography for breast cancer is a controversial topic. Experts disagree on when to begin screening, when to end screening, and how often to screen. Compared with standard screening schedules, risk-based or personalized breast cancer screening has the potential to reduce false-positive recalls while detecting clinically significant cancers earlier.
Several breast-cancer risk prediction models are in routine clinical use, each developed to support a different decision point in the screening and prevention pathway. The Gail model (BCRAT) and the BCSC calculator are typically used at routine screening encounters to identify women whose 5- or 10-year risk crosses established thresholds for management decisions (i.e., chemoprevention discussion or supplemental imaging). Tyrer-Cuzick (IBIS) and the comprehensive-pedigree models (BRCAPRO, BOADICEA/CanRisk) are typically used in genetic-counseling and high-risk-clinic settings, where they support both estimation of residual lifetime risk and probability of carrying pathogenic variants in BRCA1/BRCA2 and related susceptibility genes. Collectively, these models share 3 important strengths. They draw on factors with established epidemiologic associations to breast cancer, they are inexpensive, non-invasive, and deployable from data routinely collected at clinical encounters, and they have been validated in large cohorts and incorporated into guideline-endorsed care pathways for decades. However, they also share 3 important limitations. First, individual-level discrimination is modest across all 5 models: reported area under the receiver operating characteristic curve (AUROC) values cluster between approximately 0.58 and 0.65, leaving substantial overlap between women who do and do not subsequently develop breast cancer at any given risk score. Second, all 5 models were developed predominantly in cohorts of European ancestry, and external validations have demonstrated calibration drift in non-European populations and in younger women. Third, the output horizons of the 5 models are heterogeneous: only the Gail model and BCSC produce 5-year invasive breast-cancer risk estimates on the scale used by the NCCN greater than or equal to 1.7% increased risk pathway, and of those 2, only BCSC incorporates mammographic image information (breast density). The remaining 3 models output 10-year, lifetime, or mutation-probability estimates that would not directly support the greater than or equal to 1.7% 5-year threshold without re-derivation.
The 5-year greater than or equal to 1.7% threshold itself is well established in breast-cancer risk management. It corresponds to the entry criterion for the National Surgical Adjuvant Breast and Bowel Project initiated Breast Cancer Prevention Trial Part 1 (NSABP P-1) tamoxifen prevention trial, which demonstrated a 49% short-term reduction in invasive breast cancer incidence in eligible participants randomized to tamoxifen, and it remains the canonical threshold above which clinicians initiate the standard Increased Risk management cascade (Fisher, 1998, Fisher 2005). For individuals who meet this threshold and are greater than or equal to 35 years of age, the cascade includes clinical encounter every 6 to 12 months, annual digital mammography with tomosynthesis, consideration of annual breast MRI (with contrast-enhanced mammography, molecular breast imaging, or whole-breast ultrasound as alternatives where MRI is contraindicated or unavailable), counseling on risk-reduction strategies, and periodic risk reassessment. Risk-reduction agents, such as tamoxifen, raloxifene, anastrozole, and exemestane are routinely offered for this threshold, with tamoxifen as the only agent indicated for premenopausal patients. In current practice, this means a 5-year breast-cancer risk estimate greater than or equal to 1.7%, produced by a tool whose output is calibrated to that horizon (i.e., Gail and BCSC), an individual is routed into the Increased Risk management cascade described above.
Breast Cancer Risk Prediction Models
Model: The Breast Cancer Surveillance Consortium (BCSC) Risk Calculator (BCSC, 2026) [Version 3 of the BCSC Risk Calculator extends the model to include BMI, extended (beyond first-degree) family history, and age at first live birth, with modestly improved discrimination (Gard, 2024). The v3 update is not yet reflected in current NCCN guidance.]
Description
Interactive tool designed to estimate 5-year risk of developing invasive breast cancer in those AFAB, including benign breast disease diagnoses and to estimate both 5-year and 10-year breast cancer risk. May have limited applicability for extensive family history beyond first-degree relatives.
Factors Included
5- and 10-year breast cancer risk calculations are based on 5 factors:
Benefits
Accessible online
Limitations
The calculator is NOT applicable to individuals who meet any of the following criteria:
Model: Gail Model (NCI, 2026)
Description
Individualized breast cancer risk assessment computed based on SEER-specific breast cancer risk data with inclusion of personalized risk factors. Provides both 5-year and lifetime risk assessment. 5-year risk assessment greater than or equal to 1.67% used to assess eligibility for a risk-reducing agent.
Factors Included
Benefits
Limitations
Model: International Breast Cancer Intervention Study (IBIS/Tyrer-Cuzick) (IBIS Breast Cancer Risk Evaluation Tool, 2026)
Description
Computerized model base on initial data from the United Kingdom Thames Cancer Registry 2005-2009. Attribution of risk based on family history data. Provides personalized breast cancer risk assessment based on individual risk factors and family history information. Both lifetime breast cancer risk (to age 85 in version 7+) and 10-year risk estimations are available.
Factors Included
Benefits:
Limitations:
Model: BRCAPro (Parmigiani, 1998)
Description
Bayesian carrier-probability model that uses Mendelian segregation analysis applied to a multi-generational family pedigree of breast and ovarian cancers. Penetrance and prevalence parameters for BRCA1/2 are derived from a systematic review of the published literature. Produces lifetime breast-cancer risk estimates derived from carrier-status modeling.
Factors Included:
Benefits
Limitations
Model: CanRisk Tool (BOADICEA v6) (Breast and Ovarian Analysis of Disease Incidence and Carrier Estimation Algorithm and CanRisk-Prostate model, 2026)
Description
Comprehensive risk prediction tool that integrates multi-gene panel results, PRS, multi-generational family history of breast and non-breast cancers, lifestyle and hormonal factors, mammographic breast density, and tumor pathology where available. Prospectively validated for both carrier-probability prediction and prediction of subsequent cancer risk. Produces lifetime breast cancer and ovarian cancer risk estimates.
Factors Included:
Benefits
Limitations:
Adapted from NCCN Guidelines v1.2026 Breast Cancer Risk Reduction (NCCN, 2026).
Clairity Allix5 (Clairity Breast)
Clairity Allix5 is an image-based deep-learning software device that generates a 5-year breast cancer risk prediction from a routine bilateral screening mammogram. The algorithm inputs the 4 standard 2D screening views (RCC, LCC, RMLO, LMLO) acquired from Hologic Selenia Dimensions or Lorad Selenia mammography systems and analyzes pixel-level imaging features and characteristics that correlate with future breast cancer development. Synthetic 2D images and images from non-Hologic acquisition platforms are not supported. The mammography-based AI algorithm provides a prediction of the percentage probability that the individual will receive a diagnosis of breast cancer or develop breast cancer within the 5-year timeframe following the screening mammogram.
Mirai
Mirai is a deep learning model that produces a personalized risk score that generates a 5-year breast cancer risk prediction from a routine bilateral screening mammogram. The algorithm inputs the 4 standard 2D screening views (RCC, LCC, RMLO, LMLO) acquired from Hologic Selenia Dimensions or Lorad Selenia mammography systems and consists of four modules: an image encoder, an image aggregator, a risk factor predictor, and an additive-hazard layer. During the risk factor predictor module, an individual's traditional risk factors (such as, age, weight, and hormonal factors) or predicted traditional risk factors are incorporated into the algorithm to allow for mammograms as the sole input. The mammography-based AI algorithm provides a prediction of the percentage probability that the individual will receive a diagnosis of breast cancer or develop breast cancer within the 5-year timeframe following the screening mammogram.
Regulatory Status
Clairity Allix5 (marketed as CLAIRITY BREAST; Clairity, Inc., Boston, MA) is a Software as a Medical Device (SaMD) authorized by the U.S. Food and Drug Administration on May 30, 2025, under De Novo classification DEN240047 (21 CFR 892.8500; product code SEZ; Class II with special controls) (CDRH of the FDA, 2024). It is the first device authorized in a new generic category that FDA defines as a radiological software device to predict future breast cancer risk. Mirai has not been approved or cleared by the U.S. Food and Drug Administration (FDA) and is for research purposes only until authorized by the FDA.
|
|
|
|
|
Policy/ Coverage: |
Effective January 1, 2027
Does Not Meet Primary Coverage Criteria Or Is Not Covered For Contracts Without Primary Coverage Criteria
Use of mammography-based AI risk prediction software for individuals undergoing bilateral screening mammography to estimate their 5-year risk of developing breast cancer does not meet member benefit certificate Primary Coverage Criteria that there be scientific evidence of effectiveness in improving health outcomes and is not covered.
For members with contracts without Primary Coverage Criteria, use of mammography-based AI risk prediction software for individuals undergoing bilateral screening mammography to estimate their 5-year risk of developing breast cancer is considered Not Medically Necessary or is investigational and is not covered. Not Medically Necessary or Investigational services are specific contract exclusions in most member benefit certificates of coverage.
Click the following link to view the specific criteria in InterQual®:
|
|
|
|
| Rationale: |
This evidence review was created in May 2026 with a search of the PubMed database. The most recent literature update was performed through May 01, 2026.
No published studies reported the validation of the mammography-based AI risk predication model, Clairity Allix5, only the pivotal study that led to the De Novo authorization. There are no RCTs, and no studies in which the Clairity Allix5 risk model was used to guide management decisions.
The pivotal study supporting authorization was a retrospective study evaluating discrimination and calibration of the device's 5-year breast cancer risk prediction (The CDRH of the FDA, 2024). The validation cohort comprised 77,511 mammographic examinations from 44,112 individuals across 5 U.S. clinical sites in Texas, California, and Georgia, with 1,723 exams (2.22%) followed by biopsy-confirmed breast cancer within 5 years. The validation cohort was independent of the algorithm's training data and included substantial representation of Black women (37.3% of validation exams). Discrimination was measured by the time-dependent cumulative-dynamic AUC at 5 years (AUC[5]) with a pre-specified performance goal of a lower 95% confidence interval bound greater than 0.64 (selected based on traditional risk model performance). Calibration was evaluated using a modified Greenwood-Nam-D'Agostino (GND) test against pre-specified tolerance bounds with a p-value threshold of less than 0.025 to declare predictions significantly within bounds. Both co-primary endpoints were met in the total study population: AUC(5) was 0.70 (95% CI 0.69 to 0.72) and the modified GND test for calibration achieved p less than .0001. The pre-specified secondary analysis evaluated performance after excluding patients with "current cancer," defined as a biopsy-proven breast cancer diagnosed within 90 days of a positive BI-RADS assessment (BI-RADS 0, 3, 4, or 5) on the index mammogram. In this secondary analysis, AUC(5) declined to 0.66 (95% CI 0.66 to 0.68). In the subgroup analyses, Allix5 met the discrimination performance goal in the total population but failed to do so when "current cancers" were excluded for individuals in the following subgroups: Black, Hispanic, aged 65 or older, and breast implants. Calibration remained within pre-specified bounds in the overall analysis that excluded individuals with current cancers (p=.0174), but the calibration plot demonstrated significant overestimation of risk in the highest decile of predicted scores (mean predicted 8.65% vs. observed 5.73%; calibration error 2.92%, 95% CI 2.29% to 3.55%). Acceptable calibration was demonstrated in most subgroups but not in Asian individuals, individuals younger than 50, individuals aged 65 or older, or individuals with breast implants. Limitations include, but are not limited to, heterogeneity in image acquisition, screening intervals, and follow-up duration given the retrospective design, the validation cohort was limited to directly acquired 2D images from two Hologic mammography systems, and that discrimination and calibration performance declined or fell below pre-specified thresholds in several clinically important subgroups when current cancers were excluded.
Schopf et al conducted a systematic review of deep learning artificial intelligence (AI) algorithms validated on real-world screening mammograms for prediction of future breast cancer risk from 16 studies (Schopt, 2024). All 16 studies used a retrospective design with risk prediction horizons across studies ranging from 1 to 5 years. Of the 14 studies that reported area under the receiver operating characteristic curve (AUC), image-only AI models demonstrated a maximum AUC range of 0.62 to 0.90 (median 0.72), while clinical risk factor–based comparators (breast density alone and traditional risk models including Tyrer-Cuzick, Gail/BCRAT, and the Breast Cancer Surveillance Consortium model) demonstrated a maximum AUC range of 0.54 to 0.69 (median 0.61). Combined image + clinical risk factor AI models (n=7 studies) demonstrated a maximum AUC range of 0.66 to 0.84 (median 0.73) with 6 of these studies reporting no significant improvement with the addition of clinical risk factors and the remaining study only showed an AUC improvement of 0.06. The Clairity Breast (Allix5) algorithm was not included in any of the 16 studies in this review. Limitations include, but are not limited to, that the underlying body of evidence consists of retrospective studies (8 of 16 with case–control or case–case designs prone to selection bias), only half of the included studies used cancer registry linkage to ascertain incident cancers, all studies were rated by QUADAS-2 as having high or unclear risk for bias or applicability concerns in at least one domain, comparisons by race and ethnicity were available in only 5 studies and were generally collapsed to White versus non-White owing to small minority subgroups, and no quantitative meta-analysis was performed.
Lowry et al conducted an updated systematic review of mammography-based AI models for future breast cancer risk prediction, extending the literature search beyond the prior Schopf et al. review through February 28, 2025, and included 41 studies (Lowry, 2026). Thirty-seven of 41 studies evaluated digital mammography algorithms, while only 3 (7%) evaluated digital breast tomosynthesis–based models. The United States was the most represented setting (17 of 41, 41%) followed by Sweden (7 of 41, 17%) and 9 of the 19 studies that reported race and ethnicity had cohorts comprising more than 75% White non-Hispanic women. Forty of the 41 studies reported area under the receiver operating characteristic curve (AUC), with values ranging from 0.61 to 0.84 and a median AUC of 0.71 overall. Median AUCs were similar across prediction horizons (0.71 for less than or equal to 2 years, 0.72 for 3-4 years, 0.71 for greater than or equal to 5 years) but differed by handling of cancers diagnosed at or near the index mammogram (median AUC 0.76 when index cancers were included vs. 0.68 when excluded), suggesting that a portion of the discriminatory signal reflects detection of prevalent disease on the index examination rather than prediction of future risk. Only 33% of the studies used true external validation populations with only 15% reported model calibration metrics. The Clairity Breast (Allix5) algorithm was not included in any of the studies for this review. Limitations include, but are not limited to, that the underlying body of evidence is dominated by retrospective studies with a high or unclear risk of bias across multiple QUADAS-2 domains, included studies disproportionately reflect digital mammography acquired with a single equipment manufacturer in U.S. and Swedish cohorts that overrepresent White non-Hispanic women, and no quantitative meta-analysis was performed.
Characteristics and findings of the Mirai validation studies, head-to-head comparator studies, and the prospective workflow study are described below.
Several retrospective validation studies reported on the mammography-based AI risk prediction model, Mirai, and examined the ability to predict 5-year breast cancer risk from a single screening mammogram (Avendano, 2024; Omoleye, 2023; Santeramo, 2024; Rothwell, 2025; Yala, 2021; Yoon, 2023; Park, 2024; Damiani, 2023). These studies cover a wide range of ethnicities that represent a diverse population and reported concordance indices and area under the receiver operating characteristic curves within the ranges denoted in the systematic reviews. Three nonrandomized studies compared mammography-based AI risk prediction models with traditional risk prediction models (e.g., Tyrer-Cuzick, BCSC, BCRAT, etc.) to assess the performance characteristics of each model (Arasu, 2023; Lehman 2022; Lamb, 2023). Lehman et al reported a higher AUC (0.68, 95% CI: 0.66 to 0.70) for the mammography-based AI risk prediction model compared with Tyrer-Cuzick (0.57, 95% CI: 0.54 to 0.60) and NCI BCRAT (0.57, 95% CI: 0.54 to 0.60) models. Additionally, the mammography-based AI risk prediction model, Mirai (8.6, 95% CI: 7.9 to 9.4), demonstrated a higher cancer detection rate (cancers detected per thousand patients screened) than either of the traditional models (Tyrer-Cuzick: 4.4, 95% CI: 3.9 to 4.9 and NCI BCRAT: 3.8, 95% CI: 3.3 to 4.3; p less than .001). Arasu et al. compared 5 mammography-based AI algorithms (Mirai, MammoScreen, ProFound AI, Mia, and the Globally Aware Multiple Instance Classifier) and the Breast Cancer Surveillance Consortium (BCSC) clinical risk model for prediction of 5-year breast cancer risk and demonstrated that all 5 mammography-based AI algorithms performed better than BCSC. Lamb et al. extended the head-to-head comparison to the supplemental MRI screening setting and identified individuals as high risk by the mammography-based AI risk prediction model, Mirai, which demonstrated a significantly higher cancer detection rate at screening MRI than patients identified as high risk with the Tyrer-Cuzick lifetime model (20.6/1,000 vs. 6.0/1,000, p less than .01) or the NCI BCRAT lifetime model (20.6/1,000 vs. 6.8/1,000, p=.04). One prospective study used the Mirai algorithm with a pre-calibrated 10th-percentile threshold applied to a 1-year breast cancer risk score and reported a significantly higher cancer detection rate in high-risk expedited women compared with non-high-risk women, as well as significantly shorter times to screening result, diagnostic evaluation, and biopsy in expedited patients versus high-risk controls (Chung, 2025).
Yala et al. developed and validated Mirai, a mammography-based deep learning model designed to predict 5-year breast cancer risk from a single screening mammogram (Yala, 2021). The model was trained on 210,819 mammograms from 56,786 patients at Massachusetts General Hospital (MGH) and validated on three independent test sets: an MGH held-out test set (25,855 mammograms, 588 followed by cancer), the Karolinska University Hospital test set in Sweden (19,328 mammograms, 1,413 followed by cancer), and the Chang Gung Memorial Hospital (CGMH) test set in Taiwan (13,356 mammograms, 244 followed by cancer) with all included mammograms acquired using Hologic equipment. Mirai achieved a concordance index (C-index) of 0.76 (95% CI, 0.74 to 0.80) at MGH, 0.81 (95% CI, 0.79 to 0.82) at Karolinska, and 0.79 (95% CI, 0.79 to 0.83) at CGMH. In a supplemental analysis excluding cancers diagnosed within 6 months of the screening mammogram, the C-indices declined to 0.71 to 0.74 across the three sites. Limitations include, but are not limited to, that the validation cohort was limited to mammograms acquired on Hologic equipment, C-indices were derived from analyses that applied a relatively narrow 6-month exclusion window of the index examination, the MGH validation cohort overlaps with the development cohort, race/ethnicity subgroups at MGH had a small sample size, and no prospective validation was performed.
Yala et al. extended the validation of Mirai to 7 hospital systems across 5 countries: MGH (USA), Novant (USA), Emory (USA), Maccabi-Assuta (Israel), Karolinska (Sweden), CGMH (Taiwan), and Barretos (Brazil) (Yala, 2022). The combined cohort comprised 128,793 mammograms from 62,185 individuals across the 7 sites, of which 3,815 were followed by a cancer diagnosis within 5 years with all mammograms acquired using Hologic equipment. Mirai achieved C-indices ranging from 0.75 (95% CI, 0.70 to 0.80) at Novant to 0.84 (95% CI, 0.81 to 0.88) at Barretos, with 5-year AUCs of 0.75 to 0.82. In a supplemental analysis excluding cancers diagnosed within 6 months of the screening mammogram, C-indices declined to 0.69 to 0.78 across the 7 sites. The model demonstrated similar performance in African American and White individuals in the Emory subgroup (C-indices 0.75 and 0.79, respectively). In a high-risk cohort analysis, Mirai-based guidelines significantly outperformed Tyrer-Cuzick lifetime risk-based guidelines at MGH (sensitivity 39.7% vs. 22.9% at matched specificity; p less than .001), and the improvement was preserved across the other 6 sites. Notable limitations are identical to those listed in the 2021 Yala et al study.
In 2023, the American College of Radiology (ACR) published clinical practice management recommendations for breast cancer screening for Women at higher-than-average risk (Monticciolo, 2023). Regarding risk prediction models that encompass image-based data and artificial intelligence the ACR states "Deep learning models using mammographic images suggest that images contain risk indicators that may not be perceptible to the human eye. In some studies, performance was found to be superior to traditional models, but additional work is needed." Based off their previous recommendations, that emphasized the importance of risk assessment at an age early enough to allow for supplemental screening the ACR makes the following claim "The identification of increasing numbers of genetic mutations, and their associated risks, suggests that risk assessment should take place by age 25 for all women, especially Black women and those of Ashkenazi Jewish descent" (Monticciolo, 2018).
The National Comprehensive Cancer Network guidelines for breast cancer screening and diagnosis v.1.2026 states the following regarding mammography-based AI algorithms: "Studies with AI image-only based models were found to have a higher median AUC performance (0.72; range, 0.62 to 0.90) compared to breast density models or clinical risk prediction tools such as Gail or Tyrer-Cuzick (0.61; range, 0.54 to 0.69). For an individual greater than or equal to 35 years of age with a 5-year risk greater than or equal to 1.7% per the Gail model or as calculated by an imaging-based risk assessment model, the NCCN panel encourages breast awareness and recommends a clinical encounter every 6 to 12 months and annual digital mammography with tomosynthesis, to begin at the age identified as being at increased risk by either model" (NCCN, 2026).
Some ongoing trials that might influence this review are listed below.
|
|
|
|
| CPT/HCPCS: | |
|
|
|
| Group specific policy will supersede this policy when applicable. This policy does not apply to the Wal-Mart Associates Group Health Plan participants. | |
| CPT Codes Copyright © 2026 American Medical Association. | |