Assessment of a Risk-Based Approach for Triaging Mammography Examinations During Periods of Reduced Capacity

Key Points Question How can imaging facilities optimize the number of breast cancers detected during periods of reduced capacity using clinical indication and individual characteristics? Findings In this cohort study including 898 415 individuals with 1 878 924 mammograms, 12% of mammograms with very high and high cancer detection rates accounted for 55% of detected cancers, while 44% of mammograms with very low cancer detection rate accounted for 13% of detected cancers. Meaning These findings suggest that triaging individuals most likely to have cancer detected during periods of reduced capacity could result in detecting the most cancers while performing the fewest examinations compared with a non–risk-based approach.


Introduction
The coronavirus disease 2019 (COVID- 19) pandemic has profoundly impacted the US health care system, including breast cancer screening, surveillance, and diagnostic services. [1][2][3][4] To rapidly respond to county and state stay-at-home orders during the initial phase of the pandemic, radiology facilities cancelled or reduced nonurgent services, including breast cancer screening and surveillance, and some diagnostic imaging, which allowed health care personnel to shift to pandemic-related efforts and preserve personal protective equipment. [5][6][7][8][9] The precipitous drop in screening and diagnostic imaging 1,4 likely reduced the number of breast cancers diagnosed. 10,11 As COVID-19 case rates continue to increase in waves, radiology facilities face significant scheduling challenges in addressing the backlog of postponed mammograms, reduced staff, and reduced number of mammogram appointment times required to maintain physical distancing and safety protocols, resulting in markedly diminished imaging availability. 12,13 During periods of reduced capacity or in settings with limited mammography availability, facilities, individuals and their physicians who are determining the urgency of mammography should consider the probability that breast cancer will be detected. Several professional associations have posted guidance for scheduling individuals for breast imaging services during the COVID-19 pandemic based on expert opinion. [13][14][15] For example, the Society of Breast Imaging 12 suggests prioritization, from highest to lowest urgency, of patients who need imaging to inform breast cancer surgery, imaging for percutaneous breast biopsy, diagnostic work-up of an abnormal screening examination, short interval follow-up or diagnostic imaging for nonurgent symptoms, and last, screening with possible further prioritization by breast cancer risk. These recommendations do not specifically address how breast cancer risk factors should be considered nor address individuals with a personal history of breast cancer.
We used data representative of the US population from the Breast Cancer Surveillance Consortium (BCSC) 16 to develop an algorithm that radiology facilities could use to optimize triaging of individuals for mammography examinations during periods of decreased mammography capacity, such as during surge periods of the COVID-19 pandemic.

Study Setting and Cohort
The BCSC registries and the Statistical Coordinating Center received institutional review board approval for active or passive consenting processes or waiver of consent to enroll participants, link and pool data, and perform analysis. All procedures adhered to the Health Insurance Portability and Accountability Act. Registries 18 and link to state or regional cancer registries and pathology databases for complete capture of breast cancer diagnoses and benign breast disease. 17 The study population included 1 878 924 mammograms from 898 415 individuals interpreted by 448 radiologists from 2014 to 2019 across 92 facilities. We excluded mammograms performed within 6 months after a breast cancer diagnosis because they were likely associated with treatment planning.

Characteristics of Individuals and Mammograms
Age, first-degree family history of breast cancer, breast symptoms, time since last mammogram, and breast cancer history were collected from self-administered questionnaires at the time of mammography or extracted from electronic health records. Self-reported race/ethnicity was categorized as White, Black, Asian or Pacific Islander, Hispanic/Latina, and other or mixed race/ ethnicity. Breast cancer history was also obtained by linkage with pathology databases and state or regional cancer registries. Diagnoses of high-risk breast lesions, defined as atypical ductal or lobular hyperplasia or lobular carcinoma in situ, were obtained from pathology databases and cancer registries (for lobular carcinoma in situ).
Clinical indication for the mammogram was assigned by the radiologist or technologist as screening, additional evaluation of a recent mammogram, short-interval follow-up, or diagnostic for clinical signs or symptoms. We further subdivided indications for clinical symptoms into presence of a lump, presence of other symptoms, and unknown. We classified surveillance mammograms (ie, asymptomatic mammograms) among individuals with a breast cancer history using a previously described algorithm with slight modifications. 19 Mammograms with a nonspecific diagnostic indication (66 382 mammograms [3.5%]) were classified as short-interval follow-up if the most recent prior mammogram was within 3 to 13 months, with a final Breast Imaging Reporting and Data System (BI-RADS) 20 assessment of 3 (ie, probably benign finding) after all imaging work-up was performed; as an additional evaluation of recent mammogram if the most recent prior mammogram occurred within 3 months and had an initial assessment of 0 (ie, need additional imaging evaluation), 3, 4 (ie, suspicious abnormality), or 5 (ie, highly suggestive of malignant neoplasm), or occurred within 3 to 6 months and had an initial assessment of 0, 4, or 5; or otherwise as a diagnostic mammogram for clinical signs or symptoms. Radiologists categorized BI-RADS breast density 20 at the time of clinical interpretation as almost entirely fatty, scattered fibroglandular densities, heterogeneously dense, or extremely dense.

Mammography Outcomes
Breast cancer diagnoses were obtained by linking with pathology databases and state or regional cancer registries. 21 The final BI-RADS assessment was considered positive if it had an assessment of 4 or 5 and negative if the assessment was 1, 2, or 3. 20 The cancer detection rate was calculated as number of mammograms with positive final assessment and invasive carcinoma or ductal carcinoma in situ diagnosed within 90 days divided by total number of mammograms.

Statistical Analysis
We summarized characteristics of the study population by breast cancer history. We estimated cancer detection rate by individual and mammographic characteristics, separately by breast cancer history, and estimated 95% CIs using generalized estimating equations with a working independence correlation structure to account for correlation among mammograms within the same facility. 22 Classification and regression trees (CART) 23 were used to identify factors most strongly associated with cancer detection rate. We fit 4 models: screening mammograms in individuals

JAMA Network Open | Oncology
Risk-Based Approach for Triaging Mammography Examinations During Periods of Reduced Capacity without a breast cancer history, surveillance mammograms in individuals with a breast cancer history, and diagnostic mammograms separately by breast cancer history. All models included age group (<40, 40-49, 50-59, 60-69, Ն70 years), time since last mammogram (first mammogram, 1 year, 2 years, Ն3 years), first-degree family history of breast cancer, and breast density as potential predictors. Models for diagnostic mammograms also included clinical indication as additional evaluation of a recent mammogram, short-interval follow-up, or clinical signs or symptoms subdivided by lump, other symptoms, and unknown symptoms. Models for individuals without a breast cancer history also included history of a high-risk lesion. Models for individuals with a breast cancer history also included years since diagnosis (<5, 5 to <10, Ն10, unknown).
We set a maximum tree depth (ie, number of levels in the decision tree) of 10 and a maximum number of leaves (ie, final subgroups) of 10. We required at least 10% of the sample in each leaf. The maximum tree depth achieved was 5, the maximum number of leaves was 7, and pruning was not required. To internally validate the model, we split the study population into 2 random samples, fit the CART models to each sample, and compared the results to the overall model; we found no clinically meaningful differences.
We grouped the final subgroups from the 4 models into 5 risk groups: very high (>50), high         BI-RADS density and family history were selected for some models, but differences by these 2 risk factors were not clinically meaningful, and subgroups were combined into the same risk groups. The

JAMA Network Open | Oncology
Risk-Based Approach for Triaging Mammography Examinations During Periods of Reduced Capacity cancer detection rate was considered very high or high for clinical indications of additional evaluation of a recent mammogram or diagnostic work-up of a breast lump in all individuals, diagnostic work-up of symptoms other than lump in individuals with a breast cancer history, and short-interval follow-up or diagnostic work-up of symptoms other than lump in individuals aged 60 years or older without a breast cancer history ( Table 3). The cancer detection rate was considered moderate for surveillance mammograms in individuals diagnosed with breast cancer 10 or more years prior, short-interval follow-up examinations in individuals aged 70 years or older with a breast cancer history, screening examinations in individuals with a history of a high-risk breast lesion, short-interval follow-up or diagnostic work-up of symptoms other than lump in individuals younger than 60 years, and surveillance examinations in individuals less than 10 years since diagnosis. The cancer detection rate was considered low for surveillance mammograms in individuals with a breast cancer history diagnosed less than 10 years prior, first screening mammograms and screens performed more than 1 year since last mammogram in individuals aged 50 years or younger, short-interval follow-up examinations in individuals younger than 70 years with a personal history of breast cancer, and annual screens in individuals 70 years or older. reduced capacity, such as that which occurred during the start of the COVID-19 pandemic. Clinicians could use our results to counsel individuals about how urgently they should seek breast imaging based on their breast symptoms, breast cancer history, age, and time since last mammogram. We demonstrate that triaging individuals at highest risk of having cancer detected could result in detecting the most cancers while performing the fewest examinations compared with a non-riskbased approach. For example, a non-risk-based approach resulted in 11.5 cancers detected per 1000 mammograms (corresponding to the overall cancer detection rate in our study) while a risk-based approach limiting to the 12.1% of mammograms with high or very high risk of cancer detection detected 55.0% of cancers and resulted in at least a 3-to 10-fold greater cancer detection rate (36-122 cancers per 1000 mammograms). In contrast, the cancer detection rate for the 44.2% of mammograms with the lowest risk was 3.8 cancers or less per 1000 and accounted for 13.1% of detected cancers. The low cancer detection rate in these individuals should be considered, along with patient preferences, when deciding about the safety of postponing imaging owing to limited capacity, such as during pandemic surges when individuals may also experience risks and anxiety about contracting an infectious disease.
The American College of Radiology posted a "Return to Mammography Care" toolkit, 24 encompassing general resources for facilities, including pamphlets and letter templates to reassure patients regarding COVID-19 safety measures. Our results and triaging flowcharts complement this resource by providing direct, detailed evidence that supports risk-based scheduling of individuals most likely to have a cancer detected based on recent data from more than 1.8 million mammograms interpreted by more than 400 radiologists from more than 90 imaging facilities. In contrast to Society of Breast Imaging guidance based on expert opinion, 12 our data-driven analysis suggests a different triage order for certain groups. For instance, while the Society of Breast Imaging recommends prioritizing short-interval follow-up examinations over screening examinations, we found that woman with breast cancer diagnosed more than 10 years prior undergoing surveillance mammography and individuals with a history of high-risk lesions undergoing screening had higher cancer detection rates than most individuals requiring short-interval follow-up.
Infrastructure should be developed and resources should be allocated to facilitate risk-based algorithm implementation. Challenges include population-based risk factor ascertainment to triage those already scheduled, and software for real-time implementation when individuals call to schedule their examinations. Electronic health records could be adapted to automate prioritization, given they typically contain risk factor information. As part of our ongoing project, we are developing scripts that schedulers can use to encourage individuals with the highest risk to receive care as soon as possible despite the pandemic, and to reassure individuals whose care may be deferred owing to their low risk.
While our analysis was motivated by limited mammography capacity at the start of the COVID-19 pandemic, our results are applicable to other situations and clinical settings. For example, several medical centers have experienced cyberattacks that limited access to electronic health records. 25,26 Some centers had to greatly reduce clinical care until records were back online. The strategy of offering and promoting medical services based on risk is relevant for other cancer screening tests, including cervical cancer screening, low-dose computed tomography lung cancer screening, prostate cancer screening, and colonoscopy. [27][28][29][30][31][32] Additionally, our approach may be applicable to other diseases, such as hypertension and diabetes, for which visits to clinicians may be limited by the pandemic or in regions where health care services and clinicians are in short supply.
Using data-driven strategies to identify subgroups at highest risk of disease during times of limited capacity may lead to less downstream strain on an already taxed health care system. When capacity is limited, health care systems must ensure that the patients with the highest risk receive care.