Breast Imaging Patient Triage

Purpose

Risk stratify patients in a breast imaging screening workflow for triaged scheduling based on their risk of breast malignancy particularly in a time of limited clinical availability due to COVID-19 limitations

Tag(s)

Breast cancer, screening, scheduling, risk assessment, mammography

Panel

Non-Interpretive, COVID-19

Define-AI ID

20160002

Originator

COVID-19 Sub-Panel
Lead  Christopher McAdams

Panel Chair

Ben Wandtke, Woojin Kim

Panel Reviewers

Chuck Kauffman, Andrea Borondy Kitts

License

Creative Commons 4.0 
Status Public Comment
                               

Clinical Implementation


Value Proposition

COVID-19 has disrupted the full spectrum of cancer care. In particular, cancer screening has been delayed including follow-up imaging and procedures for potentially malignant findings. The first wave of COVID-19 cases has peaked and healthcare facilities are now able to schedule regular healthcare procedures. Algorithms to risk stratify patients will inform scheduling priority for delayed screening, follow-ups, and unperformed breast biopsies.  

Narrative(s)

Given that breast imaging exams are performed and interpreted entirely on an outpatient, non-emergent basis, a number of patients who have been recalled from screening for an additional imaging evaluation (BI-RADS 0) (e.g., a finding needs further workup or a technical repeat of imaging), those who are being followed for probably benign finding(s) (BI-RADS 3), and potentially those for whom biopsy or other appropriate action (BI-RADS 4 or BI-RADS 5) was recommended have had significant postponement of their follow-up imaging and care due to COVID-19-related scaling down and cessation of outpatient, non-emergent services. This has resulted in a significant backlog of patient exams and, in many instances, delayed beyond the expected norm due to unforeseen circumstances of the COVID-19 pandemic. While BI-RADS category 3 findings have a less than 2% likelihood of being caused by a breast malignancy, recent work highlights the importance of short term follow-ups, especially at the beginning of the prescribed follow-up period1. As practices enter staged recovery from the initial peak pandemic phase, patient navigators, business managers, and schedulers need assistance in optimizing their workflows, staffing, and schedules with a priority towards those women at greatest risk of having an undiagnosed breast malignancy, especially those breast malignancies at risk of being particularly aggressive or rapidly morbid. 


Workflow Description

Compile a list of breast cancer screening (BI-RADS 0), follow-up patients (BI-RADS 3), and patients with BI-RADS categories 4 and 5 that have pending, delayed, or deferred exams. Most screening programs have incorporated either the use of an electronic medical record with pertinent demographic information and risk factors and/or breast-specific reporting software incorporating many of the elements required to conventionally assess breast cancer risk averaged over the course of a woman’s lifetime (e.g. Tyrer-Cuzick risk calculator). Non-image related risk assessment variables include sex, age, body mass index (BMI), age at menarche, hormone replacement use, genetic mutation status (e.g. BRCA1, BRCA2, ATM, etc.), Ashkenazi Jewish descent, ethnicity (black vs non-black, etc.), prior history of mantle radiation therapy, prior personal history of breast and/or ovarian cancer, family history of breast and/or ovarian cancers (2 or more 1st-degree relatives), history of having a high-risk breast lesion (e.g. ADH, lobular neoplasia, papillomas), breast density (AI-quantified or otherwise), pregnancy history and status, and high anxiety over any delay in their breast care.  

The risk factor demographic data would be ingested into the AI engine then returned with a priority level attached to allow the end-user to order the patient list and begin setting up follow-up appointments. 


Within the scheduling system (HIS/RIS/EHR):

  1. An API call out to an AI engine that will assign a Priority value to each open item.

  2. Priority value can be used to sort the patient list for scheduling.

From outside of the scheduling system:

  1. A list of patients can be extracted to a CSV file with the data points and patient identifiers (Medical Record Number (MRN)).

  2. File sent into the AI engine and returned with a Priority Value.

 

An additional option would be to extract a patient's current COVID-19 status (from the laboratory records) if that information will influence scheduling or workflows (i.e., a designation of a COVID-19-negative workflow vs COVID-19-positive workflow or clinic sites). 


While the unforeseen constraints and rationing of resources and limited availability of outpatient imaging services have been highlighted due the medical communities’ experiences with COVID-19, these AI models are potentially also beneficial in times in which we are not operating under pandemic conditions (for example, in clinical areas with critical or limited access to breast health programs, radiologist, and/or mammography technologist staffing, etc).

Additional potential risk identifying data could potentially be ascertained by an AI tool by analyzing the need for follow-up for those called back from screening (BI-RADS 0) or the finding(s) being followed (BI-RADS 3). Previous DSI AI Use cases have addressed this sort of function (i.e., REF 2 ,REF 3 , REF 4 ). 

If a simple AI model were being considered based on systems factors (availability of data, etc.) then this might be limited to the assembly of the aforementioned demographic risk factors with AI computed estimate of risk to be used in conjunction with scheduling to ensure those at estimated higher risks of breast cancers be scheduled with some element of priority as schedule slots are available across the enterprise. 

 

Figure 1. Simple AI Model

 

The simple AI model would entail feeding in the list of follow-up patient demographics into the AI engine. The engine would evaluate the risk factors and output a priority assignment to the patient.

 

Advanced AI Model

A more novel and valuable model (albeit, more complex) would integrate the aforementioned demographic factors with an independent analysis of lesion(s) in question for their likelihood of malignancy, resulting in a summative overall risk estimate of that individual patient’s risk of malignancy per that lesion. 

 

The advanced model would build on the simple model and add the imaging element to the prioritization.  By scanning the image, the engine can pick up on findings that would elevate the patient’s prioritization.


Figure 2. Advanced AI Model

 

Additionally, we could build in a feedback element where the scheduler or physician could provide feedback on the prioritization if they promote or demote the patient. Specific feedback about which data elements caused the user to change prioritization could drive the engine’s machine learning for retraining.


Workflow

When a patient is moved up or down in prioritization the system prompts for feedback, “What elements caused you to move this patient (Up/Down)?”.  Users then would select one of the data points available on the account. This feedback would go back into the engine to tune the AI model for accuracy.


 


Data Elements


 Inputs

 

Data Element

Data Type

Description

Notes

BI-RADS Category

Categorical

Indicates BI-RADS category at last exam


BI-RADS 0 findings (RDE 799/800/801/204/205)




Non-circumscribed masses

Categorical

Indicates the patient has non-circumscribed masses. (Y/N)


Masses + Calcifications

Categorical

Indicates the patient has breast mass(es) and calcification(s) (Y/N)


Calcifications

Categorical

Identifies the patient has breast calcifications (Y/N)


Asymmetries

Categorical

Identifies asymmetries (developing asymmetries, global asymmetries, focal asymmetries)


BI-RADS 3 findings




Oval circumscribed masses

Categorical

Identifies the presence of oval circumscribed masses (Y/N)


Complicated cysts

Categorical

Identifies the presence of complicated cysts (Y/N)


Microcalcifications

Categorical

Identifies the presence of microcalcifications (Y/N)


Asymmetries

Categorical

Identifies asymmetries (global asymmetries, focal asymmetries)


COVID-19 Status

Categorical

Patient's COVID-19 status (active, recovered, unknown, negative test result within last week, etc.)


Non-imaging inputs




Duration of delay since recommended call back (BI-RADS 0), from recommended follow-up exam (BI-RADS 3), or since the last BI-RADS 4 or 5 exam

Time

Duration of delay since recommended call back (BI-RADS 0), from recommended follow-up exam (BI-RADS 3), or from the last BI-RADS 4 or 5 exam


Sex

Categorical

Patient's sex


Age

Numeric

Patient's age


Age of menarche

Numeric

Age of menarche


Age of first live birth

Numeric

Age of first live birth


Genetics status

Categorical

Indicates associated genetic status' (BRCA1/2 or carrier; other genetic mutations)


Ashkenazi descent

Categorical

Ashkenazi descent (Y/N)


Prior history of breast or ovarian cancer

Categorical

Prior history of breast or ovarian cancer (Y/N)


Family history of breast cancer

Categorical

Family history of breast cancer (Y/N)


Hormone replacement therapy use

Categorical

Patient use of hormone replacement therapy (Y/N)


Risk assessment calculator (such as the Tyrer-Cuzick Risk Model)

Numeric

Predict a percent likelihood of lifetime malignancy (See RDE 783, 784, 785)


Prior high-risk lesion (from prior biopsy)

Categorical

Presence of a prior high-risk lesion (Y/N)


Breast density

Categorical

Quantified/non-quantified (See RDE 341)


Ethnicity

Categorical

The ethnicity of the patient. American Indian or Alaskan Native, Asian, Black or African American, Hispanic or Latino, Native Hawaiian or Other Pacific Islander, White.


Pregnant

Categorical

Pregnancy status and association with BI-RADS. BIRADS 0 and pregnant, pregnant


High patient anxiety over delay in care

Categorical

(Y/N)


 

Primary Outputs


BI-RADS Category

Definition

BI-RADS category (0, 3 (sub-stratify 6 mo follow-ups), 4, 5)

Data Type

Categorical

Value Set

REF 5 

Units

N/A


Relative Risk Score

Definition

Relative risk score based on risk factors

Data Type

Numeric

Value Set

N/A

Units

N/A


COVD-19 Status

Definition

Patient’s COVID-19 status

Data Type

Categorical

Value Set

  • Active

  • Recovered

  • Unknown

  • Negative test result within last week

Units

N/A

 

 

 

Future Development Ideas


  1. Optimize risk score estimate function based on results as received (machine learning) - stability, biopsy results, interval cancers, etc. (see Figure 3. Future State Feedback AI Model below). Future state feedback would feedback the final patient diagnosis after a follow-up appointment. The final diagnosis would be correlated to other demographic factors to drive evaluating additional data points to be added to the prioritization engine.

  2. Integrate with breast clinic open appointments schedule for all breast clinical sites across the enterprise, populate with patient name and contact info by triage order with the earliest appointment for the first patient on the triage list, etc.  

  3. Generate automatic messaging to ordering clinicians and for patients with their scheduled appointment date and time, instructions on coming in for screening, and the option to change appointment time or date or opt-out of screening. 

 

Figure 3. Future State Feedback AI Model

References


  1. Berg WA, Berg JM, Sickles EA, Burnside ES, Zuley ML, Rosenberg RD, Lee CS. Cancer Yield And Patterns Of Follow-Up For BI-RADS Category 3 After Screening Mammography Recall In The National Mammography Database Radiology. 2020;296(1):32-41. doi: 10.1148/radiol.2020192641

  2. https://www.acrdsi.org/DSI-Services/Define-AI/Use-Cases/Breast-Lesion-Malignancy-Evaluation

  3. https://www.acrdsi.org/DSI-Services/Define-AI/Use-Cases/Calcification-morphology-and-follow-up

  4. https://www.acrdsi.org/DSI-Services/Define-AI/Use-Cases/Classifying-Suspicious-Microcalcifications

  5. https://www.acr.org/-/media/ACR/Files/RADS/BI-RADS/Mammography-Reporting.pdf