AMBER (AI-based Model for Breast Cancer Risk Assessment) for TNBC

Breast cancer is the second most common cancer among women in the U.S.  About 10-15% of all breast cancers are found to be triple negative breast cancer (TNBC).  Unlike other types of invasive breast cancer, TNBC tends to grow and spread faster, has fewer treatment options, and tends to have a worse prognosis.  Given its aggressivity and lack of response to hormone therapy or targeted HER2 drugs, it is often treated with neoadjuvant chemotherapy (NACT) followed by surgery.  Unfortunately, 60-70% of patients will not have a pathologic completed response following NACT.  Similarly, 5-10% of the patients will end up having disease recurrence within 3 years of surgery. 

The numbers are grim for Black and Hispanic patients who tend to have a higher incidence of and often more aggressive TNBC with worse outcomes compared to White patients.  While part of this has to do with socio-economic differences, recent evidence has suggested molecular differences in the TNBC for such populations.  Furthermore, risk calculators like Oncotype DX, which is primarily used for ER+ HER2- breast cancer, have been shown to be less accurate in under-served populations like Black women. 

The AMBER [Artificial Intelligence-based Model for Breast Cancer Risk Assessment] for TNBC will address three important problems:

1. Currently, no molecular tests for TNBC exist to identify patients at risk of recurrence or predict non-response to treatments like chemotherapy or immunotherapy. Accurate patient selection is vital to ensure TNBC patients receive beneficial therapy while sparing those unlikely to respond from associated toxicities, directing them toward alternative treatments. 
2. While molecular tests such as Oncotype DX are also extremely expensive costing more than $4,000, the AMBER tool works with H&E slides, which are routinely used in the medical workflow and whose cost is usually less than $20
3. In the TNBC domain, the density of tumor-infiltrating lymphocytes (TILs) has shown to be a potential prognostic and predictive marker. However, assessing TIL density by human readers is challenging due to high intra-observer variability and the inability to account for TIL interactions with various cell groups in the tumor microenvironment. Image analysis and AI offer objective quantification of TILs and enable the characterization of spatial patterns not easily discerned by human readers. Additionally, AMBER will be validated on a diverse range of patients from various ethnicities (e.g., Asian, Black, Hispanic, White) to demonstrate equitable prognostic and predictive capabilities across different populations.

The increased accuracy of the AMBER tool, the characterization of complex patterns, and the dramatic cost savings (98% lower than current tests), make it possible for AMBER to be available to low- and middle-income countries where cancer diagnoses are rising.  

There is increasing evidence suggesting that interaction of tumor cells with immune cells has a significant association with the likelihood of disease progression and influences tumor development, invasion, metastasis, and patient outcome. From immune cells, tumor-infiltrating lymphocytes (TILs) have been reported by several independent studies as outcome predictors in diverse solid tumor types. In the TNBC domain, density of TILs has been shown as a potential prognostic marker; unfortunately, this metric is difficult to assess by human readers, is prone to high intra-observer variability, and ignores the interaction between TILs and different cell groups in the tumor microenvironment. In our previous work, we developed and validated an artificial intelligence (AI)-based classifier that characterizes the spatial architecture of TILs in digitized H&E samples with residual TNBC after NACT aiming to find the individuals that are at higher risk of death and recurrence. This approach was trained and validated on digitized H&E samples from 92 patients with TNBC. Patients in the validation set (n=47) identified as “high risk” had a significantly shorter survival time. The model was prognostic of overall and disease-free survival with hazard ratios of 2.57 (95% confidence interval (CI):1.07-6.16, p=0.03) and 2.38 (CI:1.01-5.62, p=0.04), respectively. Additionally, have shown that population-specific computational pathology models are more accurate for risk stratification of breast cancer compared to population-agnostic models (Black and South Asians). In the present project, we seek to build on our promising previous results to further develop and validate a computational pathology platform for automated characterization of the tumor micro-environment on digital pathology images of TNBC. Our approach captures not only TIL density but also enables the quantitative characterization of the spatial arrangement and interplay of multiple different primitives and cell types simultaneously (e.g., TILs, neutrophils, cancer cells). Additionally, our approach will integrate into the analysis other biomarkers such as the organization of collagen fibers, which has also shown prognostic value in breast cancer. This could potentially provide both a more comprehensive portrait of the tumor microenvironment and a more accurate prognostic assessments of disease outcome. The platform will be used to discover morphology feature differences between different populations (Asian, Black, Hispanic, and White patients) by analyzing H&E tissue samples and construct population-specific computational histology imaging models. These models will identify 1) patients at higher risk of death and recurrence and 2) patients who will benefit from treatments such as chemotherapy and immunotherapy.

Innovation: Unlike many AI approaches that are opaque black box models such as deep learning, our approach distinguishes itself with its interpretability, which arises from its handcrafted nature that facilitates more accessible individual feature analysis. Since our approach is modeled on H&E images, it will be “non-destructive of tissue” and significantly lower in cost (~$30-50) compared to extant assays like Oncotype DX (>$4,000). Critically, we will seek to intentionally develop and validate our approach in a way that ensures that the test is equitably prognostic and predictive across populations with different ancestry (Asian, Black, Hispanic, White).

The AMBER Model will help:  Any TNBC patient who is Black, Hispanic or Asian American; healthcare patients in LMI countries who will have access to a more accurate, less costly tool to determine the best treatment approach to their TNBC diagnosis. 

Breast cancer ranks as the most prevalent cancer and the second leading cause of cancer-related deaths among women in Georgia. The absence of molecular tests for TNBC poses a challenge for identifying high-risk patients and guiding treatment decisions. This study focuses on creating computational tools to assess the tumor immune microenvironment in Stage I-III TNBC patients, aiming to distinguish those at elevated risk of death or recurrence and identify individuals benefiting from chemotherapy and immunotherapy. Collaboration with Emory University Hospital will involve validating our models with a substantial sample of Georgia patients.

A 2011 study, utilizing Georgia Comprehensive Cancer Registry data, compared invasive breast cancer frequencies and triple-negative tumor presence. Stratifying by race and age group revealed higher breast cancer incidence rates in White women but increased mortality in Black women, especially among the younger demographic. Additionally, triple-negative tumors are twice as prevalent in Black populations. While causative factors remain unclear, lifestyle, socioeconomic status, and access to care, alongside biological considerations, are posited contributors. Employing computerized image analysis of TNBC histological samples enables the examination of tissue morphology and architecture, potentially uncovering biological distinctions between populations. This project extends its scope to scrutinize morphological differences among racial groups and develop distinct prognostic models, offering more tailored disease outcome predictions and potentially benefiting underserved communities.

The team members working on the AMBER are diverse:

Principal Investigator:  German Corredor Prada – US Resident from Colombia

Postdoc Fellow:  Sahar Almahfouz Nasser – Arab

Ph.D. Student:  Satvika Bharadwaj – South Asian – Indian

Pathologist/Data:  Tilak Pathak – South Asian – Nepal

Finance Manager:  Brittany Brown – White – United States

80% of our team are from or are connected to family members from the global south or LMI countries.  Our approach to this research is informed by our lived experience and those of our family and friends as well.

In addition, our lab operates in Atlanta, Georgia which is a highly diverse, large urban center.  The breadth of Emory’s Clinical footprint in the Atlanta area means patients from every socioeconomic status, and ethnic background, and stage of life come to Emory to find care.  The Emory Health System also includes two large public hospitals in metro Atlanta that provide care to the indigent. 

We have access to the Emory Clinic which conducts clinical trials and draws diverse participants from across metro Atlanta. 

We are at the concept stage because we have not yet completed the creation of specific models of TILS for Black, Asian and Hispanic patients.  Once we have models, we can then test them for accuracy in determining the prognosis for each patient. 

Once our models accurately indicate prognosis of over-all and disease-free survival and predictive response to therapy, we will then move into the prototype stage to develop our AMBER tool to be easily implemented into clinical settings across the globe.  

I have the lab personnel and resources I need to complete the specific computational models for each category (White, Black, Hispanic, Asian) for AMBER, but I will need resources to conduct clinical trials to test the models for accuracy.  Once that stage is complete, I will need partners who can help me to operationalize the diagnostic tools I develop.  Specifically, I will need collaborative partners who can help me create diagnostic tools that are easy to implement and easy to distribute within healthcare systems across the globe.  The goal will be to operationalize these diagnostic tools in a manner that is easily adoptable in LMI as well as wealthier nations.    

AMBER is innovative because it uses the power of AI to generate accurate characterization of the tumor microenvironment across different racial groups (Black, Asian, Hispanic).  In addition, AMBER uses AI to read digitized biopsy slides and compute risk scores for predicting 1) the likelihood of developing recurrence (prognostic) and 2) the potential benefit of chemotherapy or immunotherapy (predictive). Physicians can leverage these risk scores to inform and guide treatment decisions.

 The specific aim of our initial research is to:

Aim 1: Identifying morphological differences between populations (Asian, Black, Hispanic, and White patients) with TNBC.

Aim 1A: Extracting features from the immune spatial organization within the tumor microenvironment in digitized H&E samples.

Aim 1B: Extracting features from collagen fibers.

Aim 1C: Conducting a statistical analysis comparing differences among the features extracted from different populations. 

Aim 2: Developing population specific models that use features extracted from histological images for predicting outcome and therapy benefit.

Aim 3: Validating the population specific models in a multi-institutional setting.

We believe the increased accuracy and decreased cost of the AMBER tool will make it a standard diagnostic tool across healthcare systems throughout the globe.  In addition to providing more accurate predictions about survival and therapeutic benefit, health outcomes will also improve as practitioners will have accurate information to determine who might benefit from chemotherapy/immunotherapy and who can avoid the difficult treatment altogether.

The AMBER model creates far less expensive, far more accurate prognostic and predictive tools to ensure that TNBC patients receive appropriate and effective treatment regardless of their race or geographic location. This approach helps practitioners deliver much more personalized and appropriate care to patients with TNBC, including improving the accuracy of prognosis for recurrence, withstanding NACT treatment.  

Cancer rates are increasing worldwide with most new cancer diagnoses coming from low- and middle-income countries (LMIC).  For these LMI countries, access to the latest technological breakthroughs in the treatment of cancer are difficult to access.  This is an especially daunting fact when evidence indicates that “by 2030, approximately three-quarters of all cancer deaths will occur in LMICs, with one in eight people experiencing a cancer diagnosis in their lifetime. (International Agency for Research on Cancer, WHO.  “Cancer Tomorrow.”  Global Cancer Observatory.  2021) 

Solutions must consider expense, the availability of pathology labs/staff, and the capacity of healthcare systems to adopt a new diagnostic tool.  The AMBER Model solution does not require additional diagnostic steps beyond digitizing biopsy slides to be read by the AI AMBER Model.  Our solution will be inexpensive ($30-$50/patient) and utilize AI to read slides, as opposed to human beings. 

We believe the low cost and increased accuracy of the AMBER Model will make it accessible to High-Income Countries (HIC) and to Low- and Middle-Income Countries (LMIC).  We believe this could become standard care for TNBC patients across the globe.  

Our project will be deemed a success when the developed model is prognostic of overall and disease-free survival and predictive of response to therapy.  The future integration of this platform into current clinical practice as an affordable precision medicine companion tool will lead to personalized and patient-centered treatment.

Our Impact Goals include:

1. To increase the accuracy of TILS models for Black, Asian and Hispanic patients diagnosed with TNBC;
2. To reduce the cost of TILS tests by over 90% (Oncotype DX >$4,000).  AMBER Model ($30 -$50)
3. To create an AMBER tool which can be easily adopted and implemented in different healthcare systems around the globe.  

The core technology is Artificial Intelligence which will analyze datasets of TNBC patients to determine a model of morphology features specific to diverse types of TNBC patients (White, Black, Asian, Hispanic).  The results will drive the development of specific computational models for each category of patients.  The models will be tested against large data samples and the final model will be AMBER. 

AI will then be utilized to analyze TNBC digitized slides to provide the most accurate prognosis for the medical professional of overall survival, disease-free survival and therapeutic benefit.

Five

FT – German Corredor Prada, PI

FT – Sahar Almahfouz Nasser, Postdoc Fellow

FT – Satvika Bharadwaj – PhD Student

FT – Tilak Pathak – Pathologist/Data Coordinator

PT – Brittany Brown – Finance Manager 

We have been working on this solution for about 8 years. We started by analyzing samples from patients with lung cancer, and we have recently started to adapt the model for triple-negative breast cancer.

Emory includes diversity as a core value and that value permeates our work, including how we treat patients and how we treat each other as co-workers.  We encourage faculty and staff to participate in one of our affinity groups (AAPI, African Americans, Women in Medicine, LGBTQ+, and Latinx). 

My goal is to build this team as we move into the operational stage of our process.  Expertise in healthcare systems across the globe will be critical to ensuring that this solution will be accessible in LMI countries.  Diversity serves our interests and will intend to continue building a diverse team throughout the entire AMBER Model process.  

Value Proposition:  TNBC patients in both HIC and LMIC can access AMBER to receive a much more accurate prognosis for disease progression and treatment benefit.  This more accurate prognosis will be accessible because the cost is only $30-$50 per patient. 

Customers:  Health insurance companies and medical professionals will value AMBER because it will deliver a much more accurate prognosis for each TNBC patient which can determine whether the patient will respond to neoadjuvant therapy, may need more aggressive neoadjuvant therapy, could benefit from additional adjuvant therapy or should not receive chemotherapy due to ineffectiveness.  Both health insurance companies and medical professionals should welcome greater accuracy and much lower cost as it will avoid unnecessary treatment and produce better health outcomes. 

Delivery of services:  AMBER is a computational model that can be utilized on digitized H&E slides.  Because AMBER is a computational model, it is possible that customers can access it from anywhere on the globe with internet capacity. 

We are not yet in the business model fundraising stage, but our team will need resources to complete the clinical trials to determine the accuracy of our models.  Once the models are finalized, we will need the expertise of someone who understands healthcare markets in LMIs as well as High-Income countries), but we do have on-staff resources to help us bring our product to market. 

Emory has a strong record of bringing biomedical innovations to market through the Biolacity Program, which is part of Emory and Georgia Tech’s joint Biomedical Engineering Department.   In FY23, this program launched 6 products, created 24 start-ups and had the FDA clear 4 products.