Award

All-of-patient Multimodal and Temporal Foundation Model for Breast Cancer

Description: Breast cancer care is not a single event. It unfolds over years and produces many kinds of information: mammograms and MRI scans, biopsy and surgical tissue, lab tests, genetic and molecular results, treatments given, and how a patient responds. Most computer tools that help doctors today examine only one of these at a time, at a single moment, missing how a patient's disease and treatment evolve. Our overarching goal is a trustworthy artificial intelligence that learns from a patient's entire medical record to support better breast cancer decisions. We previously built APOLLO, an all-of-patient model trained on more than 25 billion medical records from 7.2 million patients across 28 data types spanning three decades. Tested on 1.4 million patients, it performed strongly across 322 tasks, including predicting new disease up to five years early, and it outperformed specialized models on cancer progression while highlighting medically meaningful patterns. Our related pathology AI models are used by laboratories worldwide, and one spinoff was acquired by a major pharmaceutical company after extended validation. In this award term, we will rigorously test these models specifically for breast cancer. We will build a large breast cancer evaluation resource and measure whether whole-patient information improves prediction of risk and early detection, tumor subtype, response to chemotherapy before surgery, recurrence, survival, and serious treatment side effects such as heart damage. We will check whether the models stay accurate over time, at outside hospitals, and across patients of different ages and backgrounds, and ask breast cancer physicians to confirm the explanations make clinical sense. This matters because breast cancer decisions, who needs closer screening, which tumors will respond, who will relapse, and who risks treatment harm, depend on many data types changing over time. A carefully validated whole-patient model could sharpen screening, reduce overtreatment, and catch recurrences and toxicity earlier. Demonstrating that such models are accurate, fair, and transportable is the essential step before they can be trusted in real breast cancer care.
Amount: $224,600
Funding type: award
Start date: 2026-10-01
End date: 2027-09-30
Investigator: Faisal Mahmood (lead)
Institutions: Brigham and Women's Hospital, Brigham and Women's Hospital, Inc.
Funded outputs count: 0
Funder award ID: BCRF-26-235
OpenAlex ID: G9000361943