BC Cancer and UBC OVCARE:when AI finds a new disease category

Provincial cancer agency + university research programProvincial scaleVancouverClinical, research, women's health oncology

What they're doing

The Ovarian Cancer Research Program (OVCARE) at BC Cancer and UBC ran a study that made a category-level discovery, not just an incremental improvement. Dr. Ali Bashashati, Director of AI Research in OVCARE, and Dr. Jessica McAlpine, a gynecologic oncologist, used AI to analyze histopathology images of endometrial cancer. They identified a previously unrecognized high-risk subgroup inside what was called "No Specific Molecular Profile" (NSMP) endometrial cancer, the largest molecular subgroup of the disease.

The model was trained on a discovery cohort of 368 patients and validated on independent cohorts of 290 and 614 from other centres. The new subgroup had markedly worse progression-free and disease-free survival, meaning these patients had been getting under-treated because clinicians could not see them as a distinct group. The findings were published in Nature Communications in 2024. Bashashati's broader lab is doing similar work in ovarian cancer subtyping, where his CNN-based models approach expert-pathologist performance and are designed to transfer across hospitals.

Why it matters

Most clinical AI in Canada is trying to make existing workflows faster or more accurate. The OVCARE work is doing something different. It is using AI to redraw the disease map: finding patient groups inside the histology that humans could not see, and changing who gets which treatment. That is a higher-order use of the technology, and it points to where some of the most consequential clinical AI value will sit over the next decade.

What's working

A clinical-AI pairing (McAlpine and Bashashati) where the gynecologic oncologist and the computational scientist are equal authors, not contractor-and-client
A Nature Communications publication with discovery plus validation cohorts, which is the credibility bar for changing clinical practice
Transferable model architecture (designed for use across hospitals) that takes the work beyond a single site
A research program (OVCARE) that has been pooling gynecologic cancer data for years, which is what made the AI work possible
A women's health focus, an area chronically under-funded and under-researched, where AI's combinatorial advantage matters most

What's harder than it looks

Translating a research-discovered subgroup into clinical practice takes years of guideline revision, payer alignment, and pathology workflow integration. The discovery is the easy part
AI-discovered disease categories raise hard questions about how clinicians explain a prognosis to a patient ("the model sees something we cannot"). The communication side is underdeveloped
Histopathology AI depends on slide digitization, which most Canadian pathology labs still do not have at scale. The bottleneck is the scanner, not the model

Three takeaways for similar orgs

01The highest-leverage AI work in clinical medicine may not be speed or efficiency. It may be category discovery: finding patient groups inside your data that humans cannot see
02Pair a clinical lead with a computational lead as equal partners. The contractor model produces papers; the partner model changes practice
03If your specialty does not have digitized images at scale, the digitization itself is your AI strategy. Do that first, then the models follow