Are you working with a machine learning (ML) model to predict disease, but having trouble bringing it to fruition? A team of researchers, which included Dr. Mia Gaudet, of NCI’s Division of Cancer Epidemiology and Genetics, recently developed a highly successful ML model (called Histomic Prognostic Signature, or HiPS).

As noted by Dr. Gaudet, “HiPS consistently outperformed even skilled pathologists in predicting how breast cancer would progress in some individuals.”

Below is some advice from the researchers that might be helpful for getting your own model off the ground:

Engage a Multidisciplinary Team

Dr. Gaudet noted, “With the increasing easy availability of data, it’s critical to remember the value that clinical experts, data scientists, epidemiologists, bioinformaticians, and statisticians bring to your team.”

Use Diverse Data

You can give your model more prognostic power by using diverse data.

Co-corresponding author, Dr. Lee A.D. Cooper, of Northwestern University’s Feinberg School of Medicine, added, “It was vital to involve patients from both large and small community hospitals, as it ensured we were using data that truly reflected today’s cancer patients.”

In total, their model captured pathology image data representing more than 3,000 patients, from large academic centers to rural hospitals.

Think Broadly When Framing Your Model

According to the researchers, their model fills in some of the gaps found with previous models—which focused on cancer cells without considering other cells and processes that are also present within the tumor.

“Biologists have known for a long time that immune and other noncancerous cells are important in cancer progression,” said Dr. Cooper. “We wanted our model to incorporate this information in a way that others, particularly pathologists, could understand.”