Are you developing machine learning and looking for ways to make your model generalizable to a diverse population? What if you could make your model applicable to more people using the data you already have?

In a new study, researchers from NCI CBIIT, Wake Forest School of Medicine, and the University of North Carolina, used an algorithm—the Gerchberg-Saxton (GS) algorithm—to help mitigate race and ethnicity bias in a deep learning model.

The GS algorithm is known for its ability to infer patterns and fill in gaps, making it popular for creating sophisticated optics, such as computer-generated holograms.

According to Dr. Umit Topaloglu, chief of CBIIT’s Clinical and Translational Research Informatics Branch, it was the algorithm’s ability to optimize the distribution of information that first sparked their interest. Using the GS algorithm, they were able to boost parity in a data set of 13,980 patients that included:

• 9,814 patients categorized as European American,
• 1,690 patients categorized as African American,
• 346 patients categorized as Eastern Asian American,
• 641 patients categorized as Hispanic American, and
• 1,489 patients who did not report.

When they tested their new data set using an ML model with a known bias, they saw an immediate boost in prediction accuracy.

Dr. Topaloglu said, “The GS algorithm is a model-agnostic technique that centers on transforming the data rather than tweaking the model. The model can’t replace diverse and large data sets, but it can, with meticulous attention, reach near-equitable representation of all racial/ethnic groups within each batch of data.”

He added, “Enhancing model performance for underrepresented populations is vital, and this approach offers one method for improving equity—both in our ML models and in the way we care for all people with cancer.”