Pearson and his colleagues studied the performance of deep learning models trained on data from the Cancer Genome Atlas, one of the largest repositories of cancer genetic and tissue image data. These models can predict survival rates, gene expression patterns, mutations, and more from the tissue histology, but the frequency of these patient characteristics varies widely depending on which institutions submitted the images, and the model often defaults to the “easiest” way to distinguish between samples – in this case, the submitting site.
For example, if Hospital A serves mostly affluent patients with more resources and better access to care, the images submitted from that hospital will generally indicate better patient outcomes and survival rates. If Hospital B serves a more disadvantaged population that struggles with access to quality care, the images that site submitted will generally predict worse outcomes.
The research team found that once the models identified which institution submitted the images, they tended to use that as a stand in for other characteristics of the image, including ancestry. In other words, if the staining or imaging techniques for a slide looked like it was submitted by Hospital A, the models would predict better outcomes, whereas they would predict worse outcomes if it looked like an image from Hospital B. Conversely, if all patients in Hospital B had biological characteristics based on genetics that indicated a worse prognosis, the algorithm would link the worse outcomes to Hospital B’s staining patterns instead of things it saw in the tissue.
“Algorithms are designed to find a signal to differentiate between images, and it does so lazily by identifying the site,” Pearson said. “We actually want to understand what biology within a tumor is more likely to predispose resistance to treatment or early metastatic disease, so we have to disentangle that site-specific digital histology signature from the true biological signal.”
The key to avoiding this kind of bias is to carefully consider the data used to train the models. Developers can make sure that different disease outcomes are distributed evenly across all sites used in the training data, or by isolating a certain site while training or testing the model when the distribution of outcomes is unequal. The result will produce more accurate tools that can get physicians the information they need to quickly diagnose and plan treatments for cancer patients.