× close
Credit: Pixabay/CC0 Public Domain
Advanced artificial intelligence (AI) systems have shown the potential to revolutionize the field of pathology by transforming disease detection, diagnosis, and treatment. However, pathology datasets used to develop AI models may underrepresent certain patient populations, limiting the overall quality of performance and potentially increasing health disparities.
A new study led by researchers at Mass General Brigham shows that standard computational pathology systems behave differently depending on the demographic profile associated with histology, but a larger “foundational model” can detect these differences. It is highlighted that it may help to partially alleviate the
Survey results announced on April 19th natural medicine, We highlight the need for more diverse training datasets and demographically stratified evaluation of AI systems to ensure that all patient groups benefit equitably from the use of AI systems.
“Comprehensive analyzes of the performance of AI algorithms in pathology stratified across diverse patient populations based on independent laboratory data have yet to be performed,” said corresponding author and U.S. Department of Pathology said Dr. Faisal Mahmood of the Department of Computational Pathology. General Brigham Mass.
“This study is based on both publicly available datasets widely used in AI research in pathology and the Mass General Brigham internal cohort, and includes patients of different races, insurance types, and age groups. We found that advanced deep learning models, when trained in a self-supervised manner known as “basic models”, mitigate these differences in performance. Accuracy can be increased. ”
Based on data from the widely used Cancer Genome Atlas and EBRAINS Brain Tumor Atlas, which contain data primarily from Caucasian patients, researchers have developed breast cancer subtype classification, lung cancer subtype classification, and glioma subtype classification. We developed a computational pathology model for tumor IDH1 mutation prediction (an important element in treatment). response).
When researchers tested the accuracy of these models using histology slides from more than 4,300 cancer patients at the Massachusetts General Brigham and Cancer Genome Atlas and stratified the results by race, they found that the models was found to work more accurately in white patients than in black patients. The models the research team tested to identify breast and lung cancer subtypes and predict IDH1 mutations in gliomas found discrepancies of 3.7, 10.9, and 16%, respectively, in making the correct classification.
The researchers attempted to reduce the observed differences using standard machine learning techniques to reduce bias, such as highlighting examples of underrepresented groups during model training. However, these methods only slightly reduced bias.
Instead, using a self-supervised underlying model reduced the disparity. This is a new form of advanced AI trained on large datasets to perform a wide range of clinical tasks. These models encode a richer representation of histology that may reduce the potential for model bias.
Despite the observed improvements, performance gaps remain evident, reflecting the need for further refinement of basic models in pathology. Additionally, this study included a small number of patients from some demographic groups.
Researchers are continuing to explore how multimodality-based models that incorporate multiple forms of data, such as genomics and electronic medical records, can improve these models.
The emergence of AI tools in healthcare has the potential to positively reshape healthcare delivery. Balancing the innovative potential of AI with a commitment to quality and safety is essential. General Brigham is a leader in responsible AI, conducting rigorous research on newly emerging technologies to inform the incorporation of AI in healthcare.
“Overall, the results of this study represent a call to action to develop more equitable AI models in healthcare,” Mahmoud said. “While this is a call to action for scientists to use more diverse datasets in their research, it is also a call to action for regulators and policy makers to evaluate these models before they are approved and deployed. We also urge guidelines to include demographic stratification of these models to ensure AI systems are safe and equitably benefit all patient groups.”
For more information:
Vaidya, A et al. Demographic bias in misdiagnosis by computational pathology models, natural medicine (2024). DOI: 10.1038/s41591-024-02885-z
Magazine information:
natural medicine