Minisymposium

This talk delves into the often underestimated yet pivotal role that social determinants of health (SDOH) play in the landscape of cancer research. Beyond the cellular and genetic intricacies, the broader context of an individual's life can significantly influence cancer outcomes. We will unravel the interconnected web of socioeconomic, cultural, and environmental factors that shape health disparities and impact cancer incidence, progression, and treatment responses. By acknowledging and understanding these determinants, we can pave the way for more inclusive and effective approaches in cancer research, fostering a comprehensive understanding of the complex interplay between biological factors and the diverse environments in which individuals live.

Recent studies have shown that both biologic and non-biologic disease-based characteristics can be predicted from medical imaging. For instance, self-reported race, sex, and age can be predicted from chest X-rays and other imaging modalities, as well as disease conditions, such as ICD code diagnoses. Additional research has indicated that disease risk, such as the risk of breast cancer, can be predicted from medical imaging with better performance than clinical and traditional scoring systems like the Tyrer-Cuzick and Gail breast cancer risk prediction. These models prove to be powerful even when they lack high precision labels from radiologists.

However, these image models face challenges due to the inadequacy of existing explanatory techniques. Furthermore, given the known issue of shortcut learning causing bias, there is increased concern over the use of image-only models.Conversely, if properly leveraged, image-only models can be successful, particularly for opportunistic screenings and mining information for population health, even if their initial intent was not for subsequent use.

I will discuss image-only models and their methodologies that have demonstrated superior performance over non-traditional imaging-only models, as well as the challenges and limitations of scaling imaging for precision medicine, specifically shortcut learning and limited explainability techniques.

Federated learning (FL) is a collaborative learning approach where multiple data owners train a model together under the orchestration of a central server by sharing the model trained on their local datasets instead of sharing the data directly. FL enables creation of more robust models without the exposure of local datasets. However, FL by itself, does not guarantee the privacy of data, because the information extracted from the communication of FL algorithms can be accumulated and utilized to infer the private local data used for training. We developed Advanced Privacy Preserving Federated Learning framework (APPFL), with advances in differential privacy, to enable Privacy-Preserving Federated Learning (PPFL). We enabled PPFL through scaled, distributed training on supercomputing resources across multiple institutions to help create robust, trust-worthy AI models in biomedicine and smart grid applications. Setting up a secure high-performance computing FL experiment requires capabilities that may not be available for all. To lower the barrier for leveraging PPFL, we created the Advanced Privacy-Preserving Federated Learning as a service (APPFLx). APPFLx enables cross-silo PPFL with easy to use web interface for managing, deploying, analyzing, and visualizing PPFL experiments. In this talk, we will describe APPFLx and its adoption to biomedical use cases.

This session will be an open discussion of developments, challenges and future directions of AI and life sciences, and how we can guide the way towards ethical, fair and effective integration.