The PROSurvival project aims to use AI models to predict prostate cancer patient survival by analyzing tissue specimens based solely on hematoxylin-eosin staining patterns, without the need for genomic analysis or explicit Gleason grade determination. Gleason grade, a surrogate marker for outcome prediction, is currently the most important prognostic factor for prostate cancer and crucial for treatment decisions, but is known to be highly subjective. This is reflected in a high variability of grades reported by pathologists, leading to under- and overdiagnosis of prostate cancer.
Research has shown that training AI models that generalize to unseen data require data from multiple sites. However, outcome data often cannot be shared due to privacy restrictions. To overcome this obstacle in data access, we will develop federated AI models that leverage patient history and on-site clinicopathologic metadata in combination with publicly available whole-slice imaging (WSI) data. PROSurvival will establish a privacy-preserving federated learning infrastructure that learns a global model from local data to leverage the trove of clinical routine data without compromising patients' privacy. We condense the image data to clinically relevant predictive pattern information, which reduces the complexity of the dataset and facilitates analysis with off-the-shelf hardware. The long-term aim is to generate a comprehensive, multi-site, digital WSI dataset of prostate cancer specimens through state-of-the-art data access techniques and a computational environment to support the collaborative development of AI for precision medicine.