Comprehensive Summary
This study developed artificial intelligence models that extract cancer outcomes from electronic health records (EHRs) while preserving patient privacy. The researchers used a teacher-student model framework. Teacher models were trained on PHI-rich data from Dana-Farber Cancer Institute (DFCI) to label public datasets (MIMIC-IV), which were then used to train student models without exposing them to PHI. These models were evaluated on test data from DFCI and Memorial Sloan Kettering Cancer Center (MSK). The student models performed exceptionally well, achieving AUROC values above 0.90 for predicting key outcomes such as cancer presence, progression, and treatment response across both institutions. Additionally, privacy risks were mitigated; teacher models showed vulnerability to membership inference attacks, but student models were resistant, reducing attack effectiveness to nearly random levels. The study also compared these models to those trained using Llama-3-70B, showing superior performance of the teacher-student models.
Outcomes and Implications
This research offers a new way to handle sensitive cancer data across institutions without compromising privacy. The ability to extract detailed clinical outcomes from unstructured EHR data addresses a long-standing gap in precision oncology. By linking this data to genomic databases like AACR Project GENIE, researchers can better understand how molecular characteristics impact cancer outcomes and improve treatment strategies. Importantly, these models performed consistently well across datasets from two major cancer centers, suggesting they can be adapted to other research environments. While additional testing in community hospitals and other settings is needed, this method has the potential to streamline how large-scale cancer research is conducted, paving the way for faster and more secure advancements in understanding and treating cancer.