Human in the Loop or Human Out of Luck

The speaker, a patient living with complex rare conditions, describes a Stanford GENE240 course in which interdisciplinary student teams used AI-assisted tools and multi-omic data, including whole-genome sequencing and proteomics, to investigate her case. The students did not receive her electronic health records, because she worried that decades of misdiagnoses would mislead the analysis. She also arranged individual consent agreements with each student after finding the institutional consent model inadequate.

The teams reached different hypotheses from the same starting data, shaped by their choice of tools, prompts, pipelines and disciplinary instincts. One team fed everything into an AI system and received an impressive first report, but over later iterations a finding it had flagged as important disappeared from the analysis, and the recommended prescriptions vanished with it. Students could go back to the patient with questions and ask for more diagnostics, and they said this contact mattered more than working with the synthetic patients common in similar exercises.

The speaker calls the experiment valid because it narrowed and eliminated possibilities rather than producing final answers. Her case remains unresolved, and clinical translation work with her physician is still ahead. The takeaway is that patients should stay inside the interpretive loop, and that data agency, including getting one's own data back from genomic programs, matters for participatory medicine.