How it works

Foundation models encode CT scans, reports, and labs. Sparse autoencoders then isolate latent features and map each one to a human-readable clinical concept.