CRIS Lab · Columbia University · Apr 2025 – present
Building an interpretable, ontology-grounded LLM for hierarchical differential diagnosis, tying MONDO and ORDO diseases to HPO phenotypes with MeSH-traceable evidence and clinical codes. The visualisation below follows one clinical phrase from parse to ranked diagnosis.
01 The scaffold
A clinical phrase is parsed, entity spans are recognised, and every concept has a home: MONDO and ORDO disease classes over HPO phenotype nodes, with ICD-11, SNOMED and LOINC codes attached. Beside them sit a vector-quantized codebook, a persistence panel and the differential tree, still at rest.
02 Grounded inference
Dependency arcs bind the phrase, spans ground to their HPO nodes, and is_a and has_phenotype edges carry them up the DAG. Concept embeddings snap to the codebook, a Vietoris-Rips filtration sweeps the persistence barcode, and the traversal ranks the differential, Ataxia-telangiectasia at P = 0.71.
The pipeline drawn above: a dependency parse with NER spans, grounded to an ontology DAG of MONDO and ORDO disease classes over HPO phenotypes via is_a and has_phenotype edges. Concept embeddings are vector-quantized onto a discrete codebook, a Vietoris-Rips filtration yields an H0 and H1 persistence barcode over the concept space, and a top-down DAG traversal ranks the differential by posterior probability.
The first scroll lays out the scaffold. The second animates the inference: every arrow drawn is a traceable step from evidence to diagnosis, which is the point of grounding an LLM in a knowledge graph rather than letting it free-associate.