Glyph splits clinical foundation-model embeddings into readable concepts, surfacing 397 reproducible prognostic markers in abdominal CT cohorts

Clinical foundation-model embeddings predict disease outcomes beyond their training objectives, but that information sits inside high-dimensional vectors that cannot be evaluated one marker at a time. Glyph decomposes these embeddings into sparse features, then filters, names and tests them against future disease, turning a whole vector into individually examinable clinical concepts. Earlier automated clinical discovery relied on manually coded phenotypes or preselected marker sets, so the associations it could find were bounded by those predefined features; agentic systems calling such tools inherit the same boundary. At the discovery site, all pairings of 4,285 features with 141 outcomes formed 604,185 candidate associations, of which 4,463 survived Bonferroni correction across imaging, report and laboratory modalities.

The concepts are not artifacts that collapse on new data: the same sparse-autoencoder weights were applied zero-shot to an internal test cohort and two external sites, and associations still reproduced. The internal test cohort reproduced 3,095 of the 4,463 significant associations (69.3%), the larger community-hospital external site reproduced 3,314 (74.3%), and the out-of-state site reproduced 1,701 of the imaging and report associations (54.7%). Reproduction required Benjamini–Hochberg significance at a 5% false discovery rate and the same effect direction, with no model recalibration at any site.

Once embeddings are split into concepts, each concept can be adjusted alongside age, sex, BMI, alcohol use, smoking, race and other concepts, separating associations that survive adjustment from those that do not. Of the 4,463 significant associations, 1,198 (26.8%) retained a significant adjusted odds ratio after double machine learning; fluid-accumulation concepts attenuated heavily, while laboratory concepts such as cholestasis, cytopenia and azotemia kept most of their associations. Adjustment drew up to ten significantly associated concepts per outcome plus six universal clinical covariates, and the same procedure was repeated at external sites.

A direct comparison with literature markers shows that Glyph concepts carry prognostic information beyond published markers. In the pooled held-out cohort, the top ten Glyph concepts per outcome reproduced at 64.8%, above the 28.2% of literature-derived markers, the 41.2% of the top ten literature markers and the 54.8% of corresponding Phecodes, with a higher median adjusted odds ratio (1.31 versus 1.16). A co-scientist agent searched 342,647 papers and reconstructed 412 markers computable in the local data; both models were fit at the discovery site and evaluated on the pooled held-out cohort.

The findings rest on cohorts of patients who underwent abdominal CT: 14,652 patients and 20,391 studies at the discovery site, plus 26,976 patients across three held-out cohorts. Outcome labels come from ICD codes mapped to Phecodes, and the associations describe adjusted statistical relationships rather than causal effects.

The authors note that the cohort is enriched for abdominal illness by virtue of undergoing abdominal CT, which may inflate the prominence of hepatobiliary and renal markers; ICD coding may involve undercoding, miscoding and delayed entry, so some associations may reflect existing but undocumented disease; the correlation threshold used for confounder selection is a heuristic and results are sensitive to it; and some concepts are not monosemantic, so automated text interpretations need not equal precise clinical meaning.