CT imaging
Visible morphology, anatomy and findings in abdominal CT volumes.
Cohort counts on the main page are reported directly from manuscript v34, Figure 2 and the cohort methods on pages 22–23. Discovery totals include training and validation; testing combines Internal Test and both external sites. Download the counts.
THE STUDY
Foundation models learn rich representations of clinical data. Glyph turns those representations into sparse, interpretable concepts—and asks how they relate to future disease.
Code and documentationVisible morphology, anatomy and findings in abdominal CT volumes.
Clinical observations documented in the language of radiology.
Patterns across biochemical, haematological and related measurements.
How closely the sparse concepts reconstruct their foundation-model embeddings.
| Modality | Discovery train | Internal Test | External Site #1 | External Site #2 |
|---|---|---|---|---|
| CT imaging (Merlin) | 0.961 | 0.949 | 0.919 | 0.923 |
| Radiology reports (Qwen3) | 0.870 | 0.805 | 0.759 | 0.647 |
| Laboratory measurements | 0.981 | 0.979 | 0.957 | — |
Mean per-study R², using 8,192 concepts with 40 active per study. The SAEs were trained on Discovery training data.
External Site #2 laboratory results are not reported. Download reconstruction statistics.
The literature timelines on the main page preserve the paper’s frozen HFpEF and sepsis figures. Their clinically adjusted odds ratios are estimated for reconstructed literature markers and Glyph concepts in the study cohort; they are not the effect estimates originally published in the cited studies. References link to the source papers.
The atlas preserves the original seed-1283 paper layout: learned concepts and clinical context variables. Coordinates and connections stay fixed as colors, outcomes and display filters change. Dashed concept outlines retain the original layout's undocumented-concept designation.
Association estimates are from the Discovery cohort and describe outcomes occurring 1–4 years after CT. Odds ratios are per standard deviation of concept activation. Disease hues identify outcome families, while darker shades indicate higher odds. Shading is scaled separately for each outcome, from white at OR 1 to the strongest significant effect. The atlas displays higher-risk estimates (OR > 1). Disease coloring uses Bonferroni-significant associations; the correction covers all 604,185 tested concept–outcome pairs. Inspect a concept's profile to see nonsignificant estimates too. These are associations, not causal effects.
The larger study in manuscript v34 reports 397 high-risk concepts and 4,463 Discovery associations; this atlas retains the original saved paper layout. Curated revised interpretations are used where available. Original AutoInterp accuracies, where shown, measure report-based discrimination. They do not validate revised wording or independently establish an imaging biomarker. The five clinical variables have newly computed unadjusted associations with all 141 outcomes, using the same isolated logistic estimator and Discovery population as the concept analysis. Clinical estimates are per standard deviation and use a Bonferroni denominator of 604,890 (the 604,185 original tests plus 705 clinical tests); frozen concept estimates retain their original correction.
THE RESEARCH
Glyph: An Interpretability Engine for AI-Driven Scientific Discovery in Clinical Data
| Measurement | DiscoveryTraining + validation | Test + external sitesInternal Test + two external sites |
|---|---|---|
| Patients | 14,652 | 26,976 |
| CT volumes / radiology reports | 20,391 | 45,139 |
| Preprocessed CT slices | 3,262,560 | 7,222,240 |
| Radiology report tokens | 8,083,882 | 14,342,345 |
| Laboratory measurements | 1,588,003 | 2,184,424 |
16,553 distinct ICD-9/10 codes mapped to 1,692 phenotype categories · 141 prognostic outcomes · 350 lab types
Cohort totals reported in manuscript v34. Study details · Counts