Glyph turns hidden medical AI features into testable disease-risk clues

What can researchers learn from a medical AI beyond its predictions? A model might recognize patterns associated with later disease, yet store them in a numerical representation that is difficult to interpret. Glyph, described by Robbie Holland and colleagues in a November 6 preprint, aims to turn those representations into measurable clinical clues. The findings are research results, not evidence that the system improves patient care.

Separating patterns before naming them

Many automated discovery methods start with a predefined list of findings or diagnosis codes. That makes analysis manageable, but limits what researchers can find. Glyph instead starts with embeddings: numerical representations produced by models processing abdominal CT scans, radiology reports and laboratory histories.

A separate sparse autoencoder processes each type of embedding. This neural network learns to reconstruct the original representation using relatively few active features for each examination. The intention is to separate overlapping information into patterns that can be measured individually. Sparsity encourages clearer features; it does not guarantee that every feature has one clinical meaning.

The researchers then use a language model to propose descriptions by comparing examples with strong feature activation against examples with little or none. They test those descriptions on unseen examples. Imaging features are interpreted through accompanying reports, which may omit relevant visual details. Radiologists directly inspected selected images when automated interpretation failed.

Thousands of associations, tested elsewhere

The study included 41,628 patients across a discovery cohort, an internal test cohort and two external cohorts. Discovery used 14,652 patients; the remaining 26,976 were held out. Patients did not cross the discovery and testing splits.

Glyph produced 4,285 candidate features. The team tested them against 141 outcomes selected because the original embeddings predicted them particularly well. Cases required a first recorded diagnosis one to four years after CT; controls had no recorded diagnosis of that outcome and at least four years of follow-up.

Across 604,185 feature–outcome tests, 4,463 associations met a stringent Bonferroni threshold, designed to limit false positives when many comparisons are made. These involved 397 features associated with higher odds of at least one outcome. They are not 397 newly discovered diseases or biological mechanisms.

Using unchanged feature extractors, 69.3% of the associations reproduced in the internal cohort and 74.3% at the first external site. At the Mayo Clinic cohort, 54.7% of imaging and report associations reproduced. Laboratory associations were not evaluated there because laboratory coverage was limited. Reproduction used a different, less conservative statistical procedure controlling the false discovery rate at 5%.

Several findings recovered familiar relationships: cirrhotic liver findings predicted later liver cancer, while laboratory patterns reflecting accumulated nitrogenous waste tracked kidney disease progression. Other features were harder to name. Radiologists associated one imaging feature with severely distended, fluid-filled bowel loops; it also showed associations with later infection-related diagnoses, although these did not reproduce consistently everywhere.

Prediction is not explanation

The researchers examined whether associations persisted after accounting for measured clinical characteristics and other extracted concepts. Their double-machine-learning analysis allowed nonlinear relationships among adjustment variables.

Only 1,198 of the original 4,463 associations—26.8%—passed the adjusted discovery-cohort filter. Fluid-accumulation features lost many associations, whereas some liver, kidney and blood-cell patterns retained more specific or broader signals. This illustrates why a useful predictor need not identify a distinct disease process: it may partly reflect illness already represented by other variables.

Adjustment still cannot establish causality. Unmeasured factors remain possible, and the authors acknowledge that results depended on a heuristic used to select adjustment variables.

Added information, within a limited comparison

An automated literature search identified 412 reconstructable markers across 21 outcomes. Adding ten selected Glyph concepts per outcome increased average held-out AUROC from 0.749 to 0.774. AUROC measures how well a model ranks cases above controls; it is not the percentage of patients diagnosed correctly.

For sepsis, AUROC rose from 0.673 to 0.706. Dementia and Alzheimer’s disease showed no significant improvement. These comparisons used pooled internal and first-external-site data, not Mayo data. Associations absent from this search are candidates for further investigation, not proof that no previous research exists.

Important limits remain. Everyone underwent abdominal CT, enriching the population for particular illnesses. Diagnosis codes can be delayed or inaccurate, so apparently future disease may already have existed. Analyses treated repeated examinations as independent observations despite some patients contributing several scans. Not every extracted feature was reliably interpretable.

This account relies on extracted paper text and captions; figure images and visual examples were not independently assessed. The central result is nevertheless clear in the reported analyses: separating AI representations can expose testable prognostic patterns. Prospective studies with carefully defined outcomes are needed before those patterns can support clinical decisions.