Stanford's Glyph Mines Clinical Concepts From Foundation Models to Accelerate Discovery

Stanford researchers have released Glyph, an open-source framework that uses sparse autoencoders to extract interpretable clinical concepts from the internal representations of medical foundation models — and then tests those concepts statistically against patient outcomes. Applied to multimodal data from 41,628 patients across multiple institutions, the system surfaced 397 predictive concepts carrying 4,463 Bonferroni-significant outcome associations, including signals for pancreatic cancer that substantially outperformed conventional clinical risk markers in predictive power.

The paper, submitted to arXiv on November 6, 2026, positions Glyph not as a tool for explaining point-of-care clinical AI decisions but as an engine for accelerating medical discovery — finding patterns inside foundation model internals that may correspond to previously uncharacterized disease mechanisms or underappreciated clinical biomarkers.

Foundation Models Encode More Than They Were Trained to Predict

Large foundation models (FMs) trained on CT volumes, radiology reports, and laboratory values may implicitly encode clinical knowledge beyond their explicit training objectives. Predicting whether a patient has a particular condition requires the model to organize its internal representations in ways that reflect clinically meaningful structure — but that structure remains buried in dense, high-dimensional embedding spaces that standard explainability tools cannot access.

Post-hoc attribution methods such as SHAP and LIME work at the level of input variables — a lab value, an image region, a word in a report. What Glyph's team at Stanford's Center for Artificial Intelligence in Medicine and Imaging (AIMI) is attempting is structurally different: reverse-engineering the semantic content of foundation model embeddings before they ever reach a downstream prediction task.

Sparse Autoencoders Convert Polysemantic Embeddings Into Labeled Clinical Concepts

Glyph's pipeline operates in three stages. It first passes multimodal clinical data — abdominal CT volumes, radiology reports, and laboratory values — through frozen foundation models to obtain embeddings. Those embeddings are typically polysemantic: each vector dimension simultaneously encodes multiple overlapping concepts, making direct interpretation impossible.

Glyph then trains sparse autoencoders (SAEs) on those embeddings. An SAE learns to reconstruct the original embedding using a much larger but sparsely activated dictionary. The architectural constraint forcing sparsity pushes learned features toward monosemanticity — each dictionary element tends to activate for a coherent, specific concept rather than a blend of unrelated patterns. SAEs are the dominant tool in the mechanistic interpretability literature for recovering features inside language models; Glyph extends that framework to medical foundation models at clinical scale. The team reports that the quality of the resulting concept space depended directly on SAE architecture — an ablation across 141 disease prediction tasks — underscoring SAE design as a critical engineering variable.

An LLM then automatically labels each sparse feature by examining the clinical cases that activate it most strongly and generating a natural-language concept description. Those labeled concepts are tested for outcome associations using Bonferroni-corrected statistics, controlling for known clinical covariates to isolate each concept's incremental predictive value.

Signals Beyond the Literature, Including a Pancreatic Cancer Predictor

Across 41,628 patients, Glyph identified 397 predictive concepts with 4,463 significant associations — most reproducible at held-out sites. The team states that Glyph concepts provided complementary predictive value beyond what literature-derived clinical markers capture, with the combination producing measurable gains in predictive accuracy.

Earlier work from the same Stanford AIMI group previewed the scale of that complementarity. For pancreatic cancer, a predecessor pipeline found a Glyph-derived concept with an odds ratio of 3.46 (95% CI 2.52–4.77), compared with 1.29 for age and 1.46 for smoking status. If the full Glyph results extend that pattern across multiple disease targets, the implications for identifying early diagnostic signals would be substantial.

Competing Approaches in an Expanding Field

At ICML 2026, a separate group applied SAEs to a transformer trained on MIMIC-IV electronic health records and found the model had learned a clinical ontology structurally distinct from the ICD coding system — suggesting FM embeddings capture disease patterns that existing taxonomies miss.

ConceptCLIP pursues interpretability through training rather than post-hoc decomposition, embedding UMLS-derived clinical concepts into a vision-language model pre-trained on 23 million image-text-concept triplets. The approaches reflect a fundamental architectural fork: Glyph mines a frozen foundation model's representations, while ConceptCLIP bakes interpretability into training from the start.

Regulatory Pressure Is Pushing in the Same Direction

The FDA's January 2026 update to its Clinical Decision Support guidance made explainability a central test for how AI software is classified under the agency's device framework. The less a tool's outputs can be independently reviewed by a clinician, the more likely the FDA is to treat it as a regulated medical device. According to a longitudinal analysis of FDA device authorizations, roughly 76 percent of authorized AI and ML medical devices are concentrated in radiology — the domain in which Glyph was developed.

What Validation Will Need to Confirm

The Glyph paper is a preprint. No independent team has yet reproduced the 397-concept and 4,463-association results using a separate dataset or codebase. The specific methodology for controlling confounders and the full benchmark comparison will require the peer-reviewed full text to evaluate. Code is publicly available at GitHub under StanfordMIMI/Glyph, making independent replication feasible.

Whether the mechanistic interpretability toolkit — originally developed to reverse-engineer language model safety properties — proves durable as a clinical biomarker discovery instrument will depend on whether Glyph concepts survive prospective validation against outcomes not available during training. The Stanford team's multi-institutional reproducibility result sets a credible baseline for that test to begin.