Glyph: An Interpretability Engine for AI-Driven Scientific Discovery in Clinical Data
Abstract
Automated scientific discovery has the potential to advance understanding of disease mechanisms and accelerate identification of early disease markers. Manual data annotation is infeasible at population scale, motivating methods that automatically extract pre-defined phenotypes in multimodal and unstructured clinical data. Such pre-defined feature sets determine and limit which associations these approaches can discover. This limitation extends to agentic AI systems that use these tools to query population-scale data. In contrast, foundation models (FMs) trained on large volumes of clinical data can perform tasks beyond their training objectives. This suggests their latent embeddings may represent novel or unexplored features that could yield insights into disease mechanisms, yet such latent features remain difficult to interpret. To address this challenge, we introduce Glyph, an automated multistage framework that extracts features from multimodal and unstructured clinical data using FMs and sparse autoencoders, interprets them as coded clinical concepts, and evaluates their value for clinical scientific discovery. We applied Glyph to abdominal CT volumes, CT radiology reports and laboratory measurements from 41,628 patients across different institutions. We identified 397 predictive concepts supported by 4,463 Bonferroni-significant associations. Moreover, a majority of discovered associations were automatically reproducible in held-out cohorts across different sites. Glyph concepts provided complementary associations to literature-derived markers, leading to gains in predictive power. We envision that interpretability can extend beyond a mechanistic understanding of neural networks to connecting implicitly learned features with clinical knowledge and establishing their significance for patient outcomes. We release Glyph at https://github.com/StanfordMIMI/Glyph.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The paper introduces a novel interpretability framework for clinical foundation models that successfully extracts and validates clinically meaningful features at scale. By combining sparse autoencoders with multi-site validation, it provides a rigorous path toward automated scientific discovery in medicine, offering both predictive gains and mechanistic insights that complement existing literature.
Methodology
The paper proposes "Glyph," a framework that leverages foundation models (FMs) trained on clinical data to extract latent features, which are then interpreted via sparse autoencoders (SAEs) as coded clinical concepts. The core methodological contribution is the pipeline connecting unstructured multimodal data (CT volumes, reports, labs) to interpretable, clinically meaningful features without manual annotation. The use of SAEs to decompose FM activations into sparse, human-interpretable components is a strong application of recent interpretability techniques to a high-stakes domain. The approach addresses the "black box" limitation of FMs by mapping latent dimensions to specific clinical concepts, allowing for hypothesis generation rather than just prediction.
Results & evaluation
The evaluation is robust, utilizing a large cohort of 41,628 patients across multiple institutions. The identification of 397 predictive concepts with 4,463 Bonferroni-significant associations demonstrates statistical rigor. The key strength is the validation of reproducibility in held-out cohorts from different sites, which mitigates concerns about overfitting to a single dataset. The comparison against literature-derived markers showing complementary associations and gains in predictive power is a compelling result that validates the utility of the discovered features.
Reproducibility
The authors release the code at the provided GitHub URL, which significantly enhances reproducibility. The use of standard statistical corrections (Bonferroni) and multi-site validation protocols provides a clear template for similar studies. However, the specific foundation models used and their pre-training details would need to be verified in the code repository to ensure full transparency.
Limitations
The primary limitation is the reliance on the quality of the underlying foundation models; if the FMs have biases, the interpreted concepts may inherit them. Additionally, the mapping from latent features to clinical concepts via SAEs may not capture all nuances of the data, and the "interpretability" is limited to the granularity of the coded concepts. The computational cost of running SAEs on large-scale clinical data may be a barrier for some groups.
Broader impact
This work has significant potential to accelerate scientific discovery in medicine by enabling automated hypothesis generation from population-scale data. It bridges the gap between deep learning's predictive power and clinical interpretability, potentially leading to new biomarkers and disease mechanisms. The framework could be adapted to other medical domains beyond abdominal CT, making it a versatile tool for clinical AI research. The paper introduces a novel interpretability framework for clinical foundation models that successfully extracts and validates clinically meaningful features at scale. By combining sparse autoencoders with multi-site validation, it provides a rigorous path toward automated scientific discovery in medicine, offering both predictive gains and mechanistic insights that complement existing literature.
Author of this paper? Send a correction.