RobbieHolland/Glyph-research

Glyph research alive features and outcomes

The 4,285 alive SAE features used in the Discovery analysis, with linked clinical outcomes for 51,331 studies.

File	Rows	Contents	Compressed size
activations/image.parquet	51,331	343 CT image concepts	15.1 MB
activations/findings.parquet	51,331	3,315 radiology report concepts	27.1 MB
activations/labs.parquet	51,331	627 laboratory concepts	17.6 MB
outcomes/outcomes.parquet	51,331	141 selected prognostic outcomes	3.8 MB
outcomes/survival_labels.parquet	51,331	1,925 event/time outcome pairs	184.1 MB
interpretations/autointerp.parquet	342	Current concept interpretations and available evaluation metadata	36 KB
outcomes/outcome_catalog.csv	141	Outcome names and prediction windows	12 KB
associations/unadjusted.parquet	604,185	Unadjusted odds ratios, p-values, and confidence intervals	35 MB
associations/adjusted.parquet	4,463	Confounder-adjusted results for Discovery-filtered pairs	0.7 MB

Site	Studies	Source split
Discovery Site	15,331	Stanford Main (train)
Internal Validation	5,060	Stanford Main (validation)
Internal Test	5,137	Stanford Main (test)
External Site #1	25,803	Stanford Health Care Tri-Valley

Study-level release scope

This dataset contains 51,331 studies from the Stanford Discovery, Validation, Internal Test and Tri-Valley cohorts. No Mayo study-level data are included in activations, outcomes or survival-label files. Mayo contributions are limited to aggregate association and interpretation results. Mayo analyses cannot be refitted from this distributed study-level dataset.

All Mayo-containing study-level large-file versions have been removed from this HF repository. Its active history starts from the cleaned snapshot. Use this non-Mayo release. The dataset remains private.

Data format

• Files are Zstandard-compressed Parquet. Concept_<index> values are float32 SAE activations; zero means the concept is inactive.
• study_id is a randomly assigned identifier shared across the study-level files. Join them on this column; row order has no meaning. Original accession keys are absent.
• Site uses the five labels above. Each modality has the Discovery analysis feature columns at every site.

Outcome labels

• outcomes/outcomes.parquet contains the 141 outcomes used in the discovery analysis, their history flags, and age and sex fields. Labels are 1 for an event in the prediction window, 0 for a negative with sufficient follow-up, and -1 when the outcome cannot be evaluated.
• outcomes/survival_labels.parquet allows analysis of outcomes beyond those 141. Each <task>_event field is nullable 0/1; its <task>_time field is nullable signed days relative to the index scan. An unavailable outcome pair is null.
• interpretations/autointerp.parquet has one row per top-k 40 concept. Its first column, name, uses R for report concepts, I for imaging, and L for labs, followed by the feature number (for example, R1). discovery_prevalence_pct is the percentage of Discovery Site studies with activation > 0. Fit/test sample counts and available computation times are included; every listed feature has a matching activation column. External Site #2 evaluation fields are null for laboratory concepts because Mayo laboratory results are not reported yet.

Association table reproduction

The association release fmr_orthogonal_paper_v34_20261104 replaces the historical fmr_sigmoid estimates. associations/unadjusted.parquet contains the full 604,185 concept–outcome grid. associations/adjusted.parquet contains the 4,463 Discovery Bonferroni-selected pairs, with corrected orthogonal FMR odds ratios, confidence intervals, p-values, eligibility flags, BH q-values and significance. The activation matrices and their study identifiers are unchanged.

Adjustment uses six clinical domains and up to ten concept covariates, with absolute correlation exclusion at 0.35. The release includes the promoted I26 hardware-exclusion sensitivity specification. Discovery, Internal Test and Tri-Valley use the frozen paper-v34 snapshot, excluding later experimental failed-fit retries. These choices are exploratory. associations/adjustment_specifications.parquet records the covariates and requested cross-fitting splits for each fitted pair; rare-class support can reduce the requested split count in the estimator. associations/manifest.json records settings, checksums and correction families. Failed fits remain unavailable and count as p=1 for the full-family BH correction.

Mayo evaluates only the 3,109 imaging and report hypotheses, using no laboratory covariates. Mayo laboratory estimates are null in both result files and excluded from its adjusted correction family. Original Mayo imaging/report unadjusted screening flags are preserved; historical unadjusted q-values are unavailable because excluded Mayo laboratory results are not used to reconstruct them. They must not be interpreted as negative findings.

The Glyph research repository regenerates numeric tables from these saved aggregate estimates without refitting patient-level models. Every table reports its tested denominator. The release matches manuscript v34 counts in every modality and cohort. It reproduces the paper snapshot, not the later experimental failed-fit retry cache. No study-level data are added by this association update.

Pretrained SAE checkpoints

Concept interpretations

The standard interpretation file contains the current description for every listed concept. Descriptions are used directly; no separate override file is required. Missing accuracy values indicate that the current wording has no matching held-out evaluation. Sample counts and timing metadata, where present, are retained as AutoInterp pipeline provenance. SAE weights, concept indices, activations and aggregate associations are unchanged by this annotation update.