TRAIL-UKB: A sparse-autoencoder atlas of disease-trajectory transformer representations

We present TRAIL-UKB, a sparse-autoencoder (SAE) feature atlas of a Delphi-2M-inspired masked-event transformer trained on 5.1 million first-occurrence ICD-10 events from 502,383 UK Biobank participants. SAEs decompose the model's internal representations into sparse, individually interpretable features. An automated pipeline produces feature cards combining clinical context, temporal patterns, genotype associations, model-output effects, logical activation rules and natural-language descriptions. A feature's activation may reflect a diagnosis, its surrounding context, or both. To distinguish additional contextual selectivity from patterns inherited from the diagnoses themselves, we compare observed feature context with diagnosis-conditioned expectations, weighting each diagnosis's usual context by its share of feature activation mass. The atlas contains 7,712 features representing diagnosis indicators, comorbidity and temporal contexts, patient strata, genotype-linked enrichments and recording artifacts. SAE reconstructions explain over 99.7\% of hidden-state variance and preserve the model's top-ranked predictions at over 98\% of evaluated positions. Across three independently trained SAE fits, 53\% of supported feature instances recur in at least two. Counterfactuals test feature activation sensitivity, while decoder-direction steering evaluates effects on model predictions. Comparisons with alternative sparse decompositions assess representation quality. Finally, retrospective, prescription-derived treatment-resistant depression classification demonstrates how downstream model scores can be connected to interpretable SAE features through decoder folding and dependence-aware analysis. We release the analysis pipeline and privacy-screened feature cards for selected TRD-associated features.