ASCEND
generative recordsTransformer models that learn how patients’ health records unfold over time, used to study risk, treatment response and patterns of disease.
Aim. To build models of routine electronic health records that can forecast a patient’s future care — diagnoses, prescriptions and laboratory values, and when they occur — and that can be asked what would be expected to happen if a different treatment were started. If such a model reproduces known drug effects from observational data, it becomes a single tool for studying treatment response across many drugs and outcomes, including questions that trials have not addressed.
Approach. Each patient’s record is represented as a time-stamped sequence of coded events and measurements. ASCENDgpt (2025) used phenotype-aware tokens and an encoder model for cardiovascular risk prediction. FlatASCEND (2026) moved to a compact autoregressive generator that predicts both the next event and its timing, and tested whether simulated treatment starts reproduce known pharmacological associations; a companion study used sparse autoencoders to decompose the model’s internal representations into interpretable clinical features. Current work (ASCEND-GEN) trains one generative model on UK primary-care records and evaluates it against target-trial-style emulations of drug starts across 15 drug classes and six routine outcomes (HbA1c, blood pressure, LDL cholesterol, eGFR, weight and potassium). Rather than fitting a separate model for each question, it builds time-dependent drug-response dynamics into the model itself. Because the model generates futures, the same history can be given twice with a different treatment started each time, and the two futures compared. Two kinds of known answer are used as checks, never for training: published patterns of differential drug response (for example from the Exeter work of Dennis and colleagues), and benchmarks in which a made-up drug effect is planted in real records. Current work asks whether the model recovers differences in treatment response between people, not just average effects.
Read the interpretability preprint →ORCA
records as languageTreating coded health records as a language, so that clinical sequences from different health systems can be read, compared and translated.
Aim. Coded health records are produced jointly by clinicians, laboratories and administrative processes. ORCA asks whether these sequences behave like a language, with measurable statistical structure, and whether that structure can be used to align records from different health systems and to translate between coded records and plain English. A shared, system-independent representation of clinical situations would make it easier to compare populations, transfer models between settings and interpret what a model has learned.
Approach. The first phase applied tools from computational linguistics and information theory — frequency and entropy analysis, cross-system alignment, and translation between coded sequences and English — to sequence models trained on records from several health systems. The current phase (ORCA-JEPA) trains joint-embedding predictive (JEPA-style) models on natively coded records from UK and US health systems together, without telling the model which codes mean the same thing. It tests whether patients in the same clinical situation land in the same region of the shared representation whichever system recorded them, using held-out code mappings only as a check, comparison groups matched on age, sex and amount of data, and links to clinical text.
PRIMA-HD
records and retinal imagesSet-upAn exemplar project in Glasgow's Trusted Research Environment that links diabetes records with eye-screening images, to test whether the two together predict outcomes better than either alone.
Aim. People with diabetes in Greater Glasgow and Clyde have rich linked records and regular retinal photographs, but these are rarely analysed together. PRIMA-HD (Pluripotent Representations for Integrated Multimodal Analysis of Healthcare Data) aims to learn each person's care journey from their records, to compare model-predicted outcomes of real treatment choices with what randomised trials found, and to test whether adding retinal images improves prediction. It is also a test of how this kind of work can be done safely inside a Trusted Research Environment.
Approach. Inside the NHS Greater Glasgow and Clyde / University of Glasgow Trusted Research Environment, SCI-Diabetes records are linked with prescribing, laboratory, hospital and death records and diabetic eye-screening images for adults with type 1 or type 2 diabetes. Each record becomes a time-ordered sequence for a generative model of the kind developed in ASCEND, with forecasts compared against statistical models chosen in advance. A retinal foundation model turns each screening photograph into a summary placed on the person's timeline, and prediction of retinopathy progression, kidney disease, cardiovascular events and death is compared with and without images. Trial results are used only to check the model, never to train it. Analysis code goes into the TRE, patient-level data never leave it, and only aggregate results and models that pass disclosure and memorisation checks can be released.
Phenotype fingerprinting
rare phenotype retrievalUsing patterns in clinical records to find patients with rare or under-recognised conditions earlier.
Aim. Many clinically important conditions are rare, inconsistently coded or recognised late. The aim is to identify patients with under-recognised phenotypes earlier, starting from small, carefully characterised case sets rather than large labelled cohorts.
Approach. Expert-characterised cases define a “fingerprint” in the representation space of a clinical sequence model. The fingerprint is then used to retrieve patients with similar trajectories in larger routine-care populations, before the condition is evident from diagnostic codes. Representations from autoregressive and JEPA-style models are to be compared with patient-similarity and weak-supervision baselines.