Health AI
Evaluate a clinical model in a national population, define the cohorts it should learn from, or test what it predicts against observed outcomes.
The Danish registers follow every resident through hospital contacts, redeemed prescriptions, cancer diagnoses and death, linked by one personal identifier over decades.
We run the work ourselves inside the approved research environment, under Frameshift's own authorisation with Statistics Denmark (no. 743). Our contribution is the epidemiology: what a register code does and does not measure, where a label leaks information from the future, and what a fair comparison looks like.
Ways to work with us
How a model performs in a population it has not seen.
We map the model's inputs to what the registers record, and fix the population, prediction horizon, outcome and protocol before the model is run. The report covers discrimination, calibration, performance over time and across patient groups, and a comparison with conventional baselines. Inputs the registers cannot represent are identified at feasibility.
What the model learns from, defined with care.
We define cohorts, phenotypes and outcomes from register data, prepare longitudinal event sequences, and set up a held-out evaluation that stays apart from anything used for training or adaptation. Training takes place inside the approved environment; see what the environment allows.
From a nominated relationship to an estimate.
When a platform nominates a drug-disease relationship, a patient subgroup or a suspected adverse effect, we assess whether the registers hold an adequate population, outcome and comparator, then run the prespecified study. Causal questions are designed as target trial emulations. Null results are reported as such.
Evaluation protocol
Each component is written down first, then mapped to what the registers can support.
| Component | What is specified | In register data |
|---|---|---|
| Intended use | The decision the model supports and the point in care at which it is applied. | Sets the index date. A prediction made at diagnosis and one made at the start of treatment are different problems with different cohorts. |
| Target population | Eligibility at the index date, on information available at that time. | Applied using pre-index register history, with the lookback window stated. |
| Inputs | Each predictor, its source and its time window. | Mapped to register variables. Hospital diagnoses, procedures and redeemed prescriptions are recorded. Laboratory results can be applied for. Imaging and clinical notes are not in the registers. |
| Outcome and horizon | The outcome definition and the prediction window. | Mapped to register definitions, whose coding accuracy varies between diagnoses. Death as a competing event is handled according to what the prediction is for. |
| Model version | One frozen model, with its preprocessing and its handling of missing inputs. | The evaluation describes that version. Recalibration or adaptation is a separate step and is labelled as one. |
| Test data | Data neither the model nor its developers have used. | Kept apart from any adaptation. Repeated rounds of feedback on one test set turn it into training data, so the number of looks is agreed in advance. |
| Metrics and baselines | Discrimination, calibration and the comparators. | Compared with a conventional regression or survival model on the same inputs, so that what the added complexity buys is measured. |
| Groups and time | Prespecified patient groups and calendar periods. | Performance by age, sex, comorbidity and period, each with its uncertainty. |
Departures from the protocol are documented as part of the deliverable.
Scope and limits
Register microdata stay on Statistics Denmark's servers. Only disclosure-cleared outputs leave: tables, figures, performance estimates and reports. Nobody downloads the records, ourselves included.
A collaboration normally runs as its own project, approved for that work and for statistical or scientific purposes. How register access works.
Imaging, clinical notes and genomic data do not come with register access. Statistics Denmark can sometimes link a partner's data for a Danish cohort. The registers cannot supply follow-up for patients outside Denmark.
A retrospective evaluation shows how a model performed in a defined population and period. On its own it does not show that a model is ready for clinical use.
Team research
Peer-reviewed research by members of the team on clinical prediction in Danish haematology. These are research publications, not Frameshift client engagements.
Get in touch
Send a non-confidential description of the model or question, the inputs it needs and the population you care about. We will tell you whether the registers can support it and what a first project would look like.