WHAT WE DO

AI for Health Economics

Evidence infrastructure for the models that decide where health budgets go.

Ministries of health, insurers and development financiers ask a version of the same question: given a budget that is fixed and usually too small, which investments produce the most health, for whom, and over what period. Answering it takes a decision-analytic model, and building one takes evidence that is scattered across ministry reports, national surveys, tariff schedules, programme evaluations and grey literature.

That evidence assembly, not the modelling itself, is what holds the work up. A single country model needs somewhere between two hundred and eight hundred parameters, each with a source, a distribution and a judgement about whether it transfers from the setting it came from. Assembled by hand, that is several months of a skilled analyst reading, and it has to be redone every time a new survey lands.

Omnimem builds the layer that compresses that step and makes it auditable. We work alongside health economics teams: they own the model, we build the evidence and reasoning infrastructure around it. It is the architecture behind the modelling programme of Movement Health Foundation, applied in Ghana and Peru, and evaluated by the University of Southern Denmark.

THE DESIGN PRINCIPLE

The model stays human. The AI works on the evidence around it.

No epidemiological result is produced by a language model. Every number that enters a model carries a traceable link back to a source document, and every number is confirmed by a named analyst before it is used. A language model that cannot show its source is treated as having produced nothing.

This is not a limitation we accept reluctantly. It is what makes an output defensible to a ministry, to a peer reviewer and to a regulator, and it is what allows the system to be classified as decision support with a human decision maker rather than as an automated decision system.

A reviewer can take any figure in a report and trace it back through the scenario, the parameter, the analyst who accepted it and the source document, to a page in an original publication.

ARCHITECTURE

Six layers, from a scanned ministry report to a question a minister can ask

Each layer is separately testable and separately versioned. Evidence enters at the top, passes a human acceptance gate, and reaches the models through a single versioned parameter store that every model run reads from.

1. Ingestion
Crawlers over defined source registries rather than open web search, with PDF and table extraction, OCR for scanned ministry reports, document type classification and deduplication. Peer-reviewed literature, national plans and tariff schedules, household and facility survey output, international reference databases, and the grey literature that carries a disproportionate share of what is usable in these settings.
2. Representation
Structure-aware chunking, so a table stays with its caption and footnotes and a methods paragraph is not cut in half. Sentence embeddings and vector indexing, with retrieval combining dense similarity and sparse lexical matching, because much of what has to be found is identified by an exact term: an ICD code, a district name, a specific indicator label. Every chunk carries source, date, publisher, country, geographic level and source class, and retrieval can be filtered on any of them.
3. Extraction
Retrieved text becomes candidate values in a typed schema the modelling team defines first. Units and denominators are normalised to the model’s conventions, a five-year cumulative probability converted to an annual rate, an incidence per 100,000 converted to a per-person probability. Each candidate carries its population and setting, study design, source location, a transferability assessment and a confidence score.
4. Agents
Multi-step analytical work is handled by specialist agents with a defined role, a defined set of tools and a defined output contract: evidence search, extraction, transferability assessment, calibration support, critique and drafting. Autonomy is bounded, and no agent writes to the parameter store.
5. Model interface
A versioned parameter store, typed schema, model runners, an experiment manager and a result store. Cohort and individual-level model layers read from the same artefact, which is what stops two models being fed from two different spreadsheets and quietly drifting apart.
6. Interrogation
A programme manager or a financier can ask a completed analysis a question in ordinary language: what happens at half the target coverage, what does this cost in year three, which region gains least. The layer does not reason about health economics. It maps the question onto a defined query over stored results, or onto a scenario the model runner executes, and returns the figure with the scenario definition and the provenance behind it. If a question cannot be answered from stored results or an executable scenario, it says so rather than producing an estimate.

Three controls on every extracted value

Independent second pass. A second extraction runs against the same source under a different prompt and, where useful, a different model. Disagreement routes the parameter to human review rather than being resolved automatically.
Consistency checking. Values are checked against internal constraints (probabilities in range, stage distributions summing to one, incidence consistent with prevalence and duration) and against the parameter set already accepted. A conflict is flagged with both values.
Mandatory human confirmation. Nothing enters a model run until an analyst accepts it, recorded with identity and timestamp. Low-confidence values, cross-pass disagreements and parameters that mattered in earlier sensitivity analysis are reviewed first rather than in arbitrary order.

What the audit trail holds

Every published result can be reconstructed. Five things are versioned together and linked to each figure:

The corpus snapshot, including which documents existed at the time of the run
The parameter set, with the accepted value, its source, the analyst who accepted it and the date
The model code under version control, with the commit recorded on the run
The scenario specification and the random seeds
The output set

This is the property we consider most important in the whole architecture, and the one that makes external evaluation possible.

The automation boundary

We state this in writing at the start of every engagement, because the boundary matters as much as the capability.

Performed or accelerated by the system	Reserved for named humans
Finding and organising evidence	Deciding the model structure and the states it contains
Extracting candidate parameters with provenance	Accepting any parameter into a model run
Flagging inconsistency and low confidence	Judging whether evidence from another setting transfers
Proposing calibration regions and summarising fit	Selecting the calibrated parameter ensemble
Running scenarios and storing results	Choosing which scenarios are decision relevant
Drafting text from stored results	Every interpretation, conclusion and recommendation
Challenging the current model in writing	Deciding what to do about the challenge

The critique agent in the last row is worth singling out. Its purpose is adversarial: it is prompted to find reasons the current model is wrong, and its output goes to the modelling team as a written challenge rather than as automatic corrections. It catches a class of error human teams are poor at catching in their own work, namely assumptions that were reasonable when made and were never revisited after the parameters around them changed.

HOW WE WORK WITH ORGANISATIONS

Movement Health Foundation

Movement Health Foundation is a not-for-profit working on health system innovation, with programmes across Latin America and sub-Saharan Africa. It builds quantitative evidence for health investment decisions in low- and middle-income countries, using cohort state transition models for disease natural history and long-run cost-effectiveness, alongside agent-based models for the parts of a health system that cohort models cannot represent: capacity limits, referral chains, geographic access and patient behaviour.

The Foundation owns the models, the structure and the conclusions. Omnimem builds the evidence and reasoning infrastructure around them, and the two model layers read from one shared, versioned parameter set rather than from separate spreadsheets. The technology has been evaluated by the University of Southern Denmark.

COUNTRY PROGRAMME

Ghana: non-communicable disease investment

The question was not which NCD intervention is cost-effective, but how a constrained budget should be split between disease-specific programmes and the shared delivery platform all of them depend on. The portfolio was modelled vertically, condition by condition, and horizontally across facility tiers, representing staffing by cadre, diagnostic equipment, medicine availability, catchment population and travel distance.

Disease-by-disease analysis systematically disadvantages platform investment, because the benefit of a shared resource is split across programmes and each fraction looks small. Modelling the platform explicitly, with the disease programmes drawing on it, attributes the full benefit to the investment that generates it and prevents it from being counted twice. In Ghana this changed the ranking of investment options.

COUNTRY PROGRAMME

Peru: the breast cancer pathway

Outcomes here are shaped less by treatment efficacy than by stage at presentation and by the attrition between a first suspicion and the start of treatment. The pathway was modelled step by step: awareness and care seeking, access to clinical breast examination, imaging, biopsy and pathology capacity, waiting time at each step, referral between levels, and explicit loss to follow-up between every consecutive pair of steps.

That structure separates two situations that look identical in aggregate data: a programme where few women are screened, and a programme where many are screened but most abnormal results never reach diagnosis. They call for different interventions, and the cost of getting that wrong is large. Geographic inequality was modelled explicitly, because access in Lima and access in the Andean and Amazonian regions differ enough that a national average conceals the decision.

What carries to the next country, and what does not

Reusable: the model structure, the extraction schema, the pipeline, the quality controls and the reporting apparatus. Not reusable, and never transferred without an explicit adjustment and a stated rationale: any country-specific epidemiological or cost value. The marginal cost of the next country is the corpus, the parameter review and the calibration, not the platform.

ENGAGEMENT

How a country programme runs

Five phases. Indicative durations assume a full national model in a new country; a repeat analysis, or an extension to a model that already exists, is substantially faster.

Phase	What happens	Indicative duration
1. Scoping	Stakeholder workshops with the ministry and partners. The decision problem, comparators, perspective, horizon and thresholds are agreed in writing before anything is built.	4 to 6 weeks
2. Evidence assembly	Corpus construction, retrieval and extraction, analyst review and acceptance, a gap register, and expert elicitation where no evidence exists.	6 to 10 weeks
3. Model build	Cohort structure, the delivery system model, coupling between the two, calibration against national targets, and validation at four levels.	8 to 12 weeks
4. Analysis	Base case and scenarios, probabilistic and deterministic sensitivity analysis, equity analysis, and value of information.	4 to 6 weeks
5. Transfer	Technical report, reporting checklist, interactive access to the model, training for national counterparts, and handover of the country parameter set. A model that lives only with the consultants who built it stops being used the moment the engagement ends.	4 to 8 weeks

Work that is built to be checked

The methodology this infrastructure supports has been evaluated by the University of Southern Denmark. External academic evaluation matters for a specific reason: an organisation that both builds a model and reports its results has an obvious conflict, and the appropriate response is to have the method examined by people with no stake in the conclusion.

Published standards are applied, and deviations from them are stated where they occur:

CHEERS 2022
ISPOR and SMDM modelling good practice
ISPOR good practice on dynamic simulation and agent-based modelling
WHO-CHOICE
GBD disability weights and life tables
GATHER
TRIPOD+AI
GDPR and the EU AI Act

Data, sovereignty and ownership

Aggregate data only. The system works with published and administrative aggregate data. It does not require and does not ingest individual patient records. Where a programme has a legitimate need for individual-level data, that work runs under a separate data sharing agreement inside the data holder’s own environment, and what leaves is aggregate.
Synthetic populations. Simulated populations are generated from published distributions. They are not derived from, and cannot be re-identified to, real individuals.
Residency. Omnimem is a Danish company and the infrastructure runs under European data protection law. Programmes that require national data residency are supported by deploying the corpus and index inside the partner’s own jurisdiction.
Access control. Role-based access to corpora, parameter stores and results, with country programme separation, so one partner’s material is not exposed to another.
Third-party models. Where commercial language models are used, they are used under agreements that exclude submitted content from provider model training.
Intellectual property. Ownership of parameter sets, model code and results is settled in writing at the start of each engagement, with the default that national data and country-specific parameter sets stay with the national partner.

Talk to us about a country programme

If you are building health economic evidence in a setting where the data are thin and scattered, the useful first conversation is about the decision the model has to inform, not about the technology. We are glad to walk through the architecture, the audit trail and what it took to stand up Ghana and Peru.

Book a demo

Why Omnimem