Prognostic digital twins for clinical trials (covariate adjustment & smaller control arms)

Randomised trials in slowly progressing or rare diseases need large control arms because outcomes vary strongly between patients. A prognostic model trained on pooled historical control data forecasts each participant's outcome under placebo or standard of care; the pre-specified analysis uses that forecast as a covariate. The treatment-effect estimate becomes more precise, so a trial can reach the same power with fewer participants (often fewer control patients), while randomisation and Type-I error control are preserved.

Problem & value

• Today: Phase 2/3 sample sizes are driven by outcome variance. Neurology (ALS, Alzheimer's, multiple sclerosis), rare disease and some oncology settings struggle to recruit, and many patients dislike being randomised to placebo. Classical adjustment uses a handful of baseline variables (age, baseline score, stratification factors).
• AI value: a model that compresses dozens of baseline variables and disease history into one prognostic score explains more outcome variance than classical covariates. Higher explained variance translates directly into lower variance of the treatment estimate, hence smaller N or higher power. Sponsors report the gain per programme; the achievable reduction depends on how well the model predicts the endpoint, so it must be simulated for each trial rather than assumed.
• Regulatory anchor: the FDA's final guidance on covariate adjustment (May 2023) endorses adjusting for prognostic baseline covariates in randomised trials to improve precision [1]. The EMA issued a qualification opinion for the PROCOVA method (Unlearn) in 2022, the first for an AI-based prognostic covariate approach in confirmatory trials [2][3].
• Important distinction: this is not an external or synthetic control arm that replaces randomisation. Randomised concurrent controls stay; the model only reduces noise. Replacing controls with historical data is a different, far more contested approach with bias risks.

Variants

• Covariate adjustment (PROCOVA-style): prognostic score as one pre-specified covariate in a linear or generalised linear model (e.g. MMRM/ANCOVA); the best-accepted variant.
• Unbalanced randomisation: keep power but allocate e.g. 2:1 or 3:1 to the treatment arm, with the prognostic score compensating for the smaller control arm.
• Design-stage simulation: digital twins used only for sample-size planning, go/no-go decisions and Bayesian borrowing; no effect on the final analysis.
• Generative patient simulators: conditional generative models producing full synthetic longitudinal trajectories, used for design exploration rather than confirmatory claims (still emerging).
• Industry variants: small biotech with a single rare-disease asset; big pharma platform using in-house historical trial data; academic consortia and registries (e.g. disease-specific data-sharing platforms).

How it works

1. pooled historical control-arm + registry data (data-use agreements, harmonised to CDISC/SDTM)
2. train & lock prognostic model (tabular / longitudinal deep model; validated on held-out trials)
3. freeze model version; document in SAP before unblinding
4. trial baseline data
5. per-participant prognostic score (blinded to treatment)
6. pre-specified covariate-adjusted analysis (ANCOVA/MMRM with score)
7. simulation of Type-I error, power, bias; sample-size reduction
8. regulatory submission (model description, validation, SAP, code)

• Model: ranges from regularised regression and gradient boosting to deep generative models (e.g. conditional restricted Boltzmann machines or transformer-style longitudinal models) trained on thousands of historical control patients. Inputs are baseline-only variables, so the score cannot leak post-randomisation or treatment information.
• Statistics: the prognostic score enters as an ordinary covariate; as in the FDA guidance, valid inference follows from randomisation if the covariate is pre-specified and measured before treatment [1]. Methodological background: linear adjustment for a prognostic score [4].
• Human-in-the-loop: the sponsor's biostatisticians and regulatory affairs pre-specify the model; an independent statistician or DMC reviews the simulation; the model is locked before database lock. No automated decisions on patient care are involved.

Data preparation

Needed are baseline covariates and longitudinal outcomes of historical control arms and registries, plus the current trial's baseline data. They typically live in the sponsor's clinical data warehouse, EDC and CRO deliveries, partly in legacy study databases and data-sharing platforms under data-use agreements. Typical gaps: heterogeneous formats and coding between studies, missing or differently defined endpoints, inconsistent visit schedules, unclear secondary-use consent and access rights, and no documented provenance.

Preparation steps: inventory the usable studies and rights (PREP-001, PREP-025); harmonise to CDISC SDTM/ADaM with a common data dictionary (PREP-012); extract from source systems via exports or APIs (PREP-005); pseudonymise or anonymise before pooling (PREP-022); define held-out trials as an evaluation set with statisticians (PREP-018). Tools: R or Python pipelines, Pinnacle 21 style validation, version-controlled data specifications. Data harmonisation and rights clearing are the largest block, roughly 60 to 70 percent of total effort; model fitting itself is comparatively small. See also GUIDE-003 for a readiness check.

Implementation paths

Which path for whom

• Enterprise (sponsor): Path B, because the model must be locked, validated and documented inside the sponsor's own statistical process and trained on its own pooled control data; Path A fits only where a vendor already covers the indication.
• Developer: Path C, because open tabular and longitudinal ML stacks on modest hardware are enough to build and validate a prognostic model with full code transparency for regulators.

Path A — Buy (fits where a vendor already covers your indication)

• Unlearn (US): digital-twin generators and the PROCOVA method with EMA qualification opinion; offered as a service to sponsors for disease areas with built models (e.g. neurodegeneration). Check coverage of your indication first.
• CRO biostatistics services: large CROs (IQVIA, ICON, Parexel, Syneos) and EU/UK statistical consultancies increasingly offer prognostic-covariate and simulation services. Cytel (software and consulting for adaptive designs) supports design simulation.
• Choose A when: your indication is covered by an existing qualified model and you lack your own historical data.

Path B — Build on cloud APIs

• Cloud compute is used for training and simulation rather than LLM APIs: Azure, AWS or Google in EU regions inside a validated GxP environment, plus secure research enclaves (e.g. EU trusted research environments).
• LLMs help only on the periphery: drafting SAP text, data-mapping to CDISC, documenting model cards (see HLTH-017, HLTH-028); never for the statistical result itself.
• Choose B when: you own historical data and a statistics team but want elastic compute.

Path C — Self-host (open source)

• Stack: R (rpact for group-sequential designs, mmrm, emmeans), Python (scikit-learn, XGBoost, PyTorch), the pharmaverse ecosystem for CDISC outputs, MLflow for model versioning.
• Hardware: HW1 suffices for tabular prognostic models; HW2 for deep longitudinal generative models on larger histories.
• Effort: a validated statistical-programming environment (21 CFR Part 11 / EU Annex 11 compliant) is the main cost, not compute.

Path D — Custom training

Worthwhile for sponsors with large proprietary control-arm histories in one disease. Needs several hundred to thousands of harmonised patient-level control records, ideally from several trials to test cross-trial generalisation. Data from data-sharing initiatives (e.g. Critical Path Institute consortia, Project Data Sphere, YODA, Vivli) can be pooled under data-use agreements.

State of the art (2026)

• Prognostic covariate adjustment is the only digital-twin use with an EMA qualification opinion and a clear statistical rationale [2][3]; adoption is growing in neurology and rare disease, but remains concentrated in Phase 2 and selected Phase 3 programmes.
• The FDA guidance explicitly excludes machine-learning covariate-adjustment methods from its scope [1], so sponsors should engage regulators early (EMA scientific advice, FDA Type C/INTERACT meetings) and pre-specify everything. The EMA reflection paper on AI in the medicinal product lifecycle sets expectations on data quality, model transparency, validation and lifecycle monitoring [5].
• Limits and failure modes: the gain disappears if the new trial population differs from the training data (distribution shift, new standard of care, different endpoint assessment); endpoint definitions across historical trials must be harmonised; poorly predictive models give little benefit; the sample-size reduction must not be claimed in advance of a validated simulation; model changes after unblinding invalidate the pre-specification.

Shared resources & combinations

• HLTH-009 (trial matching and optimisation) supplies recruitment; HLTH-017 (regulatory and medical writing) drafts SAP, CSR and qualification dossiers; HLTH-028 (terminology and interoperability, OMOP/CDISC) prepares the pooled data; HLTH-029 (GMP/CAPA) and HLTH-010 (pharmacovigilance) share the validated, audit-trailed pharma data platform.
• Shared capabilities: tabular-ml, time-series, causal-inference and simulation stacks also serve population-health risk models (HLTH-020).

Effort, cost & KPIs

• Path A: service contract per trial plus regulatory dialogue; weeks to set up if a model exists, quarters including agency interaction.
• Path B/C: 6-12 months to build, validate and document a model and analysis pipeline, requiring biostatistics, data engineering and regulatory expertise.
• Path D: data access agreements and harmonisation dominate cost.
• KPIs: proportion of outcome variance explained (R^2) by the score on held-out trials, simulated vs observed variance reduction, achieved sample-size or control-arm reduction, Type-I error in simulation (must be at the nominal level), regulator acceptance, time saved in recruitment.

People & change

Involve management and the clinical development lead at the start (budget, regulatory strategy), data owners and the DPO early for secondary-use rights and the DPIA question (Art. 35 GDPR: health data at scale, so a DPIA is likely), and legal/compliance for data-use agreements and AVV with CROs (Art. 28 GDPR). IT supports secure pooled environments. Biostatisticians and regulatory affairs pre-specify the model in the SAP and talk to the agency early.

Works council: normally not triggered, since the system analyses trial participants, not employee behaviour (§87(1) no. 6 BetrVG). If staff performance data would ever be included, or work processes of biostatistics teams change materially, inform the works council under §90(1) no. 3 BetrVG; public-sector sponsors involve the Personalrat.

AI literacy: support staff with training on prognostic-model limits and Type-I error control (Art. 4 AI Act, measures to support literacy; see GUIDE-006). Start with one retrospective pilot indication, show simulation results to the team and the DMC, and keep statisticians in control: the model is locked before unblinding and no patient-care decision is automated. Typical concerns are regulatory acceptance and bias; address them with transparent simulations. See GUIDE-009, GUIDE-005 and GUIDE-008.

Risks, regulation & compliance

• EU AI Act: minimal for the statistical tool in a typical research deployment. It is not itself a medical device or a patient-facing system, and Annex III has no matching category. AI literacy (Art. 4) applies; if the score were used for individual treatment decisions, MDR/IVDR software qualification and high-risk status (Art. 6, Annex I) would have to be assessed [6].
• Clinical trials rules: EU CTR 536/2014 and ICH E9(R1)/E6(R3) GCP; the SAP must pre-specify the model, version and analysis; ICH E9 principles on covariate adjustment apply.
• GDPR: historical patient-level data is special-category health data (Art. 9). Needs a legal basis and safeguards for secondary use (research exemption, Art. 89), pseudonymisation, DPIA, and respect for original consent scope. The European Health Data Space (EHDS) will gradually regulate secondary use.
• Data governance: data-use agreements with data holders, IP and licence terms for derived models, model-lifecycle documentation.
• Mitigations: blinded, pre-specified, locked models; independent simulation; early regulator dialogue; fall back to a classical analysis if the model underperforms (the pre-specified sensitivity analysis).

Getting started

1. Pick one indication where recruitment is hard and historical control data exists; define the primary endpoint and estimand.
2. Inventory accessible historical control data and its legal basis for secondary use.
3. Run a simulation: how much variance does a prognostic score explain on held-out trials, and what N reduction follows?
4. Request scientific advice from EMA or national agencies (and FDA meetings) on the planned adjustment before finalising the protocol.
5. Pre-specify model version and covariate adjustment in the SAP; lock the model before unblinding.

Sources

1. FDA — Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products (final guidance, May 2023) — https://www.fda.gov/regulatory-information/search-fda-guidance-documents/adjusting-covariates-randomized-clinical-trials-drugs-and-biological-products
2. EMA — Qualification of novel methodologies for medicine development (process, CHMP qualification opinions) — https://www.ema.europa.eu/en/human-regulatory-overview/research-development/scientific-advice-protocol-assistance/qualification-novel-methodologies-medicine-development
3. Unlearn — AI-generated digital twins for clinical trials — https://www.unlearn.ai/
4. Schuler et al. — Increasing the efficiency of randomized trial estimates via linear adjustment for a prognostic score — https://arxiv.org/abs/2012.09935
5. EMA — Reflection paper on the use of AI in the medicinal product lifecycle — https://www.ema.europa.eu/en/documents/scientific-guideline/reflection-paper-use-artificial-intelligence-ai-medicinal-product-lifecycle_en.pdf
6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj