AI patient stratification in clinical trials: how it works and how it is regulated

What patient stratification and enrichment mean

Patient stratification divides trial participants into groups by a characteristic, and enrichment uses such a characteristic to choose who enters the trial at all. The US Food and Drug Administration defines enrichment as "the prospective use of any patient characteristic to select a study population in which detection of a drug effect (if one is in fact present) is more likely than it would be in an unselected population", in its final guidance Enrichment Strategies for Clinical Trials of March 2019.

The FDA guidance sorts enrichment into three broad categories:

• Decreasing variability: choosing patients whose baseline measurements fall in a narrow range, which increases study power.
• Prognostic enrichment: choosing patients more likely to have a disease-related event or a marked worsening, the high-risk patients.
• Predictive enrichment: choosing patients more likely to respond to the drug, for example through a biomarker linked to its mechanism.

The FDA names the main concern plainly: the "generalizability and applicability of the study results". In the EU, the statistical rules sit in ICH E9, Statistical principles for clinical trials, in effect since 1 September 1998, and its E9(R1) addendum on estimands and sensitivity analysis, in effect from 30 July 2020 (EMA ICH E9 page). Any AI method used to pick or group patients is judged against these same principles.

AI patient stratification in clinical trials: how AI selects patients

AI is used to select patients for trials in three main ways: screening records to find eligible people, testing how eligibility criteria change the trial population, and predicting which patients are likely to benefit. Most published work is still retrospective or simulated, and the reviews say so.

Finding eligible patients

A 2024 study in Nature Communications described TrialGPT, a large language model framework for matching patients to trials (doi:10.1038/s41467-024-53081-z). Tested on 183 synthetic patients, it reached 87.3% accuracy on 1,015 patient-criterion pairs, and a user study found it cut screening time by 42.6%.

Testing eligibility criteria

A 2021 Nature study, Trial Pathfinder, emulated completed lung cancer trials using health records of 61,094 patients (doi:10.1038/s41586-021-03430-5). Broadening restrictive criteria with a data-driven method more than doubled the pool of eligible patients on average, while many common exclusions had little effect on the trial hazard ratios.

Predictive enrichment

A 2023 npj Digital Medicine paper tested machine-learning "phenomaps" for adaptive predictive enrichment in simulations of two completed cardiovascular trials, IRIS and SPRINT (doi:10.1038/s41746-023-00963-z). Conditioning enrolment on predicted benefit would have reduced the final trial size by an estimated 14.8% and 17.6%, while keeping the original average treatment effect. These are simulations, not prospective trials.

What the reviews conclude

A 2025 scoping review in npj Digital Medicine analysed 142 studies from 2013 to 2024 on AI for clinical trial risk assessment and listed selection bias, few prospective studies and data quality as open problems (doi:10.1038/s41746-025-01886-7). An earlier review in Trials found peer-reviewed evidence "lacking in several areas" (doi:10.1186/s13063-021-05489-x).

Regulation of AI in clinical trials in the EU

In the EU, AI used in clinical trials of medicines must meet the existing good clinical practice and statistics rules, as set out in the European Medicines Agency's Reflection paper on the use of AI in the medicinal product lifecycle. The final paper (EMA/CHMP/CVMP/83833/2023) is dated 9 September 2024. It is a reflection paper, not binding law.

The paper sorts uses by "high patient risk" and "high regulatory impact". Its main points for patient selection:

• AI in trials should meet ICH E6 good clinical practice. For high-risk uses not already qualified by the EMA, the model architecture, development logs, training data and data pipeline may be requested at the trial application, marketing authorisation or a GCP inspection.
• In early-phase trials, AI for "treatment assignment or dosing" can be high patient risk.
• In pivotal trials, "incremental learning approaches are not accepted", and models must be frozen and documented in the statistical analysis plan before database lock and unblinding.

On 14 January 2026 the EMA and FDA published ten joint principles for good AI practice in medicine development. The EMA says further EU guidance is "already underway". ICH E6(R3) GCP took effect in the EU on 23 July 2025 (EMA ICH E6 page).

The FDA draft guidance on AI and its status

The FDA's January 2025 guidance Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (docket FDA-2024-D-4689) was still marked "Draft" on the FDA's page when we checked on 18 October 2026. It sets out a 7-step, risk-based credibility assessment for an AI model's context of use.

Model risk combines "model influence" and "decision consequence". The draft's own clinical example is patient stratification: an AI model sorts trial participants into low and high risk of a serious adverse reaction to decide who is monitored as an inpatient. Because the model alone decides and an error could harm a participant, the FDA rates that model risk as high.

The EU AI Act: what applies and when

The EU AI Act, Regulation (EU) 2024/1689, applies to AI tools used in trials only in some cases. It does not apply to AI "specifically developed and put into service for the sole purpose of scientific research and development" (Article 2(6)). An AI system that is itself, or is a safety component of, a medical device or in vitro diagnostic needing third-party conformity assessment is high-risk under Article 6(1), because those device Regulations are listed in Annex I.

The dates were changed by the Digital Omnibus on AI, Regulation (EU) 2026/1744, published on 24 July 2026 (EUR-Lex). The main dates in the AI Act text as amended:

• 2 February 2025: general provisions and prohibited practices.
• 2 August 2025: rules for general-purpose AI models and governance.
• 2 August 2026: most remaining provisions.
• 2 December 2027: high-risk systems listed in Annex III (was 2 August 2026).
• 2 August 2028: high-risk systems under Article 6(1) and Annex I, including AI in medical devices (was 2 August 2027).

Qualified AI tools and key documents

One AI tool has been qualified by both the EMA and the FDA for use in trials, and it includes patient enrolment. The EMA's human medicines committee adopted a qualification opinion for AIM-NASH on 27 February 2025 (EMA qualification opinion). Its agreed context of use is as an aid to "a single central pathologist" for "enrolment/inclusion of patients into clinical phase 2 and phase 3 trials in MASH" and for evaluating outcomes. The EMA called it its first qualification opinion on an AI-based methodology (EMA news, 20 March 2025), and the FDA qualified it as its first AI drug development tool on 8 December 2025 (FDA announcement).

Sources: EMA, FDA and EUR-Lex documents linked in this guide, checked 18 October 2026. Dates are the document date or the EU effective date.

What this means for clinical trials in Ireland

Irish trial sponsors follow the same EU rules, and Ireland has named the HPRA among the authorities for oversight and enforcement of the EU AI Act. The AI Office of Ireland lists the HPRA alongside bodies such as the Data Protection Commission and the HSE. We found no HPRA statement on AI in trial patient selection as of 18 October 2026. It publishes guidance for sponsors on the Clinical Trials Regulation and CTIS.

The pool AI tools would screen is sizeable. On 30 September 2026 our Irish clinical trial register listed 2,691 ClinicalTrials.gov studies with an Irish site, of which 907 are Phase 3 and 360 were open to new participants. Our clinical trials in Ireland report breaks these down by year, sponsor and city, and the Irish trials dashboard shows them live.

Ireland's public trial infrastructure, as described by the Health Research Board in May 2023, includes the National Clinical Trials Office, the HRB Trials Methodology Research Network led from the University of Galway, Cancer Trials Ireland, five clinical research facilities or centres and seven clinical trial networks. Companies that run trials here are listed in our CRO directory.