One of the most immediately operational applications of artificial intelligence in clinical research is the use of Natural Language Processing (NLP) to extract eligibility criteria from electronic health records (EHRs).
The AIS/F-CRIN report highlights that NLP-based tools can identify potential trial candidates within minutes from unstructured medical data — a task that traditionally requires weeks of work by clinical research associates.
Hospital data warehouses significantly facilitate this automation. Some systems can even suggest that a patient may be eligible for other ongoing trials based on their clinical profile.
However, the report also stresses current limitations: data heterogeneity, fragmented IT systems, and inconsistent structuring still restrict large-scale deployment.
Selecting the right patients using powerful prognostic models
Prognostic models aim to identify patients most likely to progress, experience an event, or respond to treatment.
In clinical trials, prognostic enrichment requires deploying such models at the inclusion stage to determine eligibility.
The report describes a significant example in Alzheimer’s disease, where a multimodal model combining imaging and clinical data predicted cognitive decline more accurately than existing approaches.
Notably, hippocampal deformation analysis played a key role.
According to the data cited, a multimodal prognostic model could reduce the required sample size by up to 40% to detect a clinical effect — a major advantage in slow-progressing or heterogeneous diseases.
Optimizing randomization with AI-based algorithms
The AIS/F-CRIN report documents the contribution of AI-driven optimization models capable of balancing treatment arms across multiple and complex clinical variables — far beyond traditional randomization methods.
Published studies suggest that such approaches can reduce trial size by 25% to 50%, particularly valuable in rare diseases or phase II trials, where each inclusion is critical.
AI-enhanced randomization is therefore not theoretical: it is methodologically validated, regulator-compatible, and addresses concrete operational constraints such as cost, slow recruitment, and arm imbalance.
Monitoring treatment adherence using AI (e.g. image recognition)
Innovative solutions now enable automated adherence monitoring through image recognition — for example, verifying drug intake via a photo or selfie analyzed by AI.
While patient acceptability remains a key point of vigilance, as noted in the report, these tools address a major methodological issue: variability in treatment adherence, which often undermines result interpretation.
In the short term, such AI tools are particularly relevant for:
- decentralized clinical trials,
- studies requiring strict dosing regularity,
- trials specifically evaluating adherence.
Using connected devices and sensors for continuous clinical monitoring
Wearables and connected sensors (smartwatches, accelerometers, patches, physiological sensors) generate massive volumes of data.
AI in clinical research now enables real-time interpretation and transformation of these signals into clinically meaningful endpoints.
This opens the door to more sensitive trials, less dependent on on-site visits, and capable of capturing micro-signals often missed by conventional assessments: mobility, gait, sleep, symptom fluctuations.
The report also notes that NLP can help adjudicate endpoints directly from clinical narratives, complementing sensor-derived data.
Strengthening statistical analysis with AI-based prognostic scores (PROCOVA™)
In randomized trials, covariate adjustment is routinely used to correct confounding effects.
The report emphasizes the value of using AI-derived prognostic variables to adjust statistical analyses without altering randomization.
This approach reduces variability and increases statistical power — a strategy viewed favorably by regulatory agencies.
The PROCOVA™ score is one of the few AI-based tools explicitly recognized by the EMA, which has confirmed its methodological guarantees.
This is a critical point: it demonstrates that AI can be integrated into strict regulatory frameworks, provided transparency and validation are ensured.
Simulating trials using in silico models and virtual patients
In silico models are increasingly used to simulate clinical trials before launch.
These approaches rely on clinical databases combined with mathematical and computational tools to establish statistical correlations.
The report distinguishes:
- Mechanistic models, based on pathophysiology,
- AI-based models, capable of generating synthetic cohorts using techniques such as variational autoencoders.
Documented use cases include:
- a validated Alzheimer’s disease simulator to optimize inclusion criteria and trial duration,
- an exploratory model in Duchenne muscular dystrophy (DMD-CTSP),
- a full simulation of a cardiovascular trial (SIRIUS study) to anticipate lipid-lowering treatment effects.
Although regulators have not yet validated trials based solely on virtual patients, they officially recognize the value of these approaches and actively encourage further development.
Conclusion: artificial intelligence in clinical research is already a reality
The AIS/F-CRIN report clearly shows that artificial intelligence in clinical research does not replace randomized trials — it enhances every stage: recruitment, randomization, endpoint measurement, statistical analysis, and even protocol design.
Several of the methods described are already used in real trials; some have received positive EMA feedback, while others are in advanced qualification stages.
The challenge is no longer whether AI will be used in clinical research, but how to integrate it responsibly, transparently, and methodologically into existing regulatory frameworks.
See article on the same subject : Artificial intelligence for clinical research: a major lever to accelerate drug development
