Clinical trial design is the most central and most challenging part of compliance for AI-assisted detection products in the EU. The key requirement is that the trial design must satisfy both the clinical evidence requirements of the medical device regulations (MDR/IVDR) and the additional provisions of the EU Artificial Intelligence Act (AIA) on algorithm transparency, data quality and human oversight.
-
ISO 14155 requires AI medical device software to demonstrate its analytical validity (whether the output for a given input is accurate), scientific validity (whether the output is associated with the intended clinical condition/physiological state) and clinical performance (whether the output yields a clinically meaningful association). -
The core logic, process framework and basic principles of clinical evaluation established in MEDDEV 2.7/1 Rev.4 fully apply to AI-assisted detection software and form the foundation for building the body of clinical evidence and writing the Clinical Evaluation Report (CER).
The three core pillars of clinical trials for AI-assisted detection products: control groups, statistics and evaluation metrics
1. Scientifically rigorous control group design: control group design is the core of validating AI accuracy. We usually compare AI results with the clinical
gold standard (e.g. pathological biopsy) or an active comparator (similar products / manual workflow). Three commonly used rigorous
design types are shown below:
2. Rigorous statistical analysis and core evaluation metrics
Three core elements supporting market authorisation submissions: data governance, transparency and explainability, and human oversight
These three elements clearly set AI trials apart from traditional medical device clinical trial requirements and directly determine the compliance and evidentiary value of the trial.
They are statutory prerequisites throughout the entire clinical trial design process for AI-assisted detection products, and the safeguard for addressing AI-specific risks (algorithmic bias, overfitting, black-box decisions, performance drift) and ensuring the clinical evidence is accepted by the notified body (NB).
1. Data governance and algorithm transparency
Trial data without compliant data governance will be judged invalid by the NB outright. The AI Act requires data across the entire lifecycle – training, validation, testing and clinical trials – to be representative, high-quality, free of bias and traceable. Data governance is the only vehicle for meeting this legal requirement.
The transparency requirement for AI medical products does not mean fully opening the algorithmic black box; it means explainability in the clinical setting – enabling clinicians and regulators to understand the core logic of AI decisions, recognise the uncertainty of decisions and trace the root causes of erroneous results. It is the bridge between AI algorithm performance and compliant clinical use, and a mandatory requirement of the AI Act.
Transparency and explainability directly shape the trial’s process design, endpoints and risk-control logic, namely:
-
They define the core human–AI interaction workflow of the trial (e.g. lesion-localisation heat maps for imaging AI, annotation of key diagnostic features for pathology AI, confidence intervals for diagnostic results, and alerts for uncertain results);
-
They anchor the trial’s compliance endpoints and validation metrics (e.g. clinicians’ understanding of the basis for AI decisions, the increase in clinicians’ diagnostic confidence due to AI explainability, the accuracy of alerts for uncertain AI results, and clinicians’ review rate based on AI explanations);
-
They support root-cause tracing and risk analysis of adverse events in the trial: when misdiagnoses, missed diagnoses or other adverse events occur, the AI’s feature weights, decision basis and confidence data can be used to pinpoint the root cause (algorithm defect / data noise / human operating error);
-
They validate the clinical applicability of AI performance (e.g. distinguishing whether the algorithm has learned “clinically recognised diagnostic features” or “spurious correlations in the data”, avoiding the problem of “high accuracy in the lab but useless in the clinic”, and ensuring the trial truly demonstrates the product’s clinical applicability – the main reason most AI product clinical trials fail).
2. Human oversight
Human oversight is an inviolable rule of EU medical AI regulation. AI-assisted detection products may only be used as tools to assist clinicians and must never replace clinicians in making the final clinical decision; this is the underlying logic of clinical trial design. Human oversight runs through the entire clinical trial design and directly determines the trial’s design type, core endpoints and evidentiary value – and thus whether the trial succeeds or fails from a compliance perspective.
In summary, building a clinical trial system that complies with both sets of regulations requires you to:
-
Clarify dual jurisdiction under the MDR and the AI Act, and complete risk classification and specific compliance for high-risk AI -
Apply ISO 14155 and MEDDEV 2.7/1 to standardise trial design and the clinical evaluation process -
Choose control groups scientifically (e.g. MRMC, multi-reader multi-case), and draw up a detailed statistical plan and evaluation metrics -
Address AI-specific characteristics: establish a robust data governance system, ensure algorithm transparency, and design effective human oversight mechanisms.
In upcoming issues of our WeChat account we will discuss the specific content and action steps of clinical trial design for an AI-assisted detection product: “image-assisted diagnosis software for electronic nasopharyngolaryngoscopes”.
This is a series of educational articles introducing FDA, NMPA, GMP, CE MDR/IVDR and MDSAP regulations and standards.
Everbiz is happy to share medical device expertise with industry clients and hopes our extensive hands-on experience in quality systems and product registration will be of help to you.