Short answer
A clinical evaluation report for medical software is not a literature summary. Under MDR Article 61 and Annex XIV, applied to software through MDCG 2020-1, it is a planned, systematic evaluation showing that your specific device is safe and achieves its clinical benefit for its defined intended purpose. MDCG 2020-1 asks for evidence of valid clinical association, technical performance and clinical performance. There is no fixed minimum: the evidence must be sufficient for your claims. Every class needs a PMCF plan or a justification why PMCF is not applicable. AI software also needs documentation of training and validation datasets and performance across clinically relevant subgroups.
Why the CER is where SaMD files most often struggle
In the technical files we review at TrustedTraceMed, the Clinical Evaluation Report is the section most likely to contain major gaps. Not because it is the hardest document to write — but because many SaMD teams misunderstand what is required.
The most common failure: treating the CER as a literature summary. Finding a few papers showing that algorithms like yours work, citing them, and calling it a clinical evaluation. This is not a CER. It is a bibliography. A clinical evaluation under the MDR (Article 61 and Annex XIV), applied to software through MDCG 2020-1, is a planned, systematic evaluation of clinical evidence that demonstrates your specific device is safe and achieves its intended clinical benefit for its specific intended purpose.
The MDCG 2020-1 structure for software clinical evaluation
Step 1: Define the intended use precisely
The clinical evaluation begins and ends with the intended use. Every piece of evidence you collect must be evaluated against the specific intended use, patient population, user, and clinical context you have defined. If your intended use is vague — "a software tool to support clinical decision-making" — you cannot evaluate whether your evidence is sufficient, because you have not defined what you are trying to demonstrate.
Intended use for SaMD must specify: the medical condition or disease targeted, the patient population (age range, disease stage, contraindications), the intended users (trained clinicians, general practitioners, patients), the clinical context (point of care, hospital, home use), and the specific clinical output (diagnosis, prognosis, treatment planning, monitoring).
Step 2: Build evidence for the three MDCG 2020-1 components
MDCG 2020-1 structures the clinical evidence for software into three components, each of which has to be addressed:
- Valid clinical association (scientific validity): the extent to which the software's output — a concept, conclusion or calculation — is associated with the targeted physiological state or clinical condition. Typical sources are a systematic literature review, professional medical society guidelines and technical standards, supplemented by your own studies where the association is not established.
- Technical performance: the capability of the software to accurately and reliably generate the intended technical output from the input data. This comes from verification and validation — unit, integration and system testing — and, for diagnostic algorithms, metrics such as sensitivity, specificity, reproducibility and robustness on appropriately designed datasets.
- Clinical performance: the ability of the software to achieve its intended purpose as claimed, leading to a clinical benefit for patients when used as intended. Evidence comes from clinical performance studies or clinical investigations with your actual software, retrospective analyses, real-world data and, where available, published data on your device. This is the component most often missing.
Literature on the clinical domain supports valid clinical association, but it cannot replace technical and clinical performance evidence for your own software. Literature searches must be systematic and reproducible: documented search strategy, inclusion and exclusion criteria, and appraisal of the included papers. MDCG 2020-1 also notes that prospective data may be needed where the software predicts future patient states or affects patient management decisions, and that validation should be considered at each new software release.
Step 3: Assess clinical equivalence (with caution)
EU MDR allows clinical evaluation to rely on clinical data from an equivalent device. MDCG 2020-5 (Clinical evaluation – Equivalence, April 2020) explains that technical, biological and clinical characteristics must all be equivalent; for software, the algorithms must be similar, although the code itself does not have to be identical where it was developed under recognised software standards. For SaMD, equivalence is difficult to establish under the MDR for three reasons:
First, access to data. For class III and implantable devices, relying on a device from another manufacturer requires a contract that allows full access to its technical documentation on an ongoing basis (Article 61(5)). For other classes, you must be able to demonstrate sufficient levels of access to the data on the equivalent device to justify the claim (Annex XIV Part A, section 3) — in practice rarely possible for a competitor's software without an agreement. Second, technical equivalence for software requires that the algorithms are similar, not just that the products sit in the same clinical domain. Third, the MDR tightened the equivalence rules compared with the Directives, and Notified Bodies examine equivalence claims closely. Unless you have a strong equivalence case with documented access to the equivalent device's data, build your CER on clinical evidence for your own device.
Step 4: Benefit-risk analysis
The CER must conclude with a structured benefit-risk analysis that explicitly states: the clinical benefits demonstrated by your evidence, the residual risks identified in your risk management file, and the conclusion that benefits outweigh risks for the intended use. This conclusion must be specific — not a generic statement that the benefits outweigh risks, but a quantified or well-substantiated argument that the performance demonstrated by your clinical evidence is sufficient for the clinical context.
Clinical evidence for AI medical software — specific requirements
AI and machine learning SaMD creates additional clinical evaluation challenges. MDCG 2020-1 does not mention AI specifically, but its three components apply in full; MDCG 2025-6 (June 2025) explains how MDR requirements interact with the EU AI Act.
Training and validation dataset documentation: AI SaMD must document the datasets used for training and validation, including size, demographic composition, labelling methodology, and the source clinical settings. NBs want to know whether your validation dataset is representative of the patient population your device will be used on. A model validated only on data from one tertiary hospital may not generalise to the community settings where it will be deployed.
Performance in subgroups: Clinical evaluation for AI SaMD should include analysis of algorithm performance in clinically relevant subgroups — by age, sex, disease severity, and device/imaging equipment. Performance disparities across subgroups are a clinical safety issue today. For AI medical devices that need Notified Body assessment, the EU AI Act's data governance requirements — including examining training, validation and testing data for possible biases — will also apply from 2 August 2028, the date set by Regulation (EU) 2026/1744. See our guide to the EU AI Act and EU MDR for the interplay.
Continuous learning models: For AI models that update based on post-deployment data, the CER must address how the clinical evaluation framework will be applied to updated model versions and how significant performance changes will trigger CER updates.
The PMCF plan — the CER's ongoing companion
Under EU MDR, the CER is not a one-time document. Article 61(11) requires the clinical evaluation and its documentation to be updated throughout the device lifecycle with clinical data from the PMCF plan and the post-market surveillance plan. The Post-Market Clinical Follow-up (PMCF) plan (Annex XIV Part B) defines how you will systematically collect post-market clinical data — through literature surveillance, registry participation, post-market studies, or analysis of real-world data from your installed base. It applies to all classes: where you consider PMCF not applicable, the post-market surveillance plan must justify why.
For most SaMD, the PMCF plan should include: systematic literature searches at a defined frequency, analysis of complaint and vigilance data for clinical performance signals, monitoring of real-world performance metrics, and a defined threshold for when PMCF data would trigger a CER update. In our experience, generic PMCF plans that do not define data collection methods or analysis frequency draw major findings.
How often to update the CER
Neither the MDR nor MDCG 2020-1 sets a fixed update interval for class IIa or IIb software. For class III and implantable devices, the MDR requires the PMCF evaluation report — and, where indicated, the SSCP — to be updated at least annually (Article 61(11)). MDCG 2020-13, the template Notified Bodies use to document their assessment of a clinical evaluation, asks the reviewer to consider an appropriate update frequency in light of the PMCF and PMS plans. Define and justify the frequency in your clinical evaluation plan and align it with your PSUR cycle (class IIa: when necessary and at least every two years; class IIb and III: at least annually). How the CER fits into the wider file is covered in our guide to Annex II technical documentation for software.
If you want an independent check of your CER and PMCF plan against MDCG 2020-1 before you submit, see our technical documentation gap analysis.