Designing a clinical investigation for a MedTech startup under MDR means scoping the study to answer one specific clinical question with the smallest viable sample, the tightest endpoint set, and the fewest sites the Regulation and the statistics will permit. The design framework runs in order: define the clinical question from the intended purpose, choose the study type that fits the question, select one primary endpoint that maps directly to a GSPR, calculate the sample size from that endpoint, pick the comparator honestly, plan recruitment against a realistic site pipeline, lock the statistical analysis plan before enrolment opens, and defend the whole design through a written risk-benefit analysis. The legal framework is Regulation (EU) 2017/745 Articles 62 to 64, the broader Articles 62 to 82, Annex XV, and the Good Clinical Practice standard EN ISO 14155:2020+A11:2024. A good startup design is the smallest study that a Notified Body and a competent authority cannot reasonably ask a second question about.

By Tibor Zechmeister and Felix Lenhard. Last updated 10 April 2026.

TL;DR

  • A clinical investigation design is a document and a decision stack. The document is the clinical investigation plan required by Annex XV. The decisions behind it are where founders save or waste hundreds of thousands of euros.
  • Design always starts from the intended purpose and the specific GSPR claims that literature, equivalence, and harmonised standards could not close. Everything else is scope creep.
  • One primary endpoint, mapped one-to-one onto one regulatory question. Secondary endpoints only if they cost nothing extra to measure.
  • The sample size is a calculation, not a convention. A competent biostatistician produces the number. A good design defends it in writing.
  • The comparator is chosen to make the scientific answer interpretable, not to make the slide deck more impressive.
  • The risk-benefit analysis under MDR Article 62(4) and the risk management file per EN ISO 14971:2019+A11:2021 are the backbone the competent authority evaluates when they decide whether to authorise the investigation.

Why design is where lean investigations are won or lost

Felix has watched more startups waste money on poorly designed clinical investigations than on any other single regulatory activity. The mistake is almost always the same. The team jumps from "we need clinical evidence" straight to "let us talk to a contract research organisation about running a study." The design phase — the stage between those two sentences — is compressed, outsourced, or skipped entirely. The result is a protocol written to a generic template, sized by folklore, endpoint-stuffed by committee, and fatally disconnected from the specific clinical question the Notified Body was going to ask.

Tibor has reviewed investigation plans as a Notified Body lead auditor where the number of patients was defensible, the execution was clean, the sites were competent, and the data were still useless for certification because the primary endpoint did not correspond to any of the claims in the intended purpose. Thousands of patient-hours, hundreds of thousands of euros, and a clean data set that answered a question nobody had asked.

Design is where this failure mode is prevented. The design phase is cheap. Biostatistician time, medical writing time, and structured sessions with clinical partners are expensive-per-hour but cheap-per-euro-saved-later. Every hour spent sharpening the scientific question and tightening the endpoints removes tens of hours of execution waste and protects the runway.

The companion post How to Run a Lean Clinical Investigation as a Startup with Limited Budget covers the operational side of running the study. This post is about the decisions that must be made before the operational work starts.

The MDR text — what the Regulation requires of the design

Clinical investigations under MDR are governed by Articles 62 to 82 of Regulation (EU) 2017/745 and by Annex XV. Three articles do most of the work on design.

Article 62 sets out the general requirements. Clinical investigations must be designed, authorised, conducted, recorded, and reported in accordance with the Regulation. Article 62(4) and the surrounding provisions require the investigation to be scientifically sound, the rights and safety of subjects to be protected, the foreseeable risks and inconveniences to be justified by the expected benefits, and the investigation to produce reliable and robust data for the intended regulatory purpose.

Article 63 governs informed consent. The design must account for the consent process — how information is given, how comprehension is verified, how the signed consent is documented, and how the process is adapted for any vulnerable subjects.

Article 64 (and the following provisions through Article 68) addresses clinical investigations on specific populations — incapacitated subjects, minors, pregnant and breastfeeding women, and subjects in emergency situations. If any of these populations are in the study, the design must build in the additional protections these articles require.

Annex XV specifies, in detail, what the clinical investigation plan must contain: the rationale, the objectives, the design, the methodology, the monitoring, the conduct, and the record-keeping. Every section of Annex XV is a design decision in disguise.

EN ISO 14155:2020+A11:2024 — Clinical investigation of medical devices for human subjects — Good clinical practice — is the harmonised standard that operationalises all of the above. The standard is the operational floor for compliance with Articles 62 to 82 and Annex XV, and Section 6 of the standard (Clinical investigation plan) is the structured checklist every design must satisfy.

None of this is optional. The design decisions below are how a startup meets these requirements without building a study that would be appropriate for a multinational.

The design framework, step by step

Step 1 — The clinical question

Write one sentence on a whiteboard. "Does this device, used in this patient population, under these conditions of use, achieve this specific clinical outcome, compared to this comparator, to a degree that supports the claims in the intended purpose?"

If the sentence is vague, the study will be vague. If the sentence has two primary outcomes, the study will have two primary outcomes and the statistics will collapse under multiple-testing correction. If the patient population is broader than the intended purpose, the recruitment will be slow and the data will be noisy.

The clinical question must trace directly to a GSPR in Annex I that literature, equivalence, and harmonised standards could not close. This is the output of Pass 3 of the Subtract to Ship framework — the Evidence Pass — and it is the input to every other design decision. Our methodology pillar The Subtract to Ship Framework for MDR Compliance walks through the Evidence Pass in full.

Step 2 — Study type choice

The type of investigation is chosen from a small menu: single-arm (pre-post or against a performance goal), parallel-group randomised controlled, crossover, cluster, and a handful of hybrid designs. Each fits a narrow class of clinical questions.

Single-arm investigations are appropriate when the effect size is large relative to the natural variation, when a well-characterised performance goal or historical control is available, and when randomisation would be unethical or infeasible. For most first-in-human startup investigations of novel devices against well-documented baseline outcomes, this is the honest choice.

Parallel-group randomised designs are appropriate when the question is comparative, the comparator is accessible, and the sample size is achievable. They are the most defensible but also the most expensive. A startup that chooses this design must be certain the question actually requires it.

Crossover designs are appropriate for chronic, stable conditions where each subject can serve as their own control. They are sample-efficient but require careful washout design.

Cluster designs are appropriate when the intervention is delivered at the site or practice level rather than the patient level. They are rarely the right choice for a startup device investigation.

The design is chosen by working backwards from the clinical question. Any study type that does not answer the question is cut. Any study type that answers the question more cheaply than another is preferred.

Step 3 — Endpoint selection

One primary endpoint. That is the rule. The primary endpoint is the single outcome measure that, if the study succeeds on it, supports the regulatory claim the investigation exists to support. If the study fails on it, the regulatory claim is not supported regardless of what the secondary endpoints show.

A good primary endpoint has four properties. It is clinically meaningful — a clinician and a patient would both recognise it as mattering. It is objective or, if subjective, is measured with a validated instrument. It is measurable within the follow-up window the runway permits. And it maps one-to-one onto a specific claim in the intended purpose.

Secondary endpoints earn their place only if they cost nothing extra to measure — the same visit, the same patient interaction, the same data collection form — and if they answer a question that the Notified Body or the competent authority is likely to raise. Exploratory endpoints that look interesting but cannot be analysed at the planned sample size are not secondary endpoints. They are noise.

Every endpoint in the protocol must be defined precisely enough that two independent investigators would collect it the same way. Vague endpoints produce vague data. Vague data produces unusable investigations.

Step 4 — Sample size

The sample size is the output of a calculation, not the output of a conversation. A competent biostatistician, given the expected effect size (from literature or pilot data), the acceptable false-positive rate (alpha), the required statistical power, and the realistic dropout rate, produces a number.

Take that number. Add the justified buffer for dropouts — usually 10 to 20 percent, depending on the follow-up length and the patient population. Stop.

A lean design resists the gravitational pull towards "a bit more, to be safe." More subjects does not make weak evidence strong. More subjects makes weak evidence expensive. The Notified Body and the competent authority review sample size for whether the study is adequately powered for the primary endpoint, not for whether it is larger than some arbitrary round number.

Document the calculation in the clinical investigation plan exactly as Annex XV and Section 6 of EN ISO 14155:2020+A11:2024 require. Show the assumptions. Show the formula. Show the software and version used. A defended sample size is a sample size that survives review.

Step 5 — Comparator choice

The comparator is what the device is measured against. The choice depends on the clinical question and the ethics of the study.

For a single-arm design, the comparator is usually a pre-specified performance goal derived from literature or from a historical control, documented in the clinical investigation plan with full citation. The goal must be defensible — chosen before the data are collected, not adjusted after.

For a comparative design, the comparator is typically the current standard of care for the indication, not a placebo (placebos in device studies raise difficult ethical questions) and not a weak competitor chosen to make the device look good. The comparator should be the treatment that a reasonable clinician would offer the patient population in the absence of the investigational device.

A dishonest comparator makes the study unpublishable, unusable for reimbursement discussions, and sometimes unacceptable to the Notified Body. An honest comparator, even when it makes the effect size harder to reach, produces evidence that is worth more than the study cost.

Step 6 — Recruitment and site planning

Recruitment is the single largest source of delay and cost overruns in startup investigations. A design is only lean if the recruitment plan is realistic.

Start with the one or two clinical partner sites that were recruited on day one of the company, not after the protocol was finalised. Felix's rule on clinical partners applies to design, not just execution — the partners' patient population defines the inclusion criteria, not the other way around. If the protocol requires a patient profile the partner sites do not see, the design is wrong.

Calculate the realistic monthly enrolment rate per site from the site's actual patient flow for the indication, not from the optimistic number the site gives in the first meeting. Divide the target sample size by the realistic monthly rate. The result is the enrolment duration. Add contingency for holidays, staffing changes, and the predictable pause after the first patient while the site works out the workflow. If the resulting duration is longer than the runway permits, the design must change — either fewer subjects, a different patient population, or additional sites (which brings its own overhead and should be a last resort).

Inclusion and exclusion criteria must be broad enough to recruit at the required rate and narrow enough to keep the data interpretable. The tension between these two is the hardest part of design, and it is where a good principal investigator earns their role.

Step 7 — Statistical analysis plan

The statistical analysis plan is written and locked before the first subject is enrolled. This is a discipline that separates credible studies from ones that can be challenged as data-dredged.

The plan specifies: the primary analysis population (intention-to-treat, per-protocol, or both), the statistical test for the primary endpoint, the method for handling missing data, the rules for interim analyses if any, the adjustment for multiple comparisons if secondary endpoints are tested, and the pre-specified subgroup analyses if any.

Locking the plan means the plan cannot be changed after the data are unblinded except through a documented protocol amendment with a regulatory justification. Post-hoc analyses are permitted as exploratory, but they cannot replace the pre-specified primary analysis.

The clinical investigation plan required by Annex XV includes the statistical analysis plan as an integral component, and EN ISO 14155:2020+A11:2024 Section 6 specifies what it must contain. A statistician, not a clinician, drafts this section.

Step 8 — Risk-benefit analysis

Article 62(4) and related provisions require the risks to be justified by the expected benefits. The risk-benefit analysis is the document that shows this. It sits at the intersection of the clinical investigation plan, the investigator's brochure, and the risk management file per EN ISO 14971:2019+A11:2021.

The analysis identifies every foreseeable risk to subjects — clinical, procedural, device-related, and data-related — and weighs each against the expected benefit to the subject directly (if any) and to the patient population more broadly. The analysis must be honest about the uncertainty of the benefits, because early investigations are by definition the first place the benefits are being measured.

The competent authority reviews the risk-benefit analysis carefully when deciding whether to authorise the investigation. A weak risk-benefit section is one of the most common reasons a submission is returned with questions. A strong one — backed by the 600 to 3,000 bench and simulated-use tests a disciplined startup has already run, by the risk management file, and by the clinical rationale — is the backbone that makes the submission approvable.

The Subtract to Ship angle

The design phase is where the Evidence Pass becomes concrete. Steps 1 to 8 above are the subtraction discipline applied one decision at a time. Every endpoint that cannot be traced to a GSPR claim is cut. Every subject beyond the statistical minimum plus justified buffer is cut. Every site that does not contribute independent recruiting capacity or clinical expertise is cut. Every secondary analysis that cannot be supported at the planned sample size is cut.

What remains is the smallest study that honestly answers the clinical question. That study is also the cheapest study that a Notified Body and a competent authority will accept. The alignment between "lean" and "approvable" is not an accident — it is the consequence of designing from the Regulation outward instead of from a template inward.

Reality Check — Where do you stand?

  1. Can you write your clinical question in one sentence, with a specific population, a specific outcome, and a specific comparator, without hedging?
  2. Does your primary endpoint map one-to-one onto a specific claim in the intended purpose and a specific GSPR in Annex I?
  3. Was your sample size produced by a biostatistician from effect size, alpha, power, and dropout rate, or suggested by a template or a consultant?
  4. Is your comparator the treatment a reasonable clinician would offer the patient population without the investigational device, or is it chosen to flatter the effect size?
  5. Is your monthly enrolment rate calculated from the real patient flow at the actual clinical partner sites, not from the optimistic first-meeting number?
  6. Is your statistical analysis plan locked in writing before the first subject is enrolled?
  7. Does your risk-benefit analysis under MDR Article 62(4) rest on a complete risk management file per EN ISO 14971:2019+A11:2021 and on real pre-clinical data?
  8. Have you mapped every section of your clinical investigation plan to the corresponding Annex XV requirement and the corresponding Section 6 clause of EN ISO 14155:2020+A11:2024?

Frequently Asked Questions

How many primary endpoints should a startup clinical investigation have? One. A single primary endpoint is the discipline that keeps the study interpretable and the sample size defensible. Two or more primary endpoints split the statistical power, force multiple-testing correction, and produce results that are harder to defend to a Notified Body. Secondary endpoints can be added when they cost nothing extra to measure, but they do not replace the focus of one primary endpoint.

Who should calculate the sample size for a MedTech startup investigation? A qualified biostatistician with medical device investigation experience. The calculation requires inputs — expected effect size, acceptable alpha, required power, realistic dropout rate — that a biostatistician knows how to estimate from literature and to defend in writing. A sample size produced by a clinician or a consultant without statistical training is a sample size the reviewer can challenge, and in our experience reviewers do challenge it.

Can a startup run a single-arm clinical investigation under MDR? Yes, when the clinical question and the expected effect size justify it. Single-arm designs against a pre-specified performance goal derived from literature or historical controls are used routinely for medical devices, particularly for first-in-human or early feasibility investigations. The design must defend the choice of performance goal in writing and demonstrate that a comparative design is not required to answer the question.

How detailed must the clinical investigation plan be? Detailed enough to satisfy every requirement of Annex XV of Regulation (EU) 2017/745 and Section 6 of EN ISO 14155:2020+A11:2024. This is not a discretionary level of detail. The plan must cover the rationale, objectives, design, methodology, endpoints, sample size justification, statistical analysis plan, risk-benefit analysis, monitoring plan, adverse event handling, and reporting structure. A plan that omits any of these sections will be returned by the ethics committee, the competent authority, or both.

How do you handle missing data in the design? The method for handling missing data is specified in the statistical analysis plan before enrolment opens. The choice depends on the mechanism of missingness and the primary analysis population. Common approaches include intention-to-treat with imputation, per-protocol with sensitivity analysis, or pattern-mixture models for informative missingness. The key is that the method is pre-specified and documented, not chosen after the data are seen.

What is the relationship between the clinical investigation plan and the risk management file? They are linked but distinct documents. The risk management file per EN ISO 14971:2019+A11:2021 identifies every risk of the device across its lifecycle and documents the controls. The clinical investigation plan references the risk management file in the risk-benefit analysis and uses it to justify why the foreseeable risks are acceptable given the expected benefits. The two documents must be consistent — an investigation plan that identifies risks the risk management file has missed, or vice versa, is a signal of poor integration that reviewers will flag.

Sources

  1. Regulation (EU) 2017/745 of the European Parliament and of the Council of 5 April 2017 on medical devices, Article 62 (general requirements regarding clinical investigations conducted to demonstrate conformity of devices), Article 63 (informed consent), Article 64 (clinical investigations on incapacitated subjects), Articles 62 to 82 (clinical investigations framework), Annex XV (clinical investigations). Official Journal L 117, 5.5.2017.
  2. EN ISO 14155:2020+A11:2024 — Clinical investigation of medical devices for human subjects — Good clinical practice.
  3. EN ISO 14971:2019+A11:2021 — Medical devices — Application of risk management to medical devices.

This post is part of the Clinical Evaluation & Clinical Investigations series in the Subtract to Ship: MDR blog. Authored by Felix Lenhard and Tibor Zechmeister. If you are designing a clinical investigation against a startup runway and need the design to survive both a Notified Body review and a realistic recruitment plan, Zechmeister Strategic Solutions works with founders on exactly that decision — where the scientific question, the Regulation, and the sample size all have to agree on the same protocol.