clinical-research · clinical study design Skill
This Skill works out endpoints, sample size and feasibility in one pass. Endpoints get picked from literature and meeting rooms, Sample sizes wait on a biostatistician's calendar. Feasibility rests on "we should be able to enrol". Every number looks precise, none of them says where it came from, and nobody has signed for it.
The values above come from one synthetic-input walkthrough run on this machine, shown to illustrate the output shape. The same input always returns the same output, and every line can be recomputed against the script that produced it.
How to read this page
Sections 02 to 06 cover what the Skill does: three shifts, the three numbers that stall a protocol, how it works, and what its three outputs actually look like. Sections 07 to 09 cover what it saves and where to get it.
Its limits are written into its own output
Every output carries an ESTIMATE banner stating that this is a first-pass estimate. A verdict is not an authorization: it is a blocker list plus the roles that must sign it. Final sample size must be confirmed by a biostatistician.
The same design move, taken from "this looks reasonable" to "we know where it will be challenged".
The candidate endpoints are listed and the primary is chosen on instinct and precedent. An unvalidated surrogate looks like a good fit, and nobody can say whether it can hold the primary endpoint position with regulators.
Every candidate is scored across five weighted dimensions (clinical relevance, measurability, regulatory acceptance, sensitivity to change, patient burden) and classified as PRIMARY, KEY-SECONDARY or EXPLORATORY. An unvalidated surrogate is flagged and demoted, and cannot become the primary endpoint.
From "this endpoint looks reasonable" to "here is where it will be challenged, and why".
Sample size is often worked backwards: see how many patients can be enrolled, then choose the effect size that makes that number work. Dropout is written into the assumptions but never enters the calculation, so n comes out too small.
A closed-form solution per design: Cohen's d for means, a normal approximation for proportions, Schoenfeld event counts for survival. The sample size is inflated by the expected dropout rate, and the printed assumption block asks you to trace the effect size back to a published or anchor source.
From "the number adds up" to "the number can be traced".
The enrolment plan comes from optimistic site estimates. A dozen sites, a dozen months, a couple of hundred patients: nobody separated "is the eligible population big enough" from "how many patients can each site enrol per month".
Feasibility is scored 0 to 100 across five dimensions (recruitment, endpoint readiness, statistical power, operational complexity, budget fit). Anything below threshold becomes a named blocker and decides the verdict directly: GO, GO-WITH-CONDITIONS, REDESIGN or NO-GO.
From "we should be able to enrol" to "the gap is on this line, and this person has to sign".
They live in different places, and the work of joining them up stays with people.
The problem is not a lack of tools. Statistics software computes sample sizes, regulatory guidance documents endpoint precedent, and a spreadsheet assembles a feasibility table. The problem is that each tool answers one segment of the question.
Worse, what comes out the other end looks a lot like a conclusion. A standard-looking sample size, a plausible enrolment window, a workable site count: none of them is marked as an assumption when it is written into the protocol.
These are the numbers that cause trouble at the phase gate. The committee rarely asks whether you ran the calculation. It asks where the effect size came from, why dropout was set at 15%, and who verified the eligible population. When nobody can answer, the protocol is sent back for what looks like a design flaw. Underneath, it is that nobody can say where the basis sits.
One risk is quieter still: an unvalidated surrogate treated as the primary endpoint. It looks sensitive, cheap and easy to measure early on, right up to the point where regulators say it cannot hold the primary position. By then, the whole design has to be redone.
Every estimate should arrive together with its assumptions, its source and its limits. Every decision should know whose signature makes it stand.
That is what this Skill does. It adds no authority to the judgement. It leaves a trace of it.
All four scripts are pure standard library. No network calls, no model calls. The same input always returns the same output, so any conclusion can be recomputed and challenged. The development area genuinely changes the result: drug, device, biologic, diagnostic and digital-therapeutic each carry their own per-patient cost benchmark and recruitment difficulty coefficient.
The three sets of values below come from one synthetic-input walkthrough run on this machine, and every one of them can be recomputed against the script that produced it.
| Endpoint | Class | Composite | Clinical relevance | Measurability | Regulatory acceptance | Sensitivity | Burden |
|---|---|---|---|---|---|---|---|
| GAD-7 responder rate at week 8 | PRIMARY | 78.2 | 88 | 80 | 60* | 82 | 85 |
| CGI-Improvement at week 8 | KEY-SECONDARY | 66.6 | 80 | 68 | 52* | 60 | 70 |
| In-app engagement sessions per week | EXPLORATORY | 38.2 | 35 | 92 | 20* | 78 | 90 |
* Regulatory acceptance is shown after the digital-therapeutic profile multiplier (x0.80) is applied.
The engagement metric has one of the better raw measurability (92) and sensitivity (78) readings in this set, but it carries the unvalidated-surrogate flag and cannot anchor a primary endpoint.
Formula n = [za*sqrt(2*pbar*qbar) + zb*sqrt(p1q1+p2q2)]^2 / (p1-p2)^2, at alpha 0.05 and power 0.80.
One flag is kept on purpose: the 35% to 55% anchor is supplied by the requester and is not yet cited to a published MCID. The tool still returns a number; the assumption block asks you to trace it to a source.
No model calls, no network, standard library only. The same input always returns the same result, and anyone can take the parameters and recompute a conclusion line by line. Where a protocol has to pass a gate, reproducibility matters more than looking clever.
All outputs carry an ESTIMATE banner stating that this is a first-pass estimate and that the final sample size must be confirmed by a biostatistician using simulation, an adaptive design or an exact method. It deliberately avoids the tone of a conclusion.
A verdict is not an authorization. It is a recommendation with a blocker list, followed by the roles that must sign: Principal Investigator, Medical Monitor, Biostatistician, Regulatory Owner, Study Director, depending on the verdict tier. Onboarding replaces those roles with real names and prints them into the output.
A surrogate endpoint with no validation evidence is flagged and demoted, and cannot become the primary endpoint. That rule moves one of the most expensive kinds of late rework back into the design stage.
drug, device, biologic, diagnostic and digital-therapeutic each adjust the per-patient cost benchmark and the recruitment difficulty coefficient. This is not a relabelled dropdown: budget fit and recruitment feasibility move with it.
What this Skill reduces is not a few calculations. It is the same thing that keeps happening in clinical design: a number is produced, written into a protocol, questioned in review, and then nobody can say where it came from.
When every estimate arrives with its assumptions, its source and its limits, and when every verdict knows whose signature makes it stand, the design process moves from "that is what we decided at the time" to "here is the basis, recompute it". The first set of parameters can carry into the next revision, the first blocker can be tracked to closure, and the first named owner can keep the same thread.
For a clinical team, what is saved is not time. It is rework.
Bring your candidate endpoints, statistical parameters and protocol parameters, and run the seven-question onboarding once to pin down the development area, default alpha, power and three named owners.
What you get back is a first-pass estimate with its basis, its limits and its sign-off chain, ready to put in front of a phase-gate discussion.