clinical-research – skill Dữ liệu & Khoa học cho agent AI

clinical-research

Prospective clinical study DESIGN: endpoints, sample size / power, and phase-gate feasibility. Every output is an estimate with stated assumptions routed to a named human owner. This skill never gives clinical advice as fact and never substitutes for a biostatistician or regulatory affairs.

Purpose

R&D clinical teams, medical monitors, and biostatistics functions live at the moment between *we-have-a-hypothesis* and *we-have-a-protocol-ready-for-submission*. This skill structures three of the hardest design decisions:

Three deterministic tools:

1. sample_size_estimator.py — Closed-form power / sample-size for two-arm means (Cohen's d), proportions (normal approximation), and survival (Schoenfeld events). Inflates for dropout. Prints an "ESTIMATE — confirm with a biostatistician" banner.

2. endpoint_selector.py — Scores candidate endpoints across 5 weighted dimensions (clinical relevance, measurability, regulatory acceptance, sensitivity-to-change, burden) and classifies each as PRIMARY / KEY-SECONDARY / EXPLORATORY. Penalizes unvalidated surrogate endpoints.

3. phase_gate_scorer.py — Scores a study plan 0-100 across recruitment feasibility, endpoint readiness, statistical power, operational complexity, and budget fit; returns GO / GO-WITH-CONDITIONS / REDESIGN / NO-GO plus the named owners who must sign.

When to use

Invoke this skill when:

• You are choosing a primary endpoint and need to defend it against surrogate-endpoint scrutiny.
• You need a defensible first sample-size estimate for a protocol synopsis.
• A study plan needs a feasibility read before a phase-gate review.
• You are pressure-testing whether the planned enrollment is achievable given the eligible population and sites.

Do NOT use this skill to: prepare a regulatory submission or clinical evaluation report (use ra-qm-team), find or position a grant (use research/grants), design a live product A/B experiment (use product-team/experiment-designer), or replace a biostatistician's final sample-size justification.

Workflow

1. Draft the synopsis — Fill assets/protocol_synopsis_template.md (objectives, design, population, endpoints, statistical plan placeholder, owners-to-sign).

2. Select the endpoint — Run endpoint_selector.py --input endpoints.json --profile {drug|device|biologic|diagnostic|digital-therapeutic}. Read the classification + surrogate flags. If >1 primary, plan multiplicity control.

3. Estimate the sample size — Run sample_size_estimator.py --design {means|proportions|survival} .... Trace the effect/difference/HR to a published or anchor-based source; inflate for dropout.

4. Score feasibility — Run phase_gate_scorer.py --input study.json --profile <same> --phase {1|2|3|4}. Read the verdict + blockers + named owners.

5. Route for sign-off — Assemble the synopsis + estimates into the gate packet. The packet is a recommendation; a biostatistician, medical monitor, and regulatory owner sign.

Scripts

| Script | Purpose | Profiles |

|---|---|---|

| scripts/sample_size_estimator.py | Power / sample-size for means, proportions, survival | n/a (design-driven) |

| scripts/endpoint_selector.py | 5-dimension endpoint scoring + classification + surrogate flag | drug, device, biologic, diagnostic, digital-therapeutic |

| scripts/phase_gate_scorer.py | Feasibility 0-100 + GO/GO-WITH-CONDITIONS/REDESIGN/NO-GO + owners | drug, device, biologic, diagnostic, digital-therapeutic |

All three: stdlib-only, --help, --sample, --output {human,json}.

Onboarding & customization

Run the onboarding questionnaire once before you start — it captures your defaults and named owners so every tool in this skill is pre-configured. Customization is the point: the answers actually change tool behavior.

python3 scripts/onboard.py            # interactive (also: --defaults, --set key=value, --reset)
python3 scripts/onboard.py --show     # see the questions + current effective config

Answers are saved to ~/.config/research-ops/clinical-research.json (global) or ./.research-ops/clinical-research.json (--scope project) and are read automatically by config_loader.py. They set the default development-area profile, default alpha / power / dropout, and the named biostatistician / medical monitor / regulatory owner printed on outputs. CLI flags always override saved config; RESEARCH_OPS_NO_CONFIG=1 ignores it entirely.

The seven questions: development area · alpha · power · dropout · biostatistician · medical monitor · regulatory owner.

Optimize with autoresearch (opt-in)

This skill ships an isolated, opt-in bridge to engineering/autoresearch-agent. Only when you ask to "optimize" / "run a loop" does an autoresearch experiment iteratively improve a study plan against this skill's own feasibility score. scripts/ar_evaluator.py is the ground-truth evaluator; it prints feasibility_composite: <0-100> (higher is better).

/ar:setup --domain custom --name trial-feasibility \
  --target study.json \
  --eval "python3 ar_evaluator.py --target study.json" \
  --metric feasibility_composite --direction higher
/ar:loop custom/trial-feasibility

Isolated: no hard dependency — autoresearch runs only on demand, and the loop edits study.json, never the evaluator (locked ground truth).

References

• references/study_design_canon.md — ICH E8(R1) general considerations; ICH E9 + E9(R1) estimand addendum; CONSORT 2010; SPIRIT 2013; FDA Multiple Endpoints guidance (2022).
• references/endpoint_and_power.md — Cohen *Statistical Power Analysis*; Schoenfeld (1983) survival sample size; FDA Surrogate Endpoint Table / BEST glossary; FDA PRO guidance (2009); Chow, Shao & Wang *Sample Size Calculations in Clinical Research*.
• references/trial_operations.md — ICH E6(R2/R3) GCP; TransCelerate risk-based monitoring; FDA RBM guidance; CTTI recruitment best practices; site-feasibility scoring literature.

Assumptions

• Sample-size formulas use normal approximations with a built-in z-table. They are first-pass estimates; a biostatistician produces the final justification (and may use simulation, adaptive designs, or exact methods).
• The endpoint scorer applies *customary* regulatory priors per development area via --profile. Company- or indication-specific precedent overrides the prior.
• The phase-gate scorer bakes in a profile cost-per-patient benchmark; pass a real budget to override the default.
• An unvalidated surrogate cannot anchor a PRIMARY endpoint — the scorer enforces this with a penalty.

Anti-patterns

• Presenting a power estimate as fact. Every output is an estimate with a named owner who must sign.
• Powering for a convenience effect size. The effect must trace to a published or anchor-based MCID, not to the n you can afford.
• Anchoring a primary on an unvalidated surrogate. Surrogate endpoints need validation evidence for the indication.
• Ignoring multiplicity. More than one primary endpoint requires pre-specified alpha allocation.
• Skipping dropout inflation. Raw n undersizes the study; inflate by 1/(1 − dropout).

Distinct from

| Sibling / neighbor | Scope | Difference |

|---|---|---|

| ra-qm-team | ISO 13485 QMS, ISO 14971 risk, EU MDR tech docs + clinical evaluation, FDA 510(k)/PMA/De Novo/QSR submission | That is the submission; clinical-research designs the study beforehand |

| research/grants | NIH funding discovery + positioning | That finds funding; this designs the trial |

| product-team/experiment-designer | Live product A/B hypothesis + sample size | That is a product experiment; this is a clinical trial |

| research-finance (sibling) | R&D program budget + burn | That funds the program; this scopes the study |

Quick examples

python3 scripts/sample_size_estimator.py --sample
python3 scripts/sample_size_estimator.py --design proportions --p1 0.30 --p2 0.45 --dropout 0.15
python3 scripts/endpoint_selector.py --sample
python3 scripts/phase_gate_scorer.py --sample --output json

The sample correctly flags an unvalidated serum-cytokine surrogate (cannot be primary) and ranks PASI-75 as the PRIMARY endpoint; the phase-gate sample returns a verdict with a named owner chain.

Forcing-question library (Matt Pocock grill discipline)

Walked one at a time by /cs-grill-research-ops or the orchestrator. Recommended answer + canon citation per question. Never bundled.

1. "Is your primary endpoint a clinical outcome or a surrogate — and if surrogate, is it on FDA's validated table?"

Recommended: clinical outcome unless the surrogate is validated for this indication.

Canon: FDA Surrogate Endpoint Table; BEST (Biomarkers, EndpointS, and other Tools) glossary.

2. "What's the minimal clinically important difference you're powering for — and where did that number come from?"

Recommend

… (nội dung đầy đủ có trong gói tải về)