version: "1.0.0" name: radiology-stats description: "Plan/compute/audit inference on defined data/effect rows: CI, AUC, MRMC, survival, meta-analysis and power. CN: 统计、样本量、AUC比较、校准"

Imaging Biostatistics for _Radiology_

Use this skill to choose the right test, run it correctly, and report it the way _Radiology_ wants — estimates with 95% CIs, exact p-values, named tests, and multiplicity handled. It covers the statistics that imaging-AI, radiomics, and reader studies live or die on.

Core stance

• Estimate + uncertainty, not just p. Every primary result gets a 95% CI. Report exact

p-values (e.g. P = .03, not P < .05); use P < .001 only below that floor.

• The test must match the design. Paired data → paired test (same patients/cases read by

both methods); clustered data (multiple lesions per patient) → account for clustering; multiple readers → MRMC, not a naive average.

• Match performance measures to use. A probabilistic diagnostic/prognostic model generally

needs discrimination and calibration. Evaluate decision-analytic utility only when an intended decision, threshold range, consequences and comparator strategies are defined; do not require a decision curve for every exploratory classifier.

• Control multiplicity honestly. Thousands of radiomic/omic features ⇒ FDR or stronger;

pre-specify primary vs exploratory.

• No fishing, no fabrication. Pre-specify the primary analysis; never invent a number,

a CI, or a p-value. If data are insufficient, say what is needed.

• Reproducible. Always return an executable statistical brief. Return runnable code and computed

results only when the required data/schema, estimand, hierarchy, environment and execution authority are available; record software/version and the exact uncertainty method.

• Evidence state is explicit. Use PLAN_ONLY, DATA_READY, COMPUTED, AUTHOR_REPORTED,

PARTLY_VERIFIED, NOT_ASSESSABLE, or BIOSTATISTICIAN_REQUIRED. A proposed model is not a run.

• No universal sample-size threshold. Power and precision depend on the estimand, endpoint

distribution, effect worth detecting, allocation unit, clustering/repeated structure, multiplicity, attrition and acceptable uncertainty. Do not substitute an EPV, event-count or replicate rule of thumb for those inputs.

Choose the route

Declare one mode:

If the request crosses routes, keep separate estimands and briefs. Open only the active specialized reference: meta-analysis, experimental inference/power, quantitative measurement, causal clinical impact, or longitudinal/multistate imaging. Persist the meta-analysis or experimental specification with the meta-analysis brief or experimental inference/power brief.

When to use

• Diagnostic accuracy: sensitivity/specificity/PPV/NPV/accuracy/likelihood ratios + CIs.
• Comparing models/readers/tests: DeLong or bootstrap for AUCs; McNemar for paired

sensitivity/specificity.

• Reader studies: kappa / weighted kappa / Fleiss / ICC / Bland-Altman; MRMC analysis and

sample sizing (clinical reader-study design stays with radiology-translation).

• Prediction models: ROC, calibration (slope/intercept, Brier), **decision-curve

analysis**, threshold selection.

• Radiomics/omics: feature reproducibility (ICC), multiple-testing correction,

cross-validation/nested CV, bootstrap optimism.

• Statistical inference for already-defined configuration, method, ablation or sensitivity

comparisons: paired deltas, cluster bootstrap, run variability and multiplicity. The scientific comparison design itself remains with radiology-method-evaluation.

• Survival/prognosis: Kaplan-Meier + log-rank, Cox, C-index, time-dependent ROC,

competing risks.

• Quantitative imaging measurement: bias, repeatability/reproducibility, wCV, repeatability

coefficient, limits of agreement, smallest detectable change and scanner/site variance.

• Causal/clinical impact: randomized/pragmatic, cluster/stepped-wedge, ITS/DiD or target-trial

analyses with explicit identification assumptions and interference/time-varying confounding.

• Longitudinal imaging: mixed/joint/landmark/multistate models, interval censoring, informative

visit/dropout processes and competing events.

• Planning: sample size/power/precision for accuracy, AUC, prediction models and experiments,

using input-driven methods rather than universal EPV, event-count or replicate cutoffs.

• 中文高频说法：“样本量/统计功效怎么算？”“两个 AUC 怎么比较（DeLong 还是 bootstrap）？”

“校准曲线/决策曲线怎么做？”“多阅片者一致性（kappa/ICC/MRMC）怎么报？”“meta 分析怎么合并？”

When to open extra files

Workflow

1. Declare route, evidence and execution state. Inventory the supplied protocol, effect rows,

raw/summary data, code/output and authority. Never turn a brief or author-reported result into COMPUTED.

1. Restate the design and hierarchy — independent experimental unit, patient/lesion/slice/

timepoint nesting, paired vs unpaired, repeated measures, readers, prevalence, primary vs secondary endpoints. Open analysis-hierarchy-and-missingness.md when clustering or missingness exists; do not count images/cells/model runs as independent patients.

1. Pick the estimand and model/test using the route reference. Name the population quantity,

contrast, endpoint, time/measurement period and intercurrent/missing-data handling explicitly.

1. Choose the uncertainty method (e.g. Wilson/Clopper-Pearson for single-study proportions; DeLong or

bootstrap for AUC; bootstrap for derived metrics).

1. Handle dependence, missingness, exclusions, attrition and deviations before looking at significance;

define the analysis population and sensitivity analyses.

1. Handle multiplicity — declare the primary analysis; correct the rest (method + family).
2. Run the stop gate. If an estimand, independent/allocation unit, dependence structure, effect

definition or required power input cannot be made defensible, return BIOSTATISTICIAN_REQUIRED with the exact blocking decision and do not present a final model or sample size.

1. Specify, then run only when ready. Complete the route template. Provide code/configuration and

compute estimate + interval (+ p where applicable) only from sufficient verified inputs.

1. Write the result — a _Radiology_-style sentence (estimate, CI, p, n) plus a Methods

sentence (test, software/version, CI method, multiplicity).

1. Sanity-check — does the interval width align with information size? is the model paired,

cluster/dependence-aware if needed? is calibration reported for a clinical model? are subgroups pre-specified? Subgroup discipline:

• A subgroup claim needs an interaction test (treatment × subgroup P) — "significant in

one subgroup, not the other" is not evidence that the effect differs.

• Multiplicity compounds: #subgroups × #outcomes is the family-wise risk. Pre-specify a

small number of subgroups; label everything else exploratory.

• Show subgroup results as a forest plot including the interaction P → hand off to

radiology-figure.

• For prediction models, prespecify clinically material subgroup/fairness evaluations when the

intended population, plausible harms and available sample sizes make them relevant. Report what was and was not assessable; do not manufacture underpowered subgroup claims merely to populate a checklist.

1. In mentor mode, do not prescribe a larger model or sample by fashion. Separate:

• minimum defensible: the narrowest estimand/model/precision that current inputs can support;
• standard publishable: the comparator, dependence handling, diagnostics and sensitivity needed

for the intended claim;

• stronger/discriminating: additional design or information that resolves the named ambiguity.

For each, state why, prerequisites, resource/time burden, expected failure signal, fallback, closure evidence and allowed wording. If key nuisance or allocation inputs remain unknown, the learner route still ends at BIOSTATISTICIAN_REQUIRED, not a guessed sample size.

Reporting templates (fill from real output — never fabricate)

• Accuracy: *"Sensitivity was 0.87 (95% CI: 0.81, 0.92; 130/149) and specificity 0.79 (95%

CI: 0.73, 0.84; 158/200)."*

• Independent binary-model AUC comparison: *"On the same independent cases, the frozen model's AUC

(0.88; 95% CI: 0.84, 0.92) exceeded the frozen comparator model's (0.81; 95% CI: 0.76, 0.86; difference 0.07; P = .004, paired DeLong)."* Reader-population contrasts require the applicable MRMC analysis rather than this two-model template.

• Agreement: *"Inter-reader agreement was substantial (ICC, 0.82; 95% CI: 0.75, 0.87;

two-way random-effects, absolute agreement, single rater)."*

• Multiplicity: *"Of [N] features, [K] differed after Benjamini-Hochberg control at FDR <

0.05."*

Output contract

1. `Route and evidence state` — route, scope, supplied artifacts, execution authority and one of

the declared states.

1. `Design/estimand read` — population quantity, contrast, endpoint/period, independent or

allocation unit, pairing/readers/clusters and claim boundary.

1. `Hierarchy/missingness read` — nested/repeated/dependent structure, analysis

population, exclusions, missing-data mechanism/handling, and sensitivity analyses.

1. `Executable statistical brief` — input schema/effect rows, model/test, dependence strategy,

uncertainty, heterogeneity or power inputs, multiplicity, diagnostics, sensitivity and success/ failure assertions. Use the active route template.

1. `Code/run record` — when authorized and ready, runnable Python/R plus versions, command,

logs and output locators; otherwise NOT_RUN with blockers.

1. `Results sentence` — _Radiology_-style, with placeholders only where the user must

supply data.

1. `Methods sentence` — for the statistical-analysis paragraph.
2. `Table handoff` — canonical estimates/value keys and the table shell needed by

radiology-table when a publication table is requested.

1. `Stop gate and caveats` — PASS, CONDITIONAL, or BIOSTATISTICIAN_REQUIRED, exact

assumptions/blocker, minimum missing input, sensitivity boundary and reviewer risk.

1. `Learner route receipt` — in mentor: minimum defensible / standard publishable /

stronger-discriminating options, prerequisites, cost, failure signal, fallback, closure evidence and the claim that survives if the stronger route is infeasible.

Integrity & handoffs

• Never invent numbers, CIs, or p-values; compute from supplied data or mark as needed.
• Never invent covariance, person-time, within-study pairing, cluster ICC, repeated-measure

correlation, baseline event rate, variance, effect worth detecting, attrition or power inputs.

• BIOSTATISTICIAN_REQUIRED stops a final model/sample-size recommendation; it does not authorize

guessing or relabel an unresolved plan as completed analysis.

• Reporting-guideline alignment of the statistics → radiology-reporting.
• Plotting the result (ROC, calibration, DCA, forest, KM) → radiology-figure.
• Publication tables (Table 1, Cox/logistic, model performance, MRMC) → radiology-table.
• High-dimensional study design (leakage, batch effects in radiogenomics) →

radiology-radiogenomics.

• Decide which parameters/methods/comparators need evaluation and whether the evidence supports the

methodological claim → radiology-method-evaluation; return here only for statistical inference.

• Sample-size numbers feeding a grant's feasibility section → radiology-grant.
• Analysis plan is locked and results are in; want a harsh pre-submission read → radiology-prereview.
• Project value registry, analysis lock, and cross-artifact state → radiology-pipeline.
• This skill is statistical guidance, not a substitute for a qualified biostatistician on

high-stakes or regulatory work.