huang-sir1 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, notP < .05); useP < .001only 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.
- Discrimination is not enough for a clinical model. Report calibration and clinical utility (decision-curve) alongside AUC.
- 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. Return runnable code (Python first; R where it is the field standard) with the software/version and the exact method for CIs.
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 design and analysis.
- 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.
- Survival/prognosis: Kaplan-Meier + log-rank, Cox, C-index, time-dependent ROC, competing risks.
- Planning: sample size for accuracy / AUC; EPV and Riley minimum sample size for prediction models.
When to open extra files
| File | Open when |
|---|---|
| references/diagnostic-accuracy.md | Sensitivity/specificity/PPV/NPV/LR, the right CI method, paired comparison (McNemar) |
| references/model-evaluation.md | ROC/AUC, DeLong vs bootstrap, thresholds, calibration, Brier, decision-curve analysis |
| references/agreement-mrmc.md | Cohen/weighted/Fleiss kappa, ICC model choice, Bland-Altman, MRMC (Obuchowski-Rockette / DBM) |
| references/high-dimensional-omics.md | Multiple testing (Bonferroni/Holm/BH-FDR/q-values), CV/nested CV, leakage, optimism, harmonisation stats |
| references/survival-prognostic.md | Kaplan-Meier, Cox PH (+ assumptions), C-index, time-dependent AUC, competing risks |
| references/sample-size.md | Sample-size for sensitivity/specificity/AUC; EPV; Riley minimum sample size for prediction models |
Workflow
- Restate the design — unit of analysis (patient/lesion/slice), paired vs unpaired, number of readers, prevalence, primary vs secondary endpoints.
- Pick the estimand and test using the reference files. Name it explicitly.
- Choose the CI method (e.g. Wilson/Clopper-Pearson for proportions; DeLong or bootstrap for AUC; bootstrap for derived metrics).
- Handle multiplicity — declare the primary analysis; correct the rest (method + family).
- Run it — provide runnable code; compute estimate + CI (+ exact p where a test applies).
- Write the result — a Radiology-style sentence (estimate, CI, p, n) plus a Methods sentence (test, software/version, CI method, multiplicity).
- Sanity-check — does the CI width match n? is the test paired if the data are? is calibration reported for a clinical model? are subgroups pre-specified?
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: 0.72, 0.85; 158/200)."
- AUC comparison: "The model's AUC (0.88; 95% CI: 0.84, 0.92) exceeded the radiologists' (0.81; 95% CI: 0.76, 0.86; difference 0.07; P = .004, DeLong)."
- 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 1218 features, 47 differed after Benjamini-Hochberg control at FDR < 0.05."
Output contract
Design read— unit, pairing, readers, prevalence, endpoints.Recommended analysis— estimand, test, CI method, multiplicity plan.Code— runnable Python (and/or R), with library versions noted.Results sentence— Radiology-style, with placeholders only where the user must supply data.Methods sentence— for the statistical-analysis paragraph.Caveats— assumptions, when the test breaks, what the reviewer may ask.
Integrity & handoffs
- Never invent numbers, CIs, or p-values; compute from supplied data or mark as needed.
- Reporting-guideline alignment of the statistics →
radiology-reporting. - Plotting the result (ROC, calibration, DCA, forest, KM) →
radiology-figure. - High-dimensional study design (leakage, batch effects in radiogenomics) →
radiology-radiogenomics. - 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. - This skill is statistical guidance, not a substitute for a qualified biostatistician on high-stakes or regulatory work.