AIRCHILL Statistical Analysis Plan
Statistics · development SAP v0.1
AIRCHILL Statistical Analysis Plan
This shared SAP framework defines common principles and indication-specific primary analyses. It must be finalised and signed before database lock and, for confirmatory work, before any sponsor access to unblinded comparative results.
1. Analysis populations
- ITT: all randomised participants analysed according to assignment; primary efficacy population.
- Safety: participants analysed according to actual AIRCHILL exposure for device-related safety, while preserving ITT safety summaries in parallel.
- Per-protocol: prespecified major-deviation-free subset, supportive only.
- Thermal-exposure set: participants with interpretable device telemetry for mechanistic dose-response analyses.
2. Primary estimands
Cardiac arrest: treatment-policy effect of assignment to early AIRCHILL strategy on the locked 90-day neurological endpoint, with death incorporated according to endpoint definition and crossover/treatment interruption retained under assignment.
Stroke: treatment-policy effect on 90-day ordinal mRS; death remains mRS 6. Crossovers and incomplete exposure are handled under assignment.
3. Primary models
CA: if dichotomous favourable outcome is confirmatory, adjusted binomial/logistic model with absolute risk difference, risk ratio/odds ratio and 95% CI reported; full ordinal mRS remains key supportive analysis. Stroke: proportional-odds ordinal logistic regression, adjusted common OR with 95% CI, formal assessment of proportional-odds assumption and prespecified robust alternative if materially violated.
4. Covariate adjustment
Use a small prespecified set of strong baseline prognostic factors. CA candidates: age, initial rhythm, witnessed arrest, bystander CPR, time-to-ROSC. Stroke candidates: age, baseline NIHSS, ASPECTS/core burden, onset/last-known-well to randomisation. Avoid adjusting the primary model for post-randomisation mediators.
5. Multiplicity
One primary endpoint per pivotal study. Key secondary outcomes are ordered in a hierarchical gatekeeping sequence if confirmatory claims are intended. Imaging, biomarkers, dose-response and most subgroup analyses remain supportive/exploratory unless promoted prospectively.
6. Missing data
Primary prevention strategy is operational: central follow-up, multiple contact routes, blinded outcome assessors and retrieval from routine records. Missing-data analysis includes multiple imputation under plausible MAR assumptions plus prespecified tipping-point/worst-reasonable-case sensitivity. Death is an outcome, not ordinary missing data.
7. Intercurrent events
Device interruption, crossover, rescue temperature therapy, withdrawal of consent, WLST and protocol deviations are explicitly classified. The primary treatment-policy estimand generally retains these events; hypothetical or principal-stratum analyses, if used, are supportive and clearly labelled.
8. Interim analysis
One formal confirmatory interim near 50% information is the base design. A group-sequential alpha-spending approach preserves overall type-I error. Futility is non-binding. Exact spending function and boundaries are generated before first unblinded DSMB review.
9. Sample-size re-estimation
Blinded nuisance-parameter updates may use pooled control/event distributions, mRS distribution or missingness if prospectively specified. Any unblinded effect-based adaptation requires a fully defined adaptive method preserving type-I error.
10. Subgroups
Interaction tests—not within-subgroup p-values—drive interpretation. Prespecified candidates include treatment-start time, achieved thermal exposure, age, sex and baseline temperature. CA additionally initial rhythm/time-to-ROSC; stroke additionally baseline core/ASPECTS, reperfusion status and collateral grade where available. Findings are supportive unless separately powered.
11. Safety analyses
Summarise treatment-emergent AE/SAE, serious adverse device effects, device deficiencies, ventilation compromise, arrhythmia/hemodynamic events, pneumonia and mortality with exposure-adjusted summaries where useful. Stroke adds sICH and treatment-attributable reperfusion delay.
12. Sensitivity analyses
- Unadjusted versus adjusted primary model.
- Per-protocol and as-treated supportive analyses.
- Alternative missing-data assumptions.
- Centre/random-effects sensitivity where site heterogeneity is material.
- CA ordinal neurological outcome sensitivity.
- Stroke generalized-odds or win/utility-weighted sensitivity if proportional odds is untenable.
13. Reporting
Effect estimates with confidence intervals take precedence over isolated p-values. All deviations from the signed SAP are dated, justified and reported. Statistical code is version controlled and independently validated for the primary analysis.