IBERIA asks whether echocardiographic radiomic features — quantitative descriptors of texture, intensity, and shape extracted from standard 2D echo views — carry prognostic information for atrial fibrillation recurrence beyond what established clinical predictors already provide.
The answer, after substantial methodological care, is no. Across nested cross-validation, Meinshausen–Bühlmann stability selection, and Bayesian variable selection with horseshoe priors, no radiomic feature achieves a posterior inclusion probability or stability frequency comparable to the routine clinical predictors. This is a characterised negative result — one in which the absence of signal has been distinguished from the absence of power, signal-to-noise has been quantified via Fisher’s discriminant ratio and Cohen’s d, and alternative explanations have been systematically ruled out.
The methodological infrastructure built for IBERIA — the stability-selection audit, the Bayesian variable-selection pipeline with prior sensitivity analysis, and the conformal prediction layer — has since been generalised into the opsis R package and is being reused across CardioNIR and DynaCARD.
Sub-project pages (study design, feature pipeline, stability-selection audit, Bayesian model card, methodological retrospective) are forthcoming.