A Random Forest classifier using amplitude-invariant features outperformed baseline (AUC 0.830 vs 0.782), driven by functional re-weighting rather than representational continuity (p=0.168).
Predictive success in transfer learning from synthetic attractors to clinical ECGs relies on domain-adaptive functional re-weighting rather than geometric invariance.
Estimación del efecto: ΔK -0.051 (95% CI -0.120, 0.017)
valor p: p=0.168
Transfer learning from synthetic dynamical systems to clinical biosignals often assumes that predictive success implies representational similarity. We test this assumption across three domains-synthetic chaotic attractors, a parametric biophysical simulation, and the MIT-BIH Arrhythmia Database-using an eight-dimensional amplitude-invariant feature space (Embedding V3). Under strict AAMI inter-patient validation, a Random Forest classifier achieved AUC=0. 830, outperforming a power spectral density baseline (AUC=0. 782). A permutation test (N=10, 000 iterations) evaluated the null hypothesis of representational continuity, yielding ΔK=-0. 051 (p=0. 168, CI95%-0. 120, 0. 017), providing no statistically significant evidence of preserved structure. A test of equivalence (TOST) confirmed that any continuity, if present, falls below a practical margin (pTOST < 0. 01). Bootstrap analysis (N=1, 000 resamples) revealed a marked asymmetry: substantial representational divergence in the feature space (Dₑmb=0. 982, CI95%0. 973, 0. 992) alongside moderate attributional reordering (Dₐttr=0. 763, CI95%0. 714, 0. 813). The non-overlapping confidence intervals establish that predictive success under domain shift does not arise from geometric invariance, but from domain-adaptive functional re-weighting of amplitude-independent dynamical features.
Alvaro Lopez Almeida (Sun,) conducted a other in Arrhythmia. Random Forest classifier using an eight-dimensional amplitude-invariant feature space (Embedding V3) vs. Power spectral density baseline was evaluated on Representational continuity (ΔK) (ΔK -0.051, 95% CI -0.120, 0.017, p=0.168). A Random Forest classifier using amplitude-invariant features outperformed baseline (AUC 0.830 vs 0.782), driven by functional re-weighting rather than representational continuity (p=0.168).