Homeostasis-Inspired, Quality-Adaptive Preprocessing for Multi-Label ECG Diagnosis: A Pre-Registered Multi-Objective Evolutionary Study with Negative Findings and Falsifiable Design Principles

Deep-learning electrocardiogram (ECG) classifiers usually apply one fixed preprocessing filter to every recording, although signal quality varies. Inspired by physiological homeostasis, ECG preprocessing was formulated as a sensor–controller–effector loop: a weakly supervised quality sensor assigns each 12-lead record to a quality tier, and a tier-gated policy, shaped by Non-dominated Sorting Genetic Algorithm II (NSGA-II) on the validation fold, selects a filtering operator and a calibration temperature for a frozen compact Convolutional Neural Networks (CNN). In a pre-registered evaluation on PTB-XL v1.0.3 (21,388 records), the adaptive policy did not outperform the best fixed band-pass filter in discrimination (test macro area under the receiver operating characteristic curve (AUROC 0.9225 vs. 0.9231), worst-tier performance, or calibration, on native data or on any of 15 synthetic and real (MIT-BIH Noise Stress Test) noise benches. Like the fixed filter, it clearly outperformed unfiltered processing under baseline wander (ΔAUROC +0.0047, Holm-adjusted p < 0.001). Post hoc analyses explain these negative results: the operator bank offered no tier-dependent trade-off, the worst-tier objective was rarely binding, the sensor saturated because its tier boundaries were fitted on clean data, the sensor cost exceeded the filtering it saved, and the selected operators were unstable across resamples, while the qualitative rule “filter only the disturbed tier” was stable. Of four predictions derived from homeostasis, one was supported, two were supported only partially, and one was refuted. This study distills six falsifiable design principles for bio-inspired adaptive signal processing.