Trust Before Compression: Towards a Trustworthy Compression Frontier for ECG Classification

Compression is essential for deploying ECG models under memory and compute constraints, yet accuracy alone provides an incomplete view of compression-induced behavioral changes in safety-sensitive biosignals. A compressed model may preserve accuracy while altering calibration, perturbation robustness, or prediction stability. We present a systematic empirical study of the \emph{trustworthy compression frontier} for ECG classifiers, evaluating each compressed model relative to its full-precision counterpart using accuracy, expected calibration error (ECE), additive-noise robustness, and prediction stability. These changes are summarized by a transparent Trust Collapse Index (TCI) and an explicit operating criterion. We evaluate a multilayer perceptron (MLP), ResNet1d, and ECG HuBERT on Chapman and CPSC 2018, covering FP32/FP16 and 1--8-bit quantization, 0 to 87.5\% unstructured magnitude pruning, and matched-budget pruning--quantization in both orders. The results reveal substantial architecture- and dataset-dependent compression boundaries. Chapman ResNet1d retains 94.71\% accuracy and 94.66\% noise robustness at 3-bit quantization using only 9.4\% of the nominal FP32 parameter-storage budget, while CPSC ResNet1d remains essentially unchanged at 5 bits, achieving 68.10\% accuracy and a trust score of 0.998 at 15.6\% budget. In contrast, aggressive pruning and lower-bit quantization can produce abrupt behavioral degradation. Sequential compression further reveals a pronounced order effect: at the same nominal budget of 0.141 on Chapman ResNet1d, Q$\rightarrow$P achieves 94.33\% accuracy and TCI 0.104, compared with 90.27\% and TCI 0.806 for P$\rightarrow$Q. These are single-seed, single-run observations rather than statistically validated population effects, and TCI is an operational evaluation heuristic rather than a clinical safety metric. Overall, the results suggest that ECG compression should be characterized not by compression ratio alone, but by the region in which predictive utility and measured behavioral trustworthiness are jointly preserved.