Nonlinear amplitude compression may improve ECG classification, but whether in-domain gains generalize to external datasets remains unclear. Using a one-dimensional ResNet-34, we evaluated eight training-time compression functions: sigmoid, tanh, arctangent, softsign, logarithmic, power-law, Huber-style soft clipping, and adaptive tanh. In-domain performance was assessed by 10-fold cross-validation on CPSC2018 (6,877 recordings; nine classes). Cross-dataset performance was assessed by training and validating on CPSC2018 and testing on 4,175 G12EC recordings from seven shared classes. F1 score was the primary metric. All eight compression strategies outperformed the in-domain baseline (0.7369); tanh performed best (0.7554; +2.51%), followed by power-law (0.7543; +2.36%) and Huber-style clipping (0.7527; +2.15%). Externally, sigmoid achieved the highest F1 (0.5904; +1.74% versus 0.5803) and was the only strategy to significantly outperform the baseline. Tanh reached 0.5863 (+1.03%), whereas Huber-style clipping slightly underperformed the baseline (0.5797; −0.10%). Thus, compression benefits under dataset shift are function-dependent and require external validation.