Time Reversal Fusion for ECG Reconstruction Before and After Adversarial Fine-Tuning

Temporal reversal preserves electrocardiogram (ECG) samples while changing their presentation to a reconstruction model. We evaluated forward/reverse (F/R) fusion against matched forward/forward (F/F) ensembles using twelve-lead, seven-second windows from the normal-labeled CODE-15% subset. Each seed block comprised two forward branches and one reverse branch, with paired initialization for the second forward and reverse branches. Reverse outputs were restored to the original time axis before fusion. Across ten seed blocks of a temporal convolutional model trained with L₁ loss, equal F/R fusion reduced test mean absolute error (MAE) by 1.437% relative to equal F/F, improving every block. Local adaptive F/R fusion reduced MAE by 2.140% against the same baseline. A matched continuation experiment compared L₁ training with least-squares adversarial fine-tuning across six seed blocks. After adversarial training, equal F/R retained a mean MAE reduction of 0.717%, improving all six blocks. Lower residual cross-moments were consistent with complementary reconstruction errors. Equal F/R also improved seed-averaged amplitude error, correlation and R-peak F₁ across the original-input condition and eighteen noise settings. Boundary-excluded evaluation and symmetric-padding controls retained mean directional benefits. These results support temporal direction as a source of reconstruction diversity beyond ordinary ensembling, with benefits retained under adversarial training.