Deep learning (DL) has achieved promising performance in detecting common abnormalities from the 12-lead electrocardiogram (ECG). However, diagnostic redundancy exists in the 12-lead ECG, which could impose a systematic overfitting on DL, causing poor generalization. We, therefore, hypothesized that finding an optimal lead subset of the 12-lead ECG to eliminate the redundancy would help improve the generalizability of DL-based models. In this study, we developed and evaluated a DL-based model that has a feature extraction stage, an ECG-lead subset selection stage and a decision-making stage to automatically interpret multiple common ECG abnormality types. The data analysed in this study consisted of 6877 12-lead ECG recordings from CPSC 2018 (labelled as normal rhythm or eight types of ECG abnormalities, split into training (approx. 80%), validation (approx. 10%) and test (approx. 10%) sets) and 3998 12-lead ECG recordings from PhysioNet/CinC 2020 (labelled as normal rhythm or four types of ECG abnormalities, used as external text set). The ECG-lead subset selection module was introduced within the proposed model to efficiently constrain model complexity. It detected an optimal 4-lead ECG subset consisting of leads II, aVR, V1 and V4. The proposed model using the optimal 4-lead subset significantly outperformed the model using the complete 12-lead ECG on the validation set and on the external test dataset. The results demonstrated that our proposed model successfully identified an optimal subset of 12-lead ECG; the resulting 4-lead ECG subset improves the generalizability of the DL model in ECG abnormality interpretation. This study provides an outlook on what channels are necessary to keep and which ones may be ignored when considering an automated detection system for cardiac ECG abnormalities. This article is part of the theme issue 'Advanced computation in cardiovascular physiology: new challenges and opportunities'.