Background and objectives: Many Deep Learning (DL) models have been implemented in the cardiovascular domain for the automated ECG diagnosis, reaching very high performances. Despite that, some of them are developed on small datasets or biased towards a certain clinical condition. In the context of automatic ECG interpretation, the identification of myocardial infarction (MI) plays an important role. In this study, we employ the so-called Adversarial Multitask Learning to identify MI from ECG signals trying to mitigate the impact of age on model predictions, since it is well known that DL models can predict patient's age from ECG signals. Method: We selected ECG recordings from healthy controls and MI patients from the PTB-XL available on Physionet. ECGs were preprocessed to generate a 12-lead average beat. These beats were split in training (80%) and validation (20%) sets and used to train two DL models having in common a few layers starting from the input. The first model was trained to identify MI. The second model was trained to predict patient's age while keeping the parameters of the common layers freezed. Finally, the parameters of the common layers were trained by minimizing the classification loss while maximizing the age prediction error. We also compared two losses that were maximized for worsening the age prediction: i) mean squared error (MSE); and ii) negative covariance (NCOV). Results: On the validation set, the first model achieved a classification accuracy of 0.87 while the second model obtained a Pearson's correlation coefficient (PCC) with age of 0.67. After the adversarial training with MSE, PCC with age was -0.78 and accuracy was 0.82, while for NCOV, PCC was -0.03 and accuracy was 0.85. Conclusion: The proposed adversarial multitask training was able to reduce the correlation between true and predicted age while keeping a good performance for MI identification.