Adversarial Multitask Learning Reduces the Correlation between Age and Deep Learning Predictions of Myocardial Infarction from Electrocardiograms

Deep Learning (DL) models have exhibited high performance in automated ECG diagnosis, yet some were developed onsmall datasets or biased towards a certain clinical condition. This study applies adversarial multitask learning (AML), a technique that trains a DL model by optimizing contrastive objectives, to identify myocardial infarction (MI) from ECG signals while mitigating the influence of age on model predictions. ECG recordings from healthy and MI subjects were extracted from the PTB-XL dataset and preprocessed to generate 12-lead average beats. Two DLmodelssharing the initial layers were trained. The first model was trained to identify MI, while the second to predict patient’s age, with the parameters of the common layers frozen. Finally, the parameters of the common layers were trained to minimize the classification loss while maximizing the age prediction error, using two loss functions: i) mean squared error (MSE); and ii) negative squared covariance (NCOV). On the validation set, the first model achieved a classification accuracy of 0.87 while the second one had a Pearson’s correlation coefficient (PCC) with age of 0.67. After adversarial training with MSE and NCOV, PCCs with age were-0.78 and-0.03, and accuracies were 0.82 and 0.85, respectively. The proposed AML reduced the correlation between true and predicted age while keeping a good performance for MI identification.