Abstract: Objectives　To establish three convolutional neural network models for automatic diagnosis of premature ventricular contractions（PVC） in children by applying the method of computer deep learning and to evaluate their diagnostic value. Methods　ECGs of 1200 children with premature ventricular contractions collected in Shanghai Children's Hospital were used as PVC group and ECGs of 1200 normal children in the same age and sex were enrolled as normal control group. The male to female ratio of the two groups was 3:2, with an average age of 6.5 ± 0.5 years. In two groups, 800 samples were randomly extracted after eliminating few ECGs that are not suitable for model training, and three convolutional neural network models for automatic diagnosis of PVC in children were trained and established by applying the method of computer deep learning. An other 200 samples were extracted from each group to test and validate the performance of each model. The diagnosis of electrocardiogram expert group of Department of Cardiology in Shanghai Children's Hospital is regarded as "gold standard" to evaluate the reliability and validity of the models using medical statistics. Results　Applying the method of computer deep learning, 2D CNN model and fine-tuned Inception-V3 model are trained with ECG waveform image, 1D CNN model are trained with ECG time-series data. The sensitivity of the 2D CNN model is 65%, the specificity 71.5%, the missed diagnosis rate 35%, the misdiagnosis rate 28.5%, the positive predictive value 69.5%, the negative predictive value 67.1%, the accuracy rate 68.2%, and the Kappa value 0.365. The sensitivity of the V3 model is 82%, specificity 85%, missed diagnosis rate 18%, misdiagnosis rate 15%, positive predictive value 84.5%, negative predictive value 82.5%, accuracy 83.5%, Kappa 0.670; The sensitivity of 1D CNN model is 87.5%, specificity 89.5%, missed diagnosis rate 12.5%, misdiagnosis rate 10.5%, positive predictive value 89.3%, negative predictive value 87.7%, accuracy rate 88.5%, Kappa 0.770. Conclusions　V3 model and 1D CNN model perform well, and the reliability and validity of 1D CNN model are especially good, Kappa value 0.77 indicates highly consistent with the diagnosis of electrocardiogram expert group of Department of Cardiology in Shanghai Children's Hospital.

Key words: premature ventricular contractions;　deep learning;　convolutional neural network models;　child