Chagas disease is a condition that can cause significant cardiac complications. Early detection through automated electrocardiogram analysis may facilitate timely intervention for treatment. In this study, we present a deep learning pipeline for the binary classification of 12-lead ECGs into Chagas and non-Chagas categories. All ECG recordings were subsampled to 100 samples per second, when necessary, and then bandpass filtered to remove baseline drift and high-frequency noise. Each lead signal was subsequently decomposed using the Discrete Wavelet Transform up to level 4, using a Daubechies 3 mother wavelet. The resulting approximation and detail coefficients from the fourth level were concatenated to feed a 1D Convolutional Neural Network (CNN), designed to capture local patterns across the multi-lead input. The neural network architecture consists of 12 parallel convolutional networks, each receiving input from different patient leads. Each of these parallel networks follows the same structure: a convolutional layer with a ReLU activation function, followed by a max-pooling layer, then another convolutional layer with ReLU activation, and another max-pooling layer. After processing through these layers, each parallel network's output is flattened into a single vector. These vectors from all 12 parallel networks are then concatenated into one comprehensive vector, which is passed through three hidden layers. Finally, the network classifies the patients into two categories: those with Chagas disease and those without. The model was trained with the Samitrop and Code 15% databases, with 16,506 patients for training and 5,126 for validation to optimize binary cross-entropy loss. The model performed with an accuracy of 0.6522, precision 0.6661, recall 0.5136 and F1-score 0.58. The proposed approach demonstrates that improvement is necessary in calculating characteristics that are given to the model, changing the mother wavelet or using parallel networks for approximation and details.