Recent years have seen significant development in end-to-end methods for the automatic classification of electrocardiograms (ECGs), providing physicians new, state-of-the-art methods supporting the diagnosis of heart conditions via non-invasive screening exams. Depending on the electrocardiograph used and the purpose of the exam, the recorded ECG tracings may have a variable duration and, especially, a variable number of leads used to record the heart activity. For example, ambulatorial electrocardiographs used to detect critical, active heart problems like atrioventricular blocks, generally consist of 8 or 12 leads (placed beneath the heart, at the wrists, and at the ankles). When, instead, arrhythmia is suspected, the patient might be monitored for several hours or days with an Holter monitor, which generally uses 2 to 3 leads. Commercial devices, like wearable devices, are often limited to 1 or 2 leads and short recordings. The aim of this project is to develop a deep learning architecture capable of handling this variability in leads configuration at three different stages: (i) at training stage, to train end-to-end classifiers based on raw signals on multiple datasets having different lead configurations; (ii) at inference stage, to classify ECG traces regardless of its leads configuration while (iii) being robust against possible lead drops (caused, for example, by a loss of electrical contact due to misplacement or patient’s sweating or movement. To this end, architectures based on Residual Convolutional Neural Networks (ResCNN) and/or Selective Structured State-Space Models (S6) will be employed.