• Arrhythmia Classification
• Convolutional Neural Network
• Deep Learning
• ECG Analysis

The Electrocardiogram (ECG) can be regarded as a prime tool in getting information on the cardiac functionality. The diagnostic power of an ECG is related to the ability to timely and accurately identify specific heartbeat alterations (arrhythmia). Several approaches to the automated analysis of ECG signals have been proposed so far, in the attempt to improve the quality of arrhythmia detection.
The ECG signal, like many other biological signals, is generally non-linear, non-stationary, dynamic and complex: this hampers the extraction of characteristics to be used for any algorithmic analysis, with a severe impact on the efficiency of the overall diagnostic procedure. Specifically, the complexity of the ECG signal is related to the high variability of wave-form morphologies both for the same patient and across different patients, plus problems in the placements of electrodes on the patient’s body.
This work aims to increase the accuracy and speed of ECG diagnostic systems with respect to the real-time detection of arrhythmia events. A deep learning (DL) approach is proposed as a means to deal with the high variability in ECG signals. The implemented neural network has shown to be able to detect, starting from raw data, high-level ECG characteristics, with good generalization ability.
The work is developed according to an "intra-patient" paradigm: the network is trained not only using generic samples, but also exploiting patient-specific data. The network consists of three convolutional layers, which are in charge of extracting features out of morphological information in ECG signals; subsequently, by considering also temporal features, the previous information is fed to a multi-layer-perceptron (MLP) to finalize the classification job.
The performance evaluation of the developed classifier has been carried out taking into account different pre-processing strategies, best representation of morphological and temporal features, data augmentation techniques and fine-tuning strategies.
The results over a classical, standard benchmark database (MIT-BIH DB), show that the network performance is comparable (and in some cases superior) to what it has been obtained in other works with the same paradigm.
In accordance with the AAMI recommendations, the most challenging classes to identify are the Supraventricular Ectopic Beat (SVEB) and the Ventricular Ectopic Beat (VEB). For this reason, the performance on this class, based on Sensitivity (Sen) and Positive predicted value (Ppv), are representative of the effective robustness of the overall approach. The network reaches 60.21% and 79.85% for the SVEB class and 91.05% and 95.34% for the VEB class in terms of Sen and Ppv respectively.
Such results show the competitive performance of this work with respect to the state-of-the-art.
Finally, it is important to underline that the shallow depth of the network, as well as the limited number of steps to get to the final classification of input data, make the proposed system particularly well suited for real-time analysis in embedded devices.