Feature Selection Based on a Sparse Neural-Network Layer With Normalizing Constraints.

Feature selection (FS) is an important step in machine learning since it has been shown to improve prediction accuracy while suppressing the curse of dimensionality of high-dimensional data. Neural networks have experienced tremendous success in solving many nonlinear learning problems. Here, we propose a new neural-network-based FS approach that introduces two constraints, the satisfaction of which leads to a sparse FS layer. We performed extensive experiments on synthetic and real-world data to evaluate the performance of our proposed FS method. In the experiments, we focus on high-dimensional, low-sample-size data since they represent the main challenge for FS. The results confirm that the proposed FS method based on a sparse neural-network layer with normalizing constraints (SNeL-FS) is able to select the important features and yields superior performance compared to other conventional FS methods.

Automatic ECG classification and label quality in training data.

Objective.Within the PhysioNet/Computing in Cardiology Challenge 2021, we focused on the design of a machine learning algorithm to identify cardiac abnormalities from electrocardiogram recordings (ECGs) with a various number of leads and to assess the diagnostic potential of reduced-lead ECGs compared to standard 12-lead ECGs.Approach.In our solution, we developed a model based on a deep convolutional neural network, which is a 1D variant of the popular ResNet50 network. This base model was pre-trained on a large training set with our proposed mapping of original labels to SNOMED codes, using three-valued labels. In the next phase, the model was fine-tuned for the Challenge metric and conditions.Main results.In the Challenge, our proposed approach (team CeZIS) achieved a Challenge test score of 0.52 for all lead configurations, placing us 5th out of 39 in the official ranking. Our improved post-Challenge solution was evaluated as the best for all ranked configurations, i.e. for 12-lead, 3-lead, and 2-lead versions of the full test set with the Challenge test score of 0.62, 0.61, and 0.59, respectively.Significance.In addition to building the model for identifying cardiac anomalies, we provide a more detailed description of the issues associated with label mapping and propose its modification in order to obtain a better starting point for training more powerful classification models. We compare the performance of models for different numbers of leads and identify labels for which two leads are sufficient. Moreover, we evaluate the label quality in individual parts of the Challenge training set.