Manual and semi-automatic identification of artifacts and unwanted physiological signals in large intracerebral electroencephalographic (iEEG) recordings is time consuming and inaccurate. To date, unsupervised methods to accurately detect iEEG artifacts are not available. This study introduces a novel machine-learning approach for detection of artifacts in iEEG signals in clinically controlled conditions using convolutional neural networks (CNN) and benchmarks the method’s performance against expert annotations. The method was trained and tested on data obtained from St Anne’s University Hospital (Brno, Czech Republic) and validated on data from Mayo Clinic (Rochester, Minnesota, U.S.A). We show that the proposed technique can be used as a generalized model for iEEG artifact detection. Moreover, a transfer learning process might be used for retraining of the generalized version to form a data-specific model. The generalized model can be efficiently retrained for use with different EEG acquisition systems and noise environments. The generalized and specialized model F1 scores on the testing dataset were 0.81 and 0.96, respectively. The CNN model provides faster, more objective, and more reproducible iEEG artifact detection compared to manual approaches.
Objective : This paper describes a method for automated discrimination of heart sounds recordings according to the Physionet Challenge 2016. The goal was to decide if the recording refers to normal or abnormal heart sounds or if it is not possible to decide (i.e. ‘unsure’ recordings). Approach : Heart sounds S1 and S2 are detected using amplitude envelopes in the band 15–90 Hz. The averaged shape of the S1/S2 pair is computed from amplitude envelopes in five different bands (15–90 Hz; 55–150 Hz; 100–250 Hz; 200–450 Hz; 400–800 Hz). A total of 53 features are extracted from the data. The largest group of features is extracted from the statistical properties of the averaged shapes; other features are extracted from the symmetry of averaged shapes, and the last group of features is independent of S1 and S2 detection. Generated features are processed using logical rules and probability assessment, a prototype of a new machine-learning method. Main results : The method was trained using 3155 records and tested on 1277 hidden records. It resulted in a training score of 0.903 (sensitivity 0.869, specificity 0.937) and a testing score of 0.841 (sensitivity 0.770, specificity 0.913). The revised method led to a test score of 0.853 in the follow-up phase of the challenge. Significance : The presented solution achieved 7th place out of 48 competing entries in the Physionet Challenge 2016 (official phase). In addition, the PROBAfind software for probability assessment was introduced.
The automated detection of arrhythmia in a Holter ECG signal is a challenging task due to its complex clinical content and data quantity. It is also challenging due to the fact that Holter ECG is usually affected by noise. Such noise may be the result of the regular activity of patients using the Holter ECG—partially unplugged electrodes, short-time disconnections due to movement, or disturbances caused by electric devices or infrastructure. Furthermore, regular patient activities such as movement also affect the ECG signals and, in connection with artificial noise, may render the ECG non-readable or may lead to misinterpretation of the ECG. Objective : In accordance with the PhysioNet/CinC Challenge 2017, we propose a method for automated classification of 1-lead Holter ECG recordings. Approach : The proposed method classifies a tested record into one of four classes—‘normal’, ‘atrial fibrillation’, ‘other arrhythmia’ or ‘too noisy to classify’. It uses two machine learning methods in parallel. The first—a bagged tree ensemble (BTE)—processes a set of 43 features based on QRS detection and PQRS morphology. The second—a convolutional neural network connected to a shallow neural network (CNN/NN)—uses ECG filtered by nine different filters (8× envelograms, 1× band-pass). If the output of CNN/NN reaches a specific level of certainty, its output is used. Otherwise, the BTE output is preferred. Main results : The proposed method was trained using a reduced version of the public PhysioNet/CinC Challenge 2017 dataset (8183 records) and remotely tested on the hidden dataset on PhysioNet servers (3658 records). The method achieved F1 test scores of 0.92, 0.82 and 0.74 for normal recordings, atrial fibrillation and recordings containing other arrhythmias, respectively. The overall F1 score measured on the hidden test-set was 0.83. Significance : This F1 score led to shared rank #2 in the follow-up PhysioNet/CinC Challenge 2017 ranking.
10 July 2017 in Journal of Interventional Cardiac Electrophysiology
Purpose The aim of this proof-of-concept study is to introduce new high-dynamic ECG technique with potential to detect temporal-spatial distribution of ventricular electrical depolarization and to assess the level of ventricular dyssynchrony. Methods 5-kHz 12-lead ECG data was collected. The amplitude envelopes of the QRS were computed in an ultra-high frequency band of 500–1000 Hz and were averaged (UHFQRS). UHFQRS V lead maps were compiled, and numerical descriptor identifying ventricular dyssynchrony (UHFDYS) was detected. Results An electrical UHFQRS maps describe the ventricular dyssynchrony distribution in resolution of milliseconds and correlate with strain rate results obtained by speckle tracking echocardiography. The effect of biventricular stimulation is demonstrated by the UHFQRS morphology and by the UHFDYS descriptor in selected examples. Conclusions UHFQRS offers a new and simple technique for assessing electrical activation patterns in ventricular dyssynchrony with a temporal-spatial resolution that cannot be obtained by processing standard surface ECG. The main clinical potential of UHFQRS lies in the identification of differences in electrical activation among CRT candidates and detection of improvements in electrical synchrony in patients with biventricular pacing.