Filip Plesinger

SignalPlant: an open signal processing software platform.

The growing technical standard of acquisition systems allows the acquisition of large records, often reaching gigabytes or more in size as is the case with whole-day electroencephalograph (EEG) recordings, for example. Although current 64-bit software for signal processing is able to process (e.g. filter, analyze, etc) such data, visual inspection and labeling will probably suffer from rather long latency during the rendering of large portions of recorded signals. For this reason, we have developed SignalPlant-a stand-alone application for signal inspection, labeling and processing. The main motivation was to supply investigators with a tool allowing fast and interactive work with large multichannel records produced by EEG, electrocardiograph and similar devices. The rendering latency was compared with EEGLAB and proves significantly faster when displaying an image from a large number of samples (e.g. 163-times faster for 75  ×  10(6) samples). The presented SignalPlant software is available free and does not depend on any other computation software. Furthermore, it can be extended with plugins by third parties ensuring its adaptability to future research tasks and new data formats.

On the Time Course of Synchronization Patterns of Neuronal Discharges in the Human Brain during Cognitive Tasks

Using intracerebral EEG recordings in a large cohort of human subjects, we investigate the time course of neural cross-talk during a simple cognitive task. Our results show that human brain dynamics undergo a characteristic sequence of synchronization patterns across different frequency bands following a visual oddball stimulus. In particular, an initial global reorganization in the delta and theta bands (2–8 Hz) is followed by gamma (20–95 Hz) and then beta band (12–20 Hz) synchrony.

Very High-Frequency Oscillations: Novel Biomarkers of the Epileptogenic Zone

In the present study, we aimed to investigate depth electroencephalographic (EEG) recordings in a large cohort of patients with drug‐resistant epilepsy and to focus on interictal very high‐frequency oscillations (VHFOs) between 500Hz and 2kHz. We hypothesized that interictal VHFOs are more specific biomarkers for epileptogenic zone compared to traditional HFOs.

Taming of the monitors: reducing false alarms in intensive care units.

False alarms in intensive care units represent a serious threat to patients. We propose a method for detection of five live-threatening arrhythmias. It is designed to work with multimodal data containing electrocardiograph and arterial blood pressure or photoplethysmograph signals. The presented method is based on descriptive statistics and Fourier and Hilbert transforms. It was trained using 750 records. The method was validated during the follow-up phase of the CinC/Physionet Challenge 2015 on a hidden dataset with 500 records, achieving a sensitivity of 93% (95%) and a specificity of 87% (88%) for real-time (retrospective) files. The given sensitivity and specificity resulted in score of 81.62 (84.96) for real-time (retrospective) records. The presented method is an improved version of the original algorithm awarded the first and the second prize in CinC/Physionet Challenge 2015.

Multichannel QRS Morphology Clustering - Data Preprocessing for Ultra-High-Frequency ECG Analysis

Ultra-high-frequency ECG (UHF-ECG) in a range of 500–1,000 Hz has been tested as a new information source for analysis of left-ventricle dyssynchrony and other myocardial abnormalities. The power of UHF signals is extremely low, for which reason an averaging technique is used to improve signal-to-noise ratio. Since ventricle dyssynchrony is different for various QRS complex types, the detected QRS complexes must be clustered into morphology groups prior to averaging. Here, we present a fully-automated method for clustering. The first goal of the method is to separate previously detected QRS complexes into different morphology groups. The second goal is to precisely fit the QRS annotation marks to the exact same position against the QRS shape. The method is based on the Pearson correlation and is optimized for parallel processing. In our application with UHF-ECG data the number of detected groups was 3.24 ± 3.41 (mean and standard deviation over 1,030 records). The method can be used in other areas also where the clustering of repetitive signal formations is needed. For validation purposes, the method was tested on the MIT-BIH Arrhythmia and INCART databases from Physionet with results of purity of 98.24 % and 99.50 %.

An additional marker of ventricular dyssynchrony

Patients suffering from heart failure with left bundle branch block (LBBB) can be effectively treated by resynchronization therapy (CRT). The ejection fraction, QRS duration (QRSd) and QRS morphology are the main selection criteria. Unfortunately, approximately one-third of CRT recipients are non-responders. Here we introduce an additional marker capable of distinguishing ventricular dyssynchrony more accurately. Methods: Ultra-high-frequency (UHF, sampling 25 kHz) 12-lead ECG, resting supine position, was measured. We analyzed 21 LBBB patients selected for CRT; the QRSd min/mean/max was 130/163/190 ms. Amplitude envelopes in the 500-1,000 Hz passband were computed and averaged with an R-wave trigger for each patient in the V1 and V6 leads. V1-V6 dyssynchrony (DYS) was computed as the time difference between UHF amplitude maximums in the V1 and V6 QRS complex region. Results: The DYS parameter min/mean/max was 1/68/115 ms. Patients with a small value of the DYS parameter, in spite of the fact that their QRS duration meets CRT criteria (> 120 ms), are not expected CRT responders. The DYS parameter indicates ventricular dyssynchrony and can potentially increase the percentage of CRT responders.

Influence of tilt load on pulse wave velocity in the lower limbs

Pulse wave velocity is a marker of the state of health of the arterial system. We have developed a device (hardware unit and software) for concurrent determination of pulse-wave velocity in several parts of body (up to 18). Using this device we measured the change of pulse-wave velocity in the lower limbs when the subject is exposed to a specific load - “Head-up Tilt Test”.

Ventricular Electrical Delay Measured From Body Surface ECGs Is Associated With Cardiac Resynchronization Therapy Response in Left Bundle Branch Block Patients From the MADIT-CRT Trial (Multicenter Automatic Defibrillator Implantation-Cardiac Resynchronization Therapy)

Background: Although cardiac resynchronization therapy (CRT) is beneficial in heart failure patients with left bundle branch block, 30% of these patients do not respond to the therapy. Identifying these patients before implantation of the device is one of the current challenges in clinical cardiology. Methods: We verified the diagnostic contribution and an optimized computerized approach to measuring ventricular electrical activation delay (VED) from body surface 12-lead ECGs. We applied the method to ECGs acquired before implantation (baseline) in the MADIT-CRT trial (Multicenter Automatic Defibrillator Implantation-Cardiac Resynchronization Therapy). VED values were dichotomized using its quartiles, and we tested the association of VED values with the MADIT-CRT primary end point of heart failure or death. Multivariate Cox proportional models were used to estimate the risk of study end points. In addition, the association between VED values and hemodynamic changes after CRT-D implantation was examined using 1-year follow-up echocardiograms. Results: Our results showed that left bundle branch block patients with baseline VED <31.2 ms had a 35% risk of MADIT-CRT end points, whereas patients with VED ≥31.2 ms had a 14% risk (P<0.001). The hazard ratio for predicting primary end points in patients with low VED was 2.34 (95% confidence interval, 1.53–3.57; P<0.001). Higher VED values were also associated with beneficial hemodynamic changes. These strong VED associations were not found in the right bundle branch block and intraventricular conduction delay cohorts of the MADIT-CRT trial. Conclusions: Left bundle branch block patients with a high baseline VED value benefited most from CRT, whereas left bundle branch block patients with low VED did not show CRT benefits.

CudaFilters: A SignalPlant library for GPU-accelerated FFT and FIR filtering

Signal filtering is one of the essential tasks in signal processing. It may become an extremely time‐consuming process, as in the case of intracranial electroencephalogram recordings (eg, 30‐min records) with a large number of channels (up to 256) and high sampling frequencies (up to 5 kHz in research related to ultra‐high‐frequency oscillations). The usual way of dealing with time consumption is process parallelization. Moreover, parallelization using graphic processing unit (GPU) allows further shortening of computing times thanks to the large number of GPU cores. This paper describes a library for GPU‐accelerated finite impulse response (FIR) and fast Fourier transform (FFT) filtering—“CudaFilters.” This library is designed for SignalPlant software—a free tool for signal analysis. The resultant acceleration in computing times was 5× to 40× depending on the task, data, and hardware configuration. The results were also compared to computing speeds in Matlab.

Intracerebral EEG Artifact Identification Using Convolutional Neural Networks

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.