CORDIS fournit des liens vers les livrables publics et les publications des projets HORIZON.
Les liens vers les livrables et les publications des projets du 7e PC, ainsi que les liens vers certains types de résultats spécifiques tels que les jeux de données et les logiciels, sont récupérés dynamiquement sur OpenAIRE .
Livrables
The frequent occurrence of artefacts in EEG signals measured under clinical circumstances remains one of the most important factors that impede the widespread use of neuro-monitoring methods in clinical patient monitoring. Artefacts in the EEG caused by disturbances from external (e.g. interference from equipment) or physiological sources (e.g. muscle artefact) can severely deteriorate the results and/or reliability of advanced bio-signal interpretation methods.
To date, no reliable and widely accepted analytical models are available to understand, interpret and validate the whole range of macroscopic EEG signals and the possible artifactual disturbances that can be encountered during patient monitoring. Consequently, empirical knowledge provided by human experts in this field has to be used to evaluate any method for automatic classification of EEG signals and artefact detection in particular.
The current result allows detection of artefacts in 10-second periods of an EEG recording based upon a series of threshold detection rules in individual EEG-derived parameters. The parameters used for this purpose are based upon fitting a 5th-order autoregressive (AR) model to 1-second EEG periods and determination of minimum, maximum and average of the AR-parameter for each period of 10 seconds. Using annotations provided by a human expert in a training subset of the available dataset, discriminate analysis was applied on the EEG-parameters to determine thresholds for the range of parameter values that are to be associated with artefacts. The method allows to control for a pre-set positive predictive accuracy (PPR) in the training set, being a measure for the number of false detections.
In current neuromonitoring equipment, artefact detection mostly consists of simple amplitude-level checks and rejecting periods of the signal with amplitudes that are physiologically not plausible. The problem with this method is two-fold: 1) it does not account for differences in characteristics of the signals measured under various conditions and 2) many types of (severe) artefacts do not manifest themselves as high amplitude signals, but as uncommon signal patterns.
The current result copes with these problems by tailoring artefact detection algorithms to specific monitoring environments and by applying more advanced pattern recognition principles to characterise and identify different types of artifactual waveforms. The basic assumption is that artefacts are associated with outliers in EEG parameter space. The method consists of a training phase in which EEG signals that are considered to be representative for a particular monitoring environment are segmented in fixed-length periods each of which are annotated by a human expert for containing artefacts or being "clean". Next, parameters are extracted from the same EEG periods which are expected to be useful in the discrimination between distorted and reliable periods. The empirical cumulative distribution function for each parameter is determined for both the complete training data as well as for the subset of data annotated is containing artefacts. As last part of the training phase, these two distribution functions are used in a non- parametric discriminate analysis to determine detection thresholds for those parameter ranges that are to be associated with artefacts. The criterion used for setting such a threshold is, that a pre-set target performance criterion in terms of a maximum fraction of false artefact-detections in the training data is met. This iteratively is repeated for subsequent EEG-parameters. The result of this analysis is a set of complementary IF-THEN rules that can be applied to a set of EEG parameters extracted from any EEG recording.
In the evaluation of the method various target performance criteria were evaluated. An illustrative example is that when the maximum number of false detections was set to 5% of all detected artifactual 10-second periods, 74% of all periods in the training set marked by the human annotator as severe artefacts were detected by the automated detection algorithm. Using the same rules on an independent test set of data, the fraction of false detections dropped to 23% of all detections, and 71% of all severe artefacts were detected.
It is expected that the performance of the method significantly can be improved by adding other EEG parameters and by generalising the discriminate analysis to a multivariate technique.
Résultats exploitables
The IMPROVE Data Library consists of digital recordings of continuous monitored signals, clinical patient data and continuous patient state annotations. The continuous signals recorded typically are ECG (2 channels), systemic blood pressure, pulmonary arterial pressure, central venous pressure, airway flow and pressure and concentration of oxygen and carbon dioxide in the airway gases. In addition, EEG (2 channels) was recorded from 7 patients. The continuous signals are stored in the European Data Format. The other clinical data contains practically all the routinely available data in the intensive care, i.e. laboratory results, fluids and medication, nursing actions, patient state assessments and scoring, trend samples from the patient monitors (temperatures, heart rate, blood pressures, etc.) and intermittently measured variables (cardiac output, pulmonary capillary wedge pressure, etc.). In addition, the data is completed by continuous annotations which are made by a physician who stayed at the patientP2s bedside throughout the recordings. The annotation include patient state assessment, observations of possible sources of artifacts for the signals (coughing, patient movements, routinely non-recorded nursing actions, etc.). The clinical data and annotations are stored in ASCII format.
Recherche de données OpenAIRE...
Une erreur s’est produite lors de la recherche de données OpenAIRE
Aucun résultat disponible