M. Strintzis

One-lead ischemia detection using a new backpropagation algorithm and the European ST-T database

A supervised neural network (NN) based algorithm was used to detect ischemic episodes from electrocardiograms (ECGs). The algorithm is tested on the European ST-T database. The algorithm is very fast in its recall state due to the NN, and uses the minimum amount of information, since it is applied on a one-lead ECG. The adaptive training backpropagation algorithm reduces dramatically the training time, and makes possible adjustment training. Even though the algorithm has some problems with detecting the exact onset and end of an ischemic episode, its performance was encouraging since it had a gross sensitivity of 84.4% for ischemia episode detection in the 60 out of 90 records on which it was initially tested. Thus, it seems to be suitable for use in critical care units due to its speed and training capabilities.<<ETX>>

Detection and modeling of infarcted myocardium regions in MRI images using a contour deformable model

This work provides a semi-automatic method for defining and modeling of the infarcted myocardial tissue in MRI images. A deformable contour model based on Fourier decomposition is used to define the border of the infarcted region in successive slice images. A new fast algorithm has been developed for fitting the curve to the borders of the region. The method includes a two-stage process which detects boundary points and directly calculates the model parameters. The method has been tested on MR images from patients with myocardial infarction. Results show that the infarcted region modeling method performs well, being fast, accurate and relatively insensitive to image noise and inhomogeneities.

ST segment nonlinear principal component analysis for ischemia detection

PCA is most effective for distributions which are close to Gaussian. However, typical ST segments are not nearly symmetric. Nonlinear principal component analysis (NLPCA) is a rather new technique for nonlinear feature extraction which is usually implemented by a 5-layer neural network. It has been observed to have better performance, compared to PCA, in complex problems where the relationships between the variables are not linear. The authors apply NLPCA techniques for ST segment feature extraction and they use the NLPCA features to classify each segment into one of 4 classes: normal, ST+, ST-, or artefact. The authors results from the European ST-T database show that using only 2 nonlinear components trained on a set of 1000 normal samples from each file they are often capable of achieving a classification rate of more than 90% with a false alarm rate of less than 10%, while the classification rate rarely falls below 80%. This is an encouraging result which can be further improved with the use of more nonlinear component features or more complex classifiers.

Ischemic classification techniques using an advanced neural network algorithm

The correct classification of the beats relies heavily on the efficiency of the features extracted from the ST-segment and on the desired abilities of algorithm on sensitivity and specificity indices. Nonlinear Principal Component Analysis (NLPCA) is a recently proposed method for nonlinear feature extraction. It has been observed to have better performance for representing complex ST segment features of normal and abnormal cases. The function of representation was created using only normal patterns from the same patient. The distribution of these patterns is modeled using a Radial Basis Function Network (RBFN). This model is capable of finding abnormal patterns with high sensitivity while the specificity is also acceptable (>70%), and the authors can accomplish correct classification rates of higher than 90% for the ischemic beats in many files of the European ST-T database. This technique may be used, in general, for other classification problems in medicine and other disciplines.

Nonstationary ECG analysis using Wigner-Ville transform and wavelets

Nonstationary analysis of ECGs (especially the ST segment) was performed using the Wigner-Ville distribution (WVD) and wavelet transforms. The analysis was done on multiple leads of the same subject and on subjects with normal ECG, ischemia, necrosis and infarct. All data came from the CSE multilead database. It was found that the spectrotemporal maps were not considerably different from lead to lead and that substantial changes in spectrotemporal maps concerning the existence of nonstationarities exist among the above-mentioned pathological states. These changes were evident mainly in the QRS complex and the ST segment. Only in the infarcted subject did such changes persist over the whole P-QRS-T complex. The WVD was found superior from the wavelet transform in having better time- and frequency-domain resolution and superior computational performance.<<ETX>>

Smart alarming scheme for ICU using neural networks

In this work a new scheme for intelligent alarming is presented. The idea is that in order for an alarming scheme to be able to be efficient, the definitions of normal, abnormal and intermediate state have to be changed many times on an hour to hour basis, since in ICU the patient state can change dramatically from day to day. In order to do so, unsupervised and supervised learning systems need to be incorporated that can be trained fast and reliably by the medical personnel. Thus the need for a system that can be trained fast and the existence of a user-friendly MMI where the doctor shall be able to modulate the boundaries between normal, abnormal and intermediate values according to the patients condition is imperative. In this paper, this approach is implemented, using neural networks (NN) for training and learning, and a user friendly MMI using colours and 2-D phase planes of parameters monitored in ICU are used to achieve more efficient alarming schemes.

Parametric 4D modeling of myocardial surface motion applied to MRI data

A deformable 4D model based on Fourier decomposition is presented which was successfully applied to the modeling of the cardiac endocardial and epicardial surfaces and their deformation in time. The proposed method automatically selects boundary points on the myocardial surfaces in the 3D space, collecting a different set of points for each phase. A 4D model is then fitted to the selected points by calculating its parameters directly using an FFT algorithm. The constructed model provides a continuous and smooth representation of the moving surfaces which is consistent with the registration between different phases. Testing on 3D multi-phase MRI data and on 2D multi-slice multi-phase data has provided satisfactory results.

Application of a 3-D ischemic heart model derived from MRI data to the simulation of the electrical activity of the heart

The propagation of the electrical activity of the heart is simulated over regions containing infarcted tissue. The method is applied to patients suffering from ischemia, in order to study the impact of the injury on the electrical function of the heart. The spatial distribution of the infarcted tissue and the shape of the endocardial and the epicardial surfaces are derived from MRI data using a semiautomatic method based on 3D Fourier parametric modeling. A 2D grid is then constructed which represents a selected part of the epicardial surface and contains information about the condition of the tissue. This grid is used as input to the simulation algorithm which estimates the ionic currents over time and the propagation of the electric impulse.

IHIS: an integrated hospital environment linking via LAN ICU with PACS and biochemical laboratories

Intensive care units are characterised by the need for real-time processing of vast amounts of data. These data can result from other laboratories as well such as Biochemical Lab (LIS) and the PACS Lab. In this work an integrated environment has been developed capable of integrating data from all three laboratories into a common workstation, located in the ICU. In this way, the doctor and nursing staff have the capability to browse important data, process them and interact with the other laboratories for better assessment of the patients condition.

Comparison of time and frequency based methods for electrode distance estimation from surviving tissue

In this work, time, frequency and time-frequency methods for distance estimation are presented and compared for the electrode distance estimation one-dimensional case, under a constant velocity assumption. Electrical sources are modelled as multipoles or sources with triangular distribution. The estimation is based on these spatial models of the electrical sources and the corresponding electrical fields. Among the methods used, non-linear distance estimation seems to be a robust method. Specific characteristics of the signals (in time domain or frequency domain) are not sufficient for successful estimation. On the other hand, wavelet analysis may describe signal behavior in more detail and thus help improve ones models.