Qiuzhen Xue

Algorithms for computerized QT analysis

Several methods for measurement of the offset, peak, and morphology of the T wave in multilead ECGs are reviewed and compared. The T wave offset is the most important and also the most difficult measurement for analysis of QT interval dispersion. Measurement methods compared here include (1) the point at which the T wave intersects the isoelectric line plus a threshold; (2) the point at which the derivative of T wave intersects the isoelectric line plus a threshold; (3) the intersection of the maximum slope of the T wave and the isoelectric line; (4) the intersection of a line fitted by least squares to the maximum slope of the T wave and an isoelectric line (LSI); and (5) the point at which the T wave area reaches 90% of the entire T wave area (TA). The reproducibility tests show that the LSI method has the best reproducibility of all the algorithms examined. Although the T wave peak is better defined than the T wave offset, it is not simple to find the right peak when there are multiple T wave peaks and when the T wave is flat and/or noisy. Methods to find T wave patterns with multiple peaks and to locate the point at which the T wave is flat and noisy are therefore reviewed here. Finally, the principal component analysis-based T wave complexity measurement and its relation to other QT interval dispersion measurements are discussed.

Analysis of high-frequency signal-averaged ECG measurements

Analysis of high frequency (150-250 Hz) in the signal-averaged electrocardiogram (SAECG) is one of the emerging methods for detecting vessel patency in acute myocardial infarction following thrombolytic therapy and angioplasty. Root-mean-square voltage (RMSV) of the filtered QRS has been used in earlier studies to detect reperfusion; however, previous analysis indicated that RMSV is sensitive to residual noise in the SAECG and errors in QRS delineation (onset/offset). A new measurement is proposed, high-frequency energy (HFQE), and the robustness of the RMSV and HFQE was evaluated for simulated errors in QRS delineation. In this study, two measures (RMSV and HFQE) were tested on 24 control subjects and 21 patients undergoing thrombolytic therapy. Results indicate that unfiltered QRS duration is more stable than filtered QRS duration for the control subjects and patients and that HFQE had less fluctuation than RMSV in thrombolytic therapy patients. In the control group, HFQE was sensitive to the amplitude variation of the filtered SAECG. Therefore, another new measurement is proposed high-frequency integral of absolute value (HFAV), for reducing the sensitivity to amplitude changes in the filtered SAECG. This new feature was tested on control subjects and was found to be more stable than HFQE. In the thrombolitic therapy group, HFAV provided similar information as HFQE. These three measurements-RMSV, HFQE, and HFAV-provide a comprehensive analysis of the high-frequency SAECG for detecting vessel patency and reocclusion. Relative merits of these measures need to be evaluated on a larger database of patients undergoing thrombolysis and angioplasty for acute myocardial infarction.

An adaptive fuzzy model for ECG interpretation

A new pattern recognition model has been designed for ECG signal classification in general and acute myocardial infarction in specific. This model combines a fuzzy logic inference system with neural network adaptive learning. In this paper, we compare the performance of the proposed system to a neural network only model and a previously designed ECG interpretation program. The initial classification results based on a chest-pain patient database show that the new model has potential for classification accuracy while retaining the knowledge which is particularly useful for clinicians to understand the process of the model.

Review of methods to predict and detect atrial fibrillation in post-cardiac surgery patients

Atrial fibrillation (AF) is the most common sustained arrhythmia after cardiac surgery. Postoperative AF is known to substantially lengthen hospital stay and affect patient recovery. Identification of those at risk of developing AF after surgery and early detection of AF during recovery would be extremely helpful in effective management of these patients, including targeting prophylactic therapy to prevent AF in high-risk patients. In this communication, diagnostic methods to identify those at risk of developing AF after surgery and early identification of AF before, during, and after surgery have been reviewed. Signal-averaged P wave analysis, done before surgery, identifies patients who are likely to develop AF during recovery. When combined with low ejection fraction, signal-averaged P wave can discriminate those who develop AF from those who do not. During recovery, AF can be detected early either from a detailed analysis of atrial activity in a 10-second electrocardiogram or an analysis of R-to-R intervals from an extended rhythm strip (1 minute or longer). Analysis of the 10-second electrocardiogram includes median QRST subtraction from rhythm data and detection and analysis of atrial signals in the resulting residual. AF is detected from extended rhythm strips by using a statistical model to identify the presence of characteristic irregular patterns of R-to-R intervals.

Late potential recognition by artificial neural networks

A neural network (NN) model is trained to neurons, but distributed in many neurons. Therefore, missed classify patients with positive electrophysiology (EPP) test for information from some of the neurons can be compensated by inducible ventricular tachycardia t o m patients with a other neurons. By utilizing this advantage, we can reduce the negative test ( E P N ) f i o m ventricular late potentials of signalburden of detecting offset point very accurately, which is averaged ECG (SAECG). Pattern recognition resultsfiom NN sometimes very difficult. model are compared with the results of Bayesian classification model based on RMS voltage of terminal 40 ms in SAECG. In order to increase the robustness of the recognition, we shifed filtered offset point 4 8 Material: SAECG data was acquired using a High resolution intentionally, which is to simulate the error in detecting offset ECG system. MACIS, (Marquette Electronics. Milwaukee) point offilrered SAECG. I n both shifed and unshifred cases, from 4 hospitals. The Raining S t Consists of 48 SmCGS, in NN model achieved berter sensitivity and specificity than 40 which 24 are and 24 are ~ N S ; test set consists of 49 ms RMS feature by using Bayesian method. SAECGs, in which 16 are EPPs and 33 are EPNs. Orthogonal XYZ signals are amplified and sampled at loo0 samples/sec. After averaging 200-300 cardiac cycles, signals are filtered by a 40-250 Hz bandpass digital filter. A vector magnitude (VM) is then computed from the filtered XYZ signals. Details of II. MATERIAL AND METHODS

Recognition of high-resolution ECGs by time-frequency representation

Three time-frequency (T-F) methods are evaluated for recognition of late potentials. These methods are: short-time Fourier transform (STFT), autoregressive modeling, and Wigner transform. For each method, the procedures include T-F processing, feature selection, and classification. In feature selection and classification, the authors applied both statistical pattern recognition methods like Bayesian and nearest neighbor and also artificial neural network (ANN) models. A combination of both T-F and time-domain features achieved better performance. The authors also analyzed critical issues influencing the performance of T-F methods, namely, low signal-to-noise ratio, accuracy of model estimation, and validity of assumption in using these methods.