Assessment of automatic strategies for combining QRS detections by multiple algorithms in multiple leads.

OBJECTIVE: To develop and evaluate an algorithm for the selection of the best-performing QRS detections from multiple algorithms and ECG leads. APPROACH: The detections produced by several publicly available single-lead QRS detectors are segmented in 20 s consecutive windows; then a statistical model is trained to estimate a quality metric that is used to rank each 20 s segment of detections. The model describes each heartbeat in terms of six features calculated from the RR interval series, and one feature proportional to the number of heartbeats detected in other leads in a neighborhood of the current heartbeat. With the highest ranked segments, we defined several lead selection strategies (LSS) that were evaluated in a set of 1754 ECG recordings from 14 ECG databases. The LSS proposed were compared with simple strategies such as selecting lead II or the first lead available in a recording. The performance was calculated in terms of the average sensitivity, positive predictive value, and F score. MAIN RESULTS: The best-performing LSS, based on wavedet algorithm, achieved an F score of 98.7, with sensitivity S = 99.2 and positive predictive value P = 98.3. The F score for the simpler strategy using the same algorithm was 92.7. The LSS studied in this work have been made available in an open-source toolbox to ease the reproducibility and result comparison. SIGNIFICANCE: The results suggest that the use of LSS is convenient for purposes of selecting the best heartbeat locations among those provided by different detectors in different leads, obtaining better results than any of the algorithms individually.

Automatic SVM classification of sudden cardiac death and pump failure death from autonomic and repolarization ECG markers.

BACKGROUND: Considering the rates of sudden cardiac death (SCD) and pump failure death (PFD) in chronic heart failure (CHF) patients and the cost-effectiveness of their preventing treatments, identification of CHF patients at risk is an important challenge. In this work, we studied the prognostic performance of the combination of an index potentially related to dispersion of repolarization restitution (Δα), an index quantifying T-wave alternans (IAA) and the slope of heart rate turbulence (TS) for classification of SCD and PFD. METHODS: Holter ECG recordings of 597 CHF patients with sinus rhythm enrolled in the MUSIC study were analyzed and Δα, IAA and TS were obtained. A strategy was implemented using support vector machines (SVM) to classify patients in three groups: SCD victims, PFD victims and other patients (the latter including survivors and victims of non-cardiac causes). Cross-validation was used to evaluate the performance of the implemented classifier. RESULTS: Δα and IAA, dichotomized at 0.035 (dimensionless) and 3.73 µV, respectively, were the ECG markers most strongly associated with SCD, while TS, dichotomized at 2.5 ms/RR, was the index most strongly related to PFD. When separating SCD victims from the rest of patients, the individual marker with best performance was Δα≥0.035, which, for a fixed specificity (Sp) of 90%, showed a sensitivity (Se) value of 10%, while the combination of Δα and IAA increased Se to 18%. For separation of PFD victims from the rest of patients, the best individual marker was TS ≤ 2.5 ms/RR, which, for Sp=90%, showed a Se of 26%, this value being lower than Se=34%, produced by the combination of Δα and TS. Furthermore, when performing SVM classification into the three reported groups, the optimal combination of risk markers led to a maximum Sp of 79% (Se=18%) for SCD and Sp of 81% (Se=14%) for PFD. CONCLUSIONS: The results shown in this work suggest that it is possible to efficiently discriminate SCD and PFD in a population of CHF patients using ECG-derived risk markers like Δα, TS and IAA.

Cross-database evaluation of a multilead heartbeat classifier.

In this paper, we studied the improvement in heartbeat classification achieved by including information from multilead ECG recordings in a previously developed and validated classification model. This model includes features from the RR interval series and morphology descriptors for each lead calculated from the wavelet transform. The experiments were carried out in the INCART database, available in Physionet, and the generalization was corroborated in private and public databases. In all databases, the AAMI recommendations for class labeling and results presentation were followed. Different strategies to integrate the additional information available in the 12-leads were studied. The best performing strategy consisted in performing principal component analysis to the wavelet transform of the available ECG leads. The performance indices obtained for normal beats were sensitivity (S) 98%, positive predictive value (P(+)) 93%; for supraventricular beats, (S) 86%, (P(+)) 91%; and for ventricular beats (S) 90%, (P(+)) 90%. The generalization capability of the chosen strategy was confirmed by applying the classifier to other databases with different number of leads with comparable results. In conclusion, the performance of the reference two-lead classifier was improved by taking into account additional information from the 12-leads.

Optimization of ECG classification by means of feature selection.

This study tackles the ECG classification problem by means of a methodology, which is able to enhance classification performance while simultaneously reducing the computational resources, making it specially adequate for its application in the improvement of ambulatory settings. For this purpose, the sequential forward floating search (SFFS) algorithm is applied with a new criterion function index based on linear discriminants. This criterion has been devised specifically to be a quality indicator in ECG arrhythmia classification. Based on this measure, a comprehensive feature set is analyzed with the SFFS algorithm, and the most suitable subset returned is additionally evaluated with a multilayer perceptron (MLP) to assess the robustness of the model. Aiming at obtaining meaningful estimates of the real-world performance and facilitating comparison with similar studies, the present contribution follows the Association for the Advancement of Medical Instrumentation standard EC57:1998 and the same interpatient division scheme used in several previous studies. Results show that by applying the proposed methods, the performance obtained in similar studies under the same constraints can be exceeded, while keeping the requirements suitable for ambulatory monitoring

Heartbeat classification using feature selection driven by database generalization criteria.

In this paper, we studied and validated a simple heartbeat classifier based on ECG feature models selected with the focus on an improved generalization capability. We considered features from the RR series, as well as features computed from the ECG samples and different scales of the wavelet transform, at both available leads. The classification performance and generalization were studied using publicly available databases: the MIT-BIH Arrhythmia, the MIT-BIH Supraventricular Arrhythmia, and the St. Petersburg Institute of Cardiological Technics (INCART) databases. The Association for the Advancement of Medical Instrumentation recommendations for class labeling and results presentation were followed. A floating feature selection algorithm was used to obtain the best performing and generalizing models in the training and validation sets for different search configurations. The best model found comprehends eight features, was trained in a partition of the MIT-BIH Arrhythmia, and was evaluated in a completely disjoint partition of the same database. The results obtained were: global accuracy of 93%; for normal beats, sensitivity (S) 95%, positive predictive value (P(+)) 98%; for supraventricular beats, S 77%, P(+) 39%; and for ventricular beats S 81%, P(+) 87%. In order to test the generalization capability, performance was also evaluated in the INCART, with results comparable to those obtained in the test set. This classifier model has fewer features and performs better than other state-of-the-art methods with results suggesting better generalization capability.

Evaluation of an ECG heartbeat classifier designed by generalization-driven feature selection.

In this work we studied the classification performance of feature models selected with a floating algorithm, focusing in the generalization capability. The features were extracted from the RR interval series, from all ECG leads and different scales of the wavelet transform. The generalization was studied using Physionet databases. In all databases the AAMI recommendations for class labeling and results presentation were followed. A floating feature selection algorithm was used to obtain the best performing and generalizing models in the training and validation sets for different search configurations. The best model found includes 8 features, was trained in a partition of the MIT-BIH Arrhythmia database, and was evaluated in a completely disjoint partition of the same database. The results obtained were: global accuracy of 93%; for normal beats, sensitivity (S) 95%, positive predictive value (P+) 98%; for supraventricular beats, S 77%, P(+) 39%; for ventricular beats S 81%, P(+) 87%. This classifier model has less features and performs better than other state of the art methods with results suggesting better generalization capability.