If fewer electrodes are used for recording, the spatial resolution of the ECG is affected. This model is learned in the sparse domain after obtaining the sparse representation of standard twelve derivatives and its subset.
Standard twelve-lead ECG
Clinical significance of standard twelve-lead ECG
The twelve standard leads can capture the electrical activities of the heart from different directions. The twelve standard leads can also be categorized into four groups based on specific anatomical regions of the heart.
Morphological features of ECG
Characteristics of ECG signals in different leads
The P wave is normally always inverted in aV R and upright in all other leads in the frontal plane. In the case of transverse plane leads, the P wave is upright in leads V3, V4, V5, and V6 for the normal ECG.
Spatially enhanced ECG systems
The difficulty of placing ten electrodes for recording the standard twelve-lead ECG during ambulatory monitoring can be made easier by using fewer electrodes. The standard twelve-lead ECG system recorded using ten electrodes is the most popular lead system among cardiologists.
Scope of the present work
Since localization of the morphological features increases the correlation between leads, it is expected that learning models for each subband can improve the quality of the signal. The ability of the derived models to retain diagnostic information depends on how accurately it can represent the morphological features.
Organization of the thesis
Lead theory in electrocardiography
The potential of the lead vector is the projection of the heart vector onto the unipolar vectors. As the single dipole location model, the heart vector~hat assumes the center of the.
Lead systems in electrocardiography
Thus, it can be shown that by using a minimum of three independent leads, the cardiac vector can be obtained, which can then be used to generate a 12-lead ECG or any other lead system. The standardization of the lead system introduced the standard 12-lead ECG system, which later became the most common method of recording the ECG [3].
Systems for enhancing the spatial resolution of ECG
- Transformation between lead systems: Linear approach
- Transformation from twelve-lead subset: Linear approach
- Transformation from twelve-lead subset: Non-linear approach
- Patient-specific models and global models
They also derived the precordial leads and compared them with those obtained from the standard twelve-lead ECG. They concluded that the EASI system could be a good alternative to the standard twelve-lead ECG. They used a universal linear transformation to derive the standard twelve-lead ECG from both of these systems.
Database, tools and performance evaluation of the models
Physikalisch-Technische Bundesanstalt (PTB) database
Implementation platform and simulation tools
Performance evaluation methods
- Non-diagnostic distortion measures
- Diagnostic distortion measures
The goodness of fit of the model can be assessed using the R2 statistic which is defined as. As the name suggests, measures of diagnostic distortion are used to evaluate the model's ability to preserve clinically relevant information in key signals during reconstruction. 2.7) where Kj denotes the number of coefficients in the wavelet sub-band and D denotes the number of wavelet decomposition levels.
Motivation for the present work
In addition to the above diagnostic measures, the ability of the models to capture diagnostic information using ECG diagnostics is evaluated. This is done by assessing the quality of useful information preserved in the derived conduction signals.
Plan of the thesis
Also, selecting the best predictive lead can improve the diagnostic quality of the derived leads. Because the features are localized, a high correlation between the leads can be achieved between the subbands of the predictive leads and the response lead [9][114]. The subband coefficients of the response line can be estimated using a linear transformation which can be defined as .
Lead selection using diagnostic similarity score
In equation 3.8, wj is the weight for the jth subband and is defined in equation 3.9. In equation 3.11 and equation 3.12, l(k) represents the original derivative, ˆl(k) represents the derived derivative, and k represents the kth sample. For these derived precordial leads, different diagnostic similarity measures (DSMs) defined in Equation 3.8, Equation 3.11 and Equation 3.12 are calculated.
Proposed DSS based lead selective multi-scale linear regression
This is performed using Equation 3.4 and repeated until all lead subband estimates have been calculated. The prediction leads are then decomposed into approximation subband and detail subbands using the wavelet function. These subbands are then passed to the inverse wavelet function to generate a response lead estimate, i.e.
Diagnosability of the derived ECG model
Results and discussions
Performance evaluation
Results of lead selective multi-scale linear regression
Leads V2, V4 and V5 show a deviation from the original in the case of the T shape shown in the figure. This change in the amplitude of the P wave is also evident from the graphs given in the figure. The change in the sign of the first peak in the QRS complex for lead V2 can be seen in fig.
Repeatability of the proposed model
Variations in W EDD for precordial leads of selected heart diseases are evaluated and shown in the figure. The model is simulated five times for the same patient record and the average time required to run it is calculated. The average time required for preprocessing, model learning, and signal reconstruction using the proposed model is 1.54 seconds.
ECG diagnosability of the proposed model
This evaluation also shows that the obtained slight variations in the clinically significant characteristics of the derived ECG are negligible. The good accuracy is due to the improvement in lead signal quality when using the optimal predictor lead. Thus, the model can generate good quality derived lead signals that can classify healthy and unhealthy data with an accuracy as good as that of the original lead signals.
Comparison with the existing models
Summary
This model uses a four-lead subset of the standard 12-lead ECG as the LR ECG. The second model uses a three-lead subset of the standard 12-lead ECG as the LR-ECG. Since the LR-ECG is a subset of the HR-ECG and the dictionaries are learned jointly, it can be assumed that atsh ≈sl.
Learning the conversion function in the sparse domain
The LR ECG is spatially aligned with the HR ECG by stacking it underneath. The HR and LR ECG, i.e. Eh and El, can be related as El=AEh, where A is a matrix that transforms HR ECG into LR ECG. This assumption can be further exploited to generate the HR-ECG from the LR-ECG.
Fine tuning the model by segmentation
Proposed joint dictionary learning models
Joint dictionary learning for spatially enhanced ECG
Multiple joint dictionary learning for spatially enhanced ECG
In the reconstruction stage, the input LR-ECG is assigned to the previously defined segments using the procedure described in section 4.3. This ensures smoothing of small discontinuities that may occur in the pilot signals, especially in the high amplitude regions. The discontinuity occurs when the dictionary fails to estimate HR samples and can appear as a notch in the signal.
Results and discussion
Performance evaluation
The variation in diagnostic content between the original and the reconstructed leads is assessed using W EDD and W DD measures [18, 104]. For the second model, a comparison of diagnostically significant features between the original and the reconstructed lead signals is also performed. The relative error is also calculated by taking the difference between the accuracy of the original and derived ECG.
Results of joint dictionary learning model
The original and reconstructed ECG signal from the inferior MI record 's0114lrem' are shown in Fig. 4.2(a), it can be observed that the original and derivative leads in the frontal plane maintain good agreement. 4.2(b), which shows the original and derived precordial leads, the P wave and the T wave are slightly overestimated in leads V5 and V6.
Results of multiple joint dictionary learning model
4.4(b), the original and the reconstructed precordial leads of the anterior MI record 's0027lrem' are shown. It can also be observed that distortion present in the original signal is missing in the derived cues, which is an advantage of the dictionary learning approach. The ST shape of the derived cues follows the original cues with a minimum error and is evident in Fig.
ECG diagnosability of multiple joint dictionary learning model
The last five seconds of the thirty second samples are then used to obtain classification accuracy of original HR ECG. 4.7, classification accuracy of the original and the spatially enhanced HR ECG between different MI cases is shown. The ECG diagnosability of the proposed model for various MI cases along with relative error is given in Table 4.3.
Comparison with the existing models
From table 4.4, it is observed that the ANN based model performs better among all the models but it takes time in learning time. It is also evident that the values of correlation coefficient, RM SE and W EDD of the proposed models are comparable with the existing models. Thus from Table 4.4, it is clear that the performance of the proposed models in storing diagnostic information is analogous or better than that of the prevalent models.
Summary
Also, from Table 4.4, it can be observed that correlation coefficient for the second model is 0.98 which is the same as ANN and LSLinR models, and better than the other models. TheRM SE of the proposed model is 54.26mV which is slightly higher than ANN and PCA based models. In the case of the PCA-based model, eight independent cues are used at the time of learning and reconstruction which automatically improves the results.
Proposed approach for exploiting spatio-temporal correlations using RNN
This capability of the RNN can be used to capture intra- and inter-lead correlation in the twelve drivers. Therefore, good cross-lead correlation can be found between spatial line samples of lead signals. 5.1, the input data can be arranged in the form number of mini-batches (M) × number of time steps (N) × size of the input vector (p).
Results and discussions
Performance evaluation
The output of the proposed models is compared with the original signals to evaluate their performance. Selected diagnostically significant features of the original and the reconstructed conduction signals are also compared. ECG diagnostics of the proposed models are also calculated by following the procedure described in section 3.4.
Results of proposed RNN models
Nine features, consisting of shape and amplitude features, are analyzed for all three models. 5.3(c), and therefore the GRU model captures the diagnostic information much better than the simple RNN model for the majority of records. This indicates that the essential diagnostic information is preserved in the majority of records for the LSTM model.
ECG diagnosability of the proposed models
The diagnosticity is above 95% and the relative error is below 4%, except between IPL-INF. From Table 5.5, it is evident that very good classification accuracies can be obtained using the spatially enhanced ECG generated by the GRU model. From Tables 5.3, 5.4 and 5.5, it is clear that the maximum ECG diagnosis is produced by the LSTM model for most cases.
Comparison with the existing models
Summary
Diagnosticity analysis indicates the ability of the second model to retain clinically relevant information. Analysis of the diagnosticity of these models shows that the best model may be different for different classes. The comparison shows that the performances of the simple RNN and GRU models are comparable to the existing models and the LSTM model.
Scope for the future work
Trobec, "Electrocardiographic systems with reduced number of leads - synthesis of the 12-lead ECG," IEEE Rev. Plonsey, “12-lead EKG system,” in Bioelectromagnetism: Principles and Applications of Bioelectric and Biomagnetic Fields, 1st ed. Invalidation of the resting electrocardiogram obtained via exercise electrode sites as a standard 12-lead recording,” Am.
Characteristics of waveforms in frontal plane leads
Characteristics of waveforms in transverse plane leads
Performance evaluation of the Proposed Model for all patient records
Repeatability of the model for patients with more than one record
ECG diagnosability of the proposed model for Myocardial Infarction, Dysrhythmia, Car-
Comparison with the existing models
Performance evaluation of the multiple joint dictionary for all patient records
Performance evaluation of the multiple joint dictionary learning model for all patient
ECG diagnosability (%) and relative error (%) (inside brackets)
Comparison with the existing models
Performance assessment of the proposed models for all patient records
W DD of the proposed models
ECG diagnosability (%) and relative error (%) (inside brackets) for Simple RNN model. 100
ECG diagnosability (%) and relative error (%) (inside brackets) for GRU model
Comparison with the existing models