Ethics statement

The entire study was approved by Ethics Committee of the Hospital General Universitario Gregorio Marañón in Madrid (Spain). Written informed consent was obtained from all participants.

Clinical Information and Database

The clinical information has been previously reported with detail in [4]. In brief, EGM were obtained from 23 patients evaluated in EPS for arrhythmia ablation therapy at the Hospital General Universitario Gregorio Marañón in Madrid (Spain). Patients included in the study had previously implanted a Medtronic™ICD, and a LV mapping procedure was performed for clinical reasons by using a 3D nonfluoroscopic Cardiac Navigation System (CNS). Patients (20 men and 3 women) were 69 ± 10 (mean±standard deviation) years old, and all of them had structural heart disease, namely, 18 post myocardial infraction, 3 idiopathic dilated cardiomyopathy, 1 bi-ventricular dysplasia, and 1 localized LV disynergy. All patients had documented previous VT, and the LV ejection fraction depression degree was estimated as mild (3), moderate (4), and severe (16 patients).

EGM were recorded by an ICD electrode located at the right ventricular apex and a left subclavicular can during fast LV pacing. Pacing protocol at each LV site consisted of trains of at least 10 beats of rapid LV bipolar pacing (cycle length of 500 or 400 milliseconds), in order to avoid both a worsening of the patients’ clinical situation, and inducing a number of VT being different to the clinical VT. Hence, the average length of the recorded signals was from 4 to 5 seconds. All paced beats from each site were visually analyzed to exclude sinus captures, fusion beats, and extrasystoles. EGM were simultaneously stored in the ICD by means of two different lead configurations, namely, far-field (can to coil) and bipolar (tip to ring). EGM and their corresponding octants were available for 23 patients, 18 ± 10 EGM per patient. In 11 out of 23 patients, EGM and spatial coordinates (20 ± 9 EGM per patient) were also available. In addition, the classification of pacing sites according to the peak-to-peak EGM voltage as normal (≥ 1.5 mV) or scar (< 1.5mV) was also provided with 12 out of 23 patients (15 ± 8 EGM per patient, 8 ± 6 EGM per patient in scar areas, and 7 ± 5 EGM per patient in normal areas). Note that this study was performed within clinical procedures of mapping and ablation, and therefore patients had a compromised cardiac health status. Since pacing and therapy had to focus as much as possible on the patient response, the distribution of pacing sites was not equally represented in all octants, as it is presented in Table 1, which shows with detail the number of EGM per octant.

Digital EGM were not available due to potential commercial interests. Therefore, the EGM were collected by telemetry from ICD during EPS and printed at a paper speed of 50 mm/s. Then, an algorithm to extract the EGM signal from printed-paper recordings was needed in order to have the EGM signals available in digital format. An algorithm for EGM digital recovery based on digital image processing techniques was used to recover the ICD-EGM from printout format. This method was carefully evaluated visually and quantitatively by comparing the gold-standard digital signal and the recovered signal with several merit figures, see [8, 9] for details.

The input space for the machine learning system was obtained both from the recordings of the can-coil and tip-ring lead configuration, corresponding to far-field and bipolar EGM, respectively. Firstly, a cubic smoothing spline filter was applied for EGM baseline removal. Then, a representative beat waveform template was obtained at every pacing site by averaging the available beats. Peak deflection of far-field EGM waveform was used to synchronize and to average the available beats. Note that there could be some small temporal variations in the waveform morphology during the pacing protocol due to small shift of the catheter because of the patients breath or the heart movements. However, the EGM was discarded when perceptible shift was detected by the electrophysiologist. Two datasets were built: (D1) the morphology of the averaged beat waveform, i.e. samples of the beat waveform template; and (D2) representative features obtained from the beat waveform template.

The averaged beat waveform of dataset D1 consisted of 150 voltage samples, with a total duration of 254 ms. In order to analyze the effect of the stimulus in the learning system, we defined two subsets for far-field EGM waveform (D1): the beat waveform template with the stimulus corresponding to the pace mapping stimulation, and the beat waveform template without the stimulus. The stimulus artifact was discarded by applying linear interpolation between previous and posterior samples to it. In addition, a new dataset was created by the concatenation of far-field and bipolar EGM waveforms, in order to evaluate whether the simultaneous consideration of both EGM provided complementary information.

Dataset D2 considered several features of far-field and bipolar EGM (see Fig 1) that have been shown significant for helping in the regionalization of the impulse site formation within the LV endocardium [4]:

• Voltages: voltage of peak deflection (v_P), which is the highest voltage above the baseline; voltage of initial deflection (v_I), which is the smallest voltage between the onset of far-field EGM and the peak deflection (note that v_I is usually a negative value); voltage of final deflection (v_F), which is the smallest voltage in an interval of 170 ms after the peak deflection instant.
• Voltage ratios: voltage ratio between final and peak deflections (v_FP); voltage ratio between initial and peak deflections (v_IP).
• Time interval and instants: t_ob, or time interval between the onset of far-field EGM and the bipolar intrinsic deflection; t_P, or time instant of v_P; t_I, or time instant of v_I; t_F, or time instant of v_F. The time origin corresponded to the mid-time between previous and current peak deflections.

According to conclusions in [4], the following feature combinations were used in dataset D2:

• 3V: v_I, v_P, and v_F.
• 3V1T: v_I, v_P, v_F, and t_ob.
• 3V3T: v_I, v_P, v_F, t_I, t_P, and t_F.
• 2VR1T: v_FP, v_IP, and t_ob.

It was convenient to normalize every feature, in order to avoid dominant input attributes due to scaling during the learning procedure [10]. A zero mean and unit standard deviation normalization was used in our experiments.

Proposed System

In this section, the theoretical framework is presented for a new data-driven system aimed to support ablation procedures by using EGM (waveform and features) recorded during EPS. Machine learning techniques are used both for system design and for operation. Emphasis is made in supervised learning and capabilities to extract information from the EGM in terms of the VT anatomical exit site. Additionally, a quality measurement stage is included in order to evaluate the statistical properties of the output provided by different machines. The proposed system is divided into three stages, namely, preprocessing, machine learning, and output quality evaluation, which are next described.

Free Parameter Tuning

In order to design the machine learning classifiers (MLP, RBFNN, PNN, and SVM) and regressor (KRR), a wide enough range of values was explored for each free parameter, as follows:

• Number of hidden neurons in MLP: 10 values, from 1 to 10.
• Regularization parameter and Gaussian width for RBFNN: 10 equally spaced values in [10⁻³, 10²].
• Gaussian width in the PNN units: 10 equally spaced values in [0.1, 1].
• Regularization parameter, Gaussian width, and polynomial degree for SVM: 10 logarithmically equally spaced values in [10, 10³], 20 logarithmically equally spaced values in [10⁻², 10], and ten integer values from 1 to 10, respectively.
• Regularization parameter and Gaussian width in KRR: 50 logarithmically spaced intervals were searched in the range from 10⁻³ to 10, and from $\log (0.1\widehat{\sigma })$ to $\log (100\widehat{\sigma })$, respectively, where $\widehat{\sigma }$ is the mean distance between pairs of instances.

The most appropriate values for the free parameters were chosen according to a cross validation methodology [14] (5—fold cross validation in our experiments). In this methodology, a set of training instances are randomly partitioned into 5 subsets of the same size, so that one subset is reserved for evaluating the model constructed with the four remaining ones. This process is repeated 5 times, setting apart each of the 5 subsets just once. Results obtained with the 5 designs are averaged, and the values providing the best free parameters for each kind of classifier/regressor are selected. In our study, 10 realizations were run to obtain the average results from different cross validation partitions and initial weights in MLP.

In order to measure the performance generalization, a leave one patient out strategy was used to provide with a patient-independent result. This methodology is similar to N—fold cross validation, with N being the number of patients, this is: (1) data corresponding to one patient are set apart; (2) the machine is trained with data from the rest of patients; and (3) the performance is evaluated in the reserved patient. This procedure was repeated once for every patient in the database.

Classification Performance

A first set of experiments was conducted with classification machines, aiming to determine the impact of the presence or absence of stimulation artifact (see Fig 1), the input space, and the classification algorithms. For this purpose, binary and multiclass classifiers were used to classify LV halves and octants, respectively. Table 2 shows the averaged accuracy rates of 10 realizations when using the far-field EGM waveform defined in dataset D1, both with and without stimulus, showing that similar performance was in general obtained. Hence, it can be concluded that no noticeable distorting effect was introduced in the classification analysis due to the pace mapping stimulus.

Table 2 also shows that multiclass and binary classifiers have some discriminant capabilities with both far-field and bipolar EGM waveform input spaces. Apart from apical/basal, far-field EGM waveform has more discriminant capabilities than bipolar EGM waveform, and in general, the best performance with dataset D1 was achieved by using the concatenation of far-field and bipolar EGM waveforms. Discrimination between septal/lateral half got the best performance using the SVM classifier with GK (SVM-GK) and far-field EGM waveform as input space, whereas poor performance was obtained for apical/basal and superior/inferior halves in binary classification. Note at this point that the averaged accuracy rate for random choice among 8 classes (octants) is 12.5%, hence octant classification was significantly improved with the use of machine learning techniques, with 25.5% of accuracy for multiclass SVM-GK classifier and combining far-field and bipolar EGM waveforms.

Table 3 shows the averaged accuracy rates of 10 realizations by using dataset D2 (input spaces given by time and voltage EGM features). In this case, the SVM classifier with PK (SVM-PK) was also considered due to the good performance of this kind of kernel when dealing with low-dimensional input spaces. Octant classification with binary classifiers obtained the highest performance, 34.9% and 31.8%, for SVM-GK and input spaces 2VR1T and 3V1T, respectively. With respect to halves classification, apical/basal accuracy increased significantly with respect to the use of EGM waveform (71.3% with 2VR1T and SVM-GK), while similar discriminant capabilities were obtained in septal/lateral (with 3V1T and 3V input spaces, SVM-GK) and in superior/inferior (with 2VR1T, SVM-GK). Note in bold the best results for datasets D1 and D2 in Table 2 and Table 3 for each input space. In spite of the imbalance in the number of pacing sites among octants, the best performance was checked not to correspond to the octant conveying a larger number of EGM.

In light of the outcome, in general terms it follows that features 2VR1T and 3V1T yield the best performance in octant and halves classification. Interestingly, in spite of more electrical information being present in the waveform, no better performance was obtained when using it, probably because of the effect of the high dimensionality of the input space. Besides, SVM-GK usually works slightly better than neural networks classifiers.

Previous results were obtained by accounting the pacing sites placed on the scar or on normal tissue. Table 4 shows the averaged accuracy rates of 10 realizations by using SVM-GK (the best classifier previously obtained) and considering separately the type of tissue when pacing, namely, scar (bipolar peak-peak voltage was < 1.5 mV), normal (bipolar peak-peak voltage was ≥ 1.5 mV), and both. The highest accuracies for each type of tissue are in bold. In general, the best performance was obtained when considering both normal and scar tissue. However, these results could be due to the machine learning nature of our system, which is able to obtain better generalization ability with a more representative design set (in general, with larger number of examples). In addition, it can be checked that septal/lateral classifier provided higher accuracy rate in normal than in scar tissue, whereas the opposite is true for superior/inferior classification. Comparing the results in Table 4 with those in Table 2 and Table 3, we may conclude that input spaces 3V1T and 2RV1T provided the best performance when considering scar+normal tissues. They also yielded the best accuracy when considering normal and scar+normal tissues, both for halves and octants classification. Better performance was obtained with 3V3T for scar tissue in septal/lateral and superior/inferior halves for multiclass classification.

Regression Performance

The usefulness of machine learning for regression was benchmarked in terms of octant regression performance. The spatial coordinates of LVT pacing sites were estimated by using KRR and datasets D1 and D2. The spatial resolution was measured as the Euclidean distance between the actual and the estimated pacing spatial coordinates.

Table 5 shows the Euclidean distance (mean and standard deviation) between the actual and the estimated spatial coordinates for 138 pacing sites in datasets D1 and D2 when pacing in scar, normal, and both tissues. Slightly larger distances were obtained for normal tissue in comparison with scar tissue. A Wilcoxon rank sum test, with significance level at α = 0.05 [22], was performed for statistical comparison of: (1) type of tissue, obtaining significant differences for normal + scar (both) vs normal, and normal vs scar pacing sites; and (2) input space, with significant differences for 2VR1T vs the rest of input spaces.

Distances in Table 5 are in agreement with the spatial resolution reported by other clinical works in the literature. Recall that those works had reported a mean spatial resolution of 8.9±9 cm² (range from 0 to 35 cm²) [5] and a distance ≥ 2 cm to distinguish EGM from two pacing sites [4].