Correlation analysis of deep learning methods in S‐ICD screening

Abstract Background Machine learning methods are used in the classification of various cardiovascular diseases through ECG data analysis. The concept of varying subcutaneous implantable cardiac defibrillator (S‐ICD) eligibility, owing to the dynamicity of ECG signals, has been introduced before. There are practical limitations to acquiring longer durations of ECG signals for S‐ICD screening. This study explored the potential use of deep learning methods in S‐ICD screening. Methods This was a retrospective study. A deep learning tool was used to provide descriptive analysis of the T:R ratios over 24 h recordings of S‐ICD vectors. Spearman's rank correlation test was used to compare the results statistically to those of a “gold standard” S‐ICD simulator. Results A total of 14 patients (mean age: 63.7 ± 5.2 years, 71.4% male) were recruited and 28 vectors were analyzed. Mean T:R, standard deviation of T:R, and favorable ratio time (FVR)—a new concept introduced in this study—for all vectors combined were 0.21 ± 0.11, 0.08 ± 0.04, and 79 ± 30%, respectively. There were statistically significant strong correlations between the outcomes of our novel tool and the S‐ICD simulator (p < .001). Conclusion Deep learning methods could provide a practical software solution to analyze data acquired for longer durations than current S‐ICD screening practices. This could help select patients better suited for S‐ICD therapy as well as guide vector selection in S‐ICD eligible patients. Further work is needed before this could be translated into clinical practice.


| INTRODUC TI ON
The subcutaneous implantable cardiac defibrillator (S-ICD) is an established totally avascular alternative to the traditional transvenous defibrillators (TV-ICDs). S-ICDs offer defibrillation therapy while avoiding lead-related complications associated with traditional ICDs (Lambiase, 2022). However, the Achilles heel of the S-ICD to date remains the relatively high rate of inappropriate shocks when compared with conventional TV-ICDs. T-wave oversensing (TWO) is still the most common cause of inappropriate shock delivery in S-ICDs (Knops et al., 2020).
Appropriate functioning of the S-ICD relies on the presence of vectors with suitable ECG morphology. As such, not all patients are eligible for S-ICDs, and pre-implant ECG screening is performed on all potential candidates to ensure they have at least one vector that meets the screening criteria (Francia et al., 2018;Groh et al., 2014;Olde Nordkamp et al., 2014;Randles et al., 2014;Rudic et al., 2018).
Surface ECGs of few seconds duration done in multiple postures are used as surrogates for the three standard S-ICD vectors. These are assessed via an automated screening tool built-into an S-ICD programmer to determine eligibility (Rudic et al., 2018). A major predictor of eligibility is the T:R ratio. ECG signals sensed by at least one vector needs to pass screening in at least two postural positions for the patient to be deemed eligible. Unfortunately, despite this screening process, TWO remains the commonest cause of inappropriate shock therapies in S-ICD patients (Kamp & Al-Khatib, 2019). This is significant as inappropriate shock therapies can have detrimental effects on the quality of life, psychological well-being, can result in the induction of ventricular arrhythmias and increase allcause mortality (Daubert et al., 2008).
T-wave morphology is dynamic and can alter with position, exercise, electrolyte disturbance, progression of myocardial diseases, and changes in autonomic function (Al-Zaiti et al., 2011;Assanelli et al., 2013;Hasan et al., 2012;Madias et al., 2001;Mayuga & Fouad-Tarazi, 2007). This can provide an explanation for the occurrence of TWO events in vectors with ECG signals that originally passed the S-ICD screening. It is important to highlight that not all TWO events result in inappropriate shocks. If the TWO episode is not sustained long enough to result in capacitator charging, it will pass unmarked, and no record of the event is made. This is because an S-ICD is only programmed to store episodes of tachycardia that result in capacitor charging. This preserves both battery life and memory capacity of the system. Therefore, the true incidence of TWO in the S-ICD population is not known.
The concept of the potential varying of S-ICD vectors eligibility over time was previously presented in a published study by Wiles et al.  The study demonstrated that the vector score which determines S-ICD eligibility is dynamic in real-life ICD population. For that study, an S-ICD simulator provided by the device manufacturer was utilized for vector assessment. The clinical significance of this dynamicity has not yet been evaluated. However, it sheds light on the possibility that acquiring screening data over a much longer period than for conventional screening can enable more reliable and descriptive screening of the ECG signals sensed by the S-ICD vectors and can aid patient and vector selection in S-ICD candidates. The ultimate goal would be reducing the risk of TWO and inappropriate shocks.
The authors of this article have also previously introduced a novel deep learning-based screening tool, which provides a detailed descriptive analysis of the behavior of T:R ratios from an S-ICD perspective could be obtained (Dunn et al., 2021). We theorize that this tool can guide patient selection as well as vector selection in S-ICD recipients. The aim of this study was to clinically apply this tool to screen a cohort of ICD patients for S-ICD vectors eligibility and assess our findings against a "gold standard"-an S-ICD simulator.

| ME THODS
This is a retrospective correlation study. In our previous study Wiles et al. (2021), adult ICD (transvenous and S-ICD) patients were asked to wear Holter monitors for 24 h to record ECG signals corresponding to their S-ICD vectors. Then, these ECG recordings were analyzed using an S-ICD simulator to assess the vector scores automatically at regular intervals. Mean vector scores were also analyzed and a new concept-Eligible Vector Time (EVT), representing the percentage of all the screening assessments with passing vector scores-was introduced in the study. The study concluded that the S-ICD vectors eligibility in an ICD population is dynamic .
The same 24 h Holter recordings, sampling rate 500/s, from our previous study were first downloaded. Then, the recordings were analyzed by our deep learning tool. First, the data is split into 10 s segments. Then, phase space reconstruction (PSR) was utilized to convert the ECG signal into compressed 32 x 32 pixel PSR images, one image for each 10 s worth of ECG data.
Phase space reconstruction is a popular technique in waveform analysis for representing nonlinear characteristics of time series set of data using delay maps. Typically, when a signal is examined, it is plotted against time. To construct the PSR of the signal, a copy of the signal, which is delayed by a given amount of time, is first created.
Then, the original signal is plotted against the delayed signal. At each time increment, a single point in the phase space reconstruction is created. Each point has an x-axis value equal to the value of the original signal at that time increment and a y-axis value equal to the value of the delayed signal at that time increment. By removing the time axis from this plot, the repetitive behaviors of the signal can be seen (Rocha et al., 2008). An example of how a signal can be transformed into its PSR image is illustrated in Figure 1.
Several techniques have been used before to analyze PSR. Box counting, where simply the image or the phase space is divided up into a grid and the number of boxes occupied by the image is counted. Also, measuring the area covered by the PSR plot and calculating summary statistics for rows and columns of the PSR matrix.
In the algorithm used here, a multilayered Convoluted Neural Network (CNN) is trained to automatically determine the optimal features to extract from the transformed ECG 32 × 32-pixel PSR images. This CNN is made up of a number of feature extraction blocks, followed by a regression block. The outputs from the preceding feature extraction blocks are flattened to a 1D vector and fed into a series of fully connected (dense) layers of neurons to arrive at the final regression output: the T:R ratio.
The end result is a plot showing the mean T:R, standard deviation (SD), and the number of 10 s segments that has T:R above a predetermined threshold for the ECG signals sensed by each lead/S-ICD vector over the recorded period. See Figure 2. For further details on the algorithm development, refer to our previously published work by Dunn et al. (Dunn et al., 2021) The time that a T:R ratio of a vector was deemed favorable (below the eligibility threshold) was calculated as a percentage of the whole recording (=number of 10 s segments with T:R below eligibility threshold/total number of 10 s segments in the recording × 100).
For this article, this was labeled as favorable ratio time or FRT.

| Statistical methodology
Data analysis was done using RStudio 1.4.1106 running R 4.0.5.
Continuous data were presented as mean ± SD. The distribution of the data was checked using normality tests and plots, and histograms. The correlation was checked using Spearman's rank coefficient among variables that fit interval, ordinal or ratio scale, and in paired observations with monotonic relation assumption.
To correlate the outcome of our deep learning tool with that of the S-ICD simulator, the following were compared statistically: (mean vector score + standard deviation of the vector score) and (mean T:R + standard deviation of the T: R, mean T:R + standard deviation of T:R) and EVT, and finally FRT and EVT.

| Patients' demographics
A total of 14 patients (mean age: 63.7 ± 5.2 years, 71.4% male) were recruited in the original study. The primary and alternate vectors for each of the patients, amounting to a total of 28 vectors were analyzed. A total of 13 (92.9%) patients had transvenous ICDs. There was a high prevalence of ischemic heart disease (42.9%) and severe (ejection fraction <35% on echocardiogram) LV dysfunction (28.6%) in the recruited cohort. The main indication for ICD therapy was secondary prevention (71.4%). See Table 1 for detailed patients' demographics.

| T:R assessment
Mean T:R was lower in the primary vectors when compared to the alternate vectors (0.20 ± 0.06 vs. 0.22 ± 0.06, p = .30). Standard deviation of the T:R (a representation of dynamicity) was also lower in the primary vectors (0.07 ± 0.02 vs. 0.09 ± 0.02, p = .11). This has translated to a higher favorable ratio time (FRT) in the primary vectors when compared to the alternate vectors (87.1 ± 14.13 vs. 70.4 ± 16.16%, p = .07).
Mean T:R for all the 28 vectors combined was 0.21 ± 0.11, standard deviation for all the 28 vectors combined was 0.08 ± 0.04, and the FRT for all the vectors combined was 79 ± 30%. For individual assessment of each vector, see Table 2. F I G U R E 1 Illustrates how a phase space reconstruction (PSR) image of the sine wave is created. As the sine wave has a repeated behavior, the sine wave signal could have any length and the phase space points generated from this signal would continuously trace and retrace the resulting circle figure which represents the PSR image of the sine wave.

TA B L E 1 Patient demographics.
F I G U R E 2 One example of the results of vector analysis produced by our tool; This is the T: R analysis of the alternate S-ICD vector of patient 03.   Subsequently, the T:R ratio is not fixed in the same individual. This could provide a potential explanation for the occurrence of TWO in S-ICD vectors that previously passed S-ICD screening.

| Correlation
Because of the crucial role of the T:R ratio in the sensing mechanism of the S-ICD and its subsequent determination of S-ICD eligibility and TWO events, it was chosen specifically as the parameter to be tracked and analyzed by our novel tool. A T:R ratio eligibility cutoff ratio of 1:3 was chosen based on the manual S-ICD screening tool following the manufacturer's guidelines (Randles et al., 2014), although the manual screening method is now highly replaced with automatic screening methods, as they follow the same principles.

| Practical considerations
Our previous work Wiles et al. (2021) has utilized a S-ICD simulatorprovided by the S-ICD manufacturer-in order to analyze the vector score over the 24 h recordings in real-time, that is, it took 24 h for the S-ICD simulator to analyze a 24 h ECG recording for one vector.
As the S-ICD simulator essentially replicates the sensing mechanism of a S-ICD in real-time, it can be considered as a "gold standard" for evaluating various durations of ECG signals sensed by S-ICD vectors. Our study demonstrated that the outcomes of our novel tool correlate strongly with those of the "gold standard" S-ICD simulator, aside from being time-efficient. The simulator-aside from being not readily available-runs in real time and analyses the ECG signals sensed by the S-ICD vectors consecutively, which can be a timeconsuming process, particularly if it is required to analyze recordings of even longer durations. Our tool can provide detailed descriptive analysis of the T:R ratios simultaneously for the ECG signals sensed by all the vectors across the recordings within a few minutes without compromising on accuracy.
Acquiring screening data for eligibility in S-ICD candidates over a longer period than for conventional screening practices seems like a reasonable approach to minimize the effect of the-at least theoretical-dynamicity of S-ICD eligibility. However, this approach increases the burden of data analysis required to assess S-ICD eligibility. Our tool represents a practical software solution that could provide detailed data analysis within minutes; thus, facilitating informed decision-making and could guide patient selection as well as vector selection in S-ICD candidates.
While the R:T ratio rather than the T:R ratio is more common in literature, the reason for choosing T:R ratio for this work can be attributed to the deep learning algorithm used for data analysis; as the T-wave amplitude approaches 0, very small changes in the amplitude can result in extreme changes in the R:T ratio (but not T:R ratio). This massive variation in R:T ratio makes it inappropriate for use as a label in the algorithm, and for this reason, the T:R ratio is used instead (Dunn et al., 2021).
The results of this study denote overall more favorable ratios in the primary vectors on average. However, this might not be true for every individual patient. This highlights the importance of individualizing S-ICD screenings and tailoring the device programming for each patient.
There are some limitations to our study; first, the relatively low number of ECG signals were analyzed in this study. Second, only the ECG signals sensed by the primary and alternate vectors were available for analysis, because the Holter that was used to collect the S-ICD vectors was limited to recording only two simultaneous channels. Also, the S-ICD simulator analyzed the data at 1 min intervals, while our novel tool provided analysis of the T:R ratios at 10 s intervals. In addition, the role of the SMART PASS algorithm that could help differentiate between R and T waves based on other characteristics rather than just their amplitudes, was not considered in this study. We propose that our algorithm is to be used as a supplement and not a replacement of the SMART pass algorithm, that is, ECG signals are processed/filtered first through the SMART pass algorithm, and then, analyzed using our algorithm as an additional step.
TA B L E 2 Results of T: R assessment using the deep learning tool. study, with a larger number of recruited patients, and a more controlled data protocol, is needed to assess the effect of different physiological states, such as exercise, sudden changes in position or mental stress on the ECG recordings, and subsequently, the outcome of our deep learning tool. In addition, our algorithm needs to be tested on ECG signals, acquired from a larger number of patients, or simulated patients, that are processed with the widely utilized SMART PASS algorithm, to assess if our algorithm could improve upon the current method of sensing vector selection.

| CON CLUS ION
T:R ratio-a crucial element in the S-ICD sensing mechanism and a major determinant of S-ICD eligibility-is dynamic in "real-life" P. Roberts involved in conceptualization, data analysis, and review and editing.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available from the corresponding author upon reasonable request

E TH I C S S TATEM ENT
This is a retrospective correlation study. National code on clinical trials has declared that ethics approval is not necessary for real retrospective studies.