Abstract
Purpose
Cardiac resynchronization therapy (CRT) has been established as an important therapy for heart failure. Mechanical dyssynchrony has the potential to predict responders to CRT. The aim of this study was to report the development and the validation of machine learning models which integrate ECG, gated SPECT MPI (GMPS), and clinical variables to predict patients’ response to CRT.
Methods
This analysis included 153 patients who met criteria for CRT from a prospective cohort study. The variables were used to model predictive methods for CRT. Patients were classified as “responders” for an increase of LVEF ≥ 5% at follow-up. In a second analysis, patients were classified as “super-responders” for an increase of LVEF ≥ 15%. For ML, variable selection was applied, and Prediction Analysis of Microarrays (PAM) approach was used to model response while Naïve Bayes (NB) was used to model super-response. These ML models were compared to models obtained with guideline variables.
Results
PAM had AUC of 0.80 against 0.72 of partial least squares-discriminant analysis with guideline variables (p = 0.52). The sensitivity (0.86) and specificity (0.75) were better than for guideline alone, sensitivity (0.75) and specificity (0.24). Neural network with guideline variables was better than NB (AUC = 0.93 vs. 0.87) however without statistical significance (p = 0.48). Its sensitivity and specificity (1.0 and 0.75, respectively) were better than guideline alone (0.78 and 0.25, respectively).
Conclusions
Compared to guideline criteria, ML methods trended toward improved CRT response and super-response prediction. GMPS was central in the acquisition of most parameters. Further studies are needed to validate the models.
Keywords: Machine learning, CRT, Heart failure, SPECT
Introduction
Heart failure continues to be an increasingly prevalent disease with risk factors that vary substantially between geographies [1]. Cardiac resynchronization therapy (CRT) has been established as one of the most important therapies in the management of advanced symptomatic patients. For those with left bundle branch block (LBBB) and QRS duration > 150 ms, a strong recommendation for CRT implantation seems to be clear [2, 3]. The reasons for the relevant percentage of patients that will not experience benefit from the procedure remain unclear.
Mechanical dyssynchrony as measured by phase analysis has emerged as a potential procedure to predict responders to CRT [4, 5] and to improve lead positioning [6, 7]. In addition, gated myocardial perfusion SPECT (GMPS) imaging demonstrated value as an “all in one” tool for CRT [8, 9]. Peix et al. have shown that changes in left ventricular phase standard deviation (PSD) before and after CRT were associated with response to the therapy [10]. However, two clinical trials were unable to demonstrate the benefits of early mechanical dyssynchrony to predict CRT response [10, 11]. Ventricular remodeling was also supposed to be a potential marker of response to CRT once its relation with cardiac function was well studied [12-14].
In fact, current research regarding about heart failure and CRT showed that many parameters are associated with response to the therapy and that modeling with traditional statistical analysis methods is very challenging [3, 10, 11, 15]. So, machine learning (ML) methods have been tested in nuclear cardiology based on their capacity to use large amounts of data in order, to recognize patterns in nonlinear systems like human biology, and to build prediction models [16, 17]. ML was also tested for CRT before showing favorable results; however, this was without including mechanical dyssynchrony analysis [18, 19].
Considering the multifaceted and heterogeneous behavior of heart failure in combination with the capacity of ML algorithms to recognize patterns, it is hypothesized that a ML model could provide better prediction of CRT response. The aim of this study was to report the development and the validation of ML models that integrate ECG, GMPS, and clinical variables to predict patients’ response to CRT.
Methods
Patient population
This is a post hoc analysis of a non-randomized, multinational, multicenter prospective cohort study: “Value of intraventricular synchronism assessment by gated-SPECT myocardial perfusion imaging in the management of HF patients submitted to CRT” (IAEA VISION-CRT) funded by the International Atomic Energy Agency (IAEA).
The inclusion criteria were as follows: symptomatic HF patients over 18 years old with NYHA functional class II, III, or ambulatory class IV HF for at least 3 months before enrollment despite optimal medical treatment according to the current guidelines; LVEF ≤ 35% from ischemic or non-ischemic causes measured according to the usual procedure at the participating center for inclusion, whereas LVEFs used for analysis came from the nuclear core laboratory; sinus rhythm with LBBB configuration defined as a wide QRS duration (≥ 120 ms). Exclusion criteria were as follows: arrhythmias that prevented the gated acquisition; major coexisting illness affecting survival less than 1 year; right bundle branch block, pregnancy or breast-feeding, acute coronary syndromes, coronary artery bypass grafting, or percutaneous coronary intervention in the last 3 months before enrollment and within 6 months of CRT implantation. The CRT devices were implanted using standard procedures. The LV lead was implanted in the posterolateral coronary vein, depending on vein availability [10, 20].
All patients provided written informed consent, patient anonymity was maintained during data analysis, and all procedures were done according to the Declaration of Helsinki.
Clinical characteristics and GMPS were assessed both at baseline (before CRT) and at follow-up (6 ± 1 months after CRT). Patients whose LVEF has improved 5 points or more were classified as “responder”; otherwise, they were classified as “non-responder.” The improvement or the deterioration for a patient was defined by the difference between LVEF at follow-up minus LVEF at baseline. A second analysis was done by classifying patients whose LVEF has improved 15 points or more as “super-responder”; otherwise, they were considered “non-super-responders” to CRT.
Clinical data
A total of 12 variables were evaluated at baseline and at follow-up: patient characteristics such as age, gender, and ethnicity; disease history such as presence of CAD, previous myocardial infarction (MI), percutaneous coronary intervention (PCI), coronary artery bypass graft surgery (CABG), NYHA class, hypertension (HTM), and diabetes (DM); smoking history; medical treatment.
SPECT MPI assessment
The GMPS was performed approximately 30 min post rest injection using 740 to 1110 MBq (20–30 mCi) of 99mTc-sestamibi or tetrofosmin. All the images were acquired on gamma cameras with high-resolution low-energy collimators, using 180° orbits, and a standard resting protocol with either 8- or 16-frame ECG-gating according to ASNC guideline [21] and a 100% (50–150%) R-to-R gating window was used. Image reconstruction was performed using the OSEM algorithm with three iterations and ten subsets, and a Butterworth filter with a power of 10 and a cutoff frequency of 0.3 cycles/mm. The resulting short-axis images were sent to Emory Cardiac Toolbox (ECTb4, Atlanta, GA) for automatized assessment of LV function including left ventricular ejection fraction (LVEF), summed rest score (SRS), left ventricular end-systolic volume (ESV), myocardial mass, stroke volume, wall thickening (WT), and scar; phase systolic and diastolic parameters including phase peak (PP), phase standard deviation (PSD), and phase bandwidth (PBW); and shape parameters such as end-systolic eccentricity (ESE), end-diastolic eccentricity (EDE), and shape index (ESSI and EDSI) analysis using the Emory Cardiac Toolbox (ECTb4, Atlanta, GA). Concordance between CRT LV lead position was recorded, and the recommended site was also considered.
Concordance was defined as the agreement between CRT LV lead position recorded and the optimal LV lead position identified by Emory Cardiac Toolbox as the latest contracting viable site [11, 22].
Machine learning
A representation of the machine learning pipeline is presented in Fig. 1. It includes an overview of data splitting (training and testing sets), feature selection, data imputation, synthetic minority oversampling technique (SMOTE), data transformations, model tuning within cross-validation, modeling, and prediction. The sample was split stratified with 80% for training and 20% for testing. In Fig. 2, the flow chart displays the initial patients entering the study and the resulting samples used.
Fig. 1.
Modeling pipeline
Fig. 2.
Study flow chart
Variable selection
Five methods were used to select relevant variables to be used in the models. Information Gain, Recursive Feature Selection, Boruta, and Relief were performed individually. Afterwards, these feature subsets were culled using Pearson correlation (0.80) and near-zero variance filters. A fifth subset was created by simply taking the union of the four aforementioned feature selection subsets, and then again applying Pearson correlation and near-zero variance filters.
Model building
Numerous ML models were implemented and evaluated: Support Vector Machine, Linear Discriminant Analysis, K-Nearest Neighbor, Neural Network, Partial Least-Squares Discriminant Analysis, Naïve Bayes, Prediction Analysis of Microarrays, and Random Forest. In addition, logistic regression model with the same parameters was evaluated. The model with the highest performance metrics and best balance was chosen as explained in “Statistical analysis” section.
We used the Prediction Analysis of Microarrays (PAM) [23] aka “Nearest shrunken centroids (NSC)” to develop the predictive model that ultimately best fit CRT response using the selected variables. PAM creates centroids for each class, i.e., “response” and “non-response” in addition to an overall centroid encompassing all classes. The “threshold” is a tuning parameter that shrinks the feature centroids of each class toward the overall centroid for all classes in doing so reducing the effect of noise and removing variables which are unable to discriminate between classes. New samples are predicted by assigning the label of the nearest centroid using distance metrics, such as Euclidean distance.
For CRT super-response prediction, a Naïve Bayes (NB) model was utilized. It is a probabilistic classifier whose foundation is based on Bayes’ theorem which also assumes a naïve assumption of feature independence. Using previously learned attributes, the model assigns class labels to problem instances using probability scores.
Cross-validation
Resampling was performed using threefold cross-validation with 25 repeats. Within each iteration, SMOTE is applied to the training fold while data transformations are applied separately afterwards to both the training and testing fold. SMOTE is used to reduce model uncertainty and overly optimistic estimates of performance from overfitting, while also combating the issue of imbalanced classes. Data transformations include spatial sign, a technique of projecting features on a multidimensional sphere such that each sample is equi-distant to the center of the distribution in order to reduce the effect of outliers, in addition to centering and scaling to pre-process the data.
Prediction models
In order to compare the proposed models (PAM for response and NB for super-response to CRT), we trained two other models considering solely the variables proposed in the guideline: QRSd, LVEF, LBBB, and NYHA [2]. Using the sample as described in the pipeline (Fig. 1), a partial least squares-discriminant model (PLSDA-Guideline) was used for response prediction and a neural network for super-response (NNET-Guideline). The AUC, accuracy, sensitivity, and specificity from each model and from guideline recommendation criteria were compared.
Statistical analysis
We compared the predictive performance of response and super-response for CRT of the ML models based on selected features with guideline recommendations and with models built over the variables proposed by the guideline [2]. Evaluation metrics used include area under the curve (AUC), accuracy, sensitivity, and specificity. For a direct comparison of the models, an AUC comparison test using Delong method was applied [24]. Statistical analysis was performed using R version 4.0.3 (2020–10-10).
Results
Ten centers from eight countries (Brazil, Chile, Colombia, Cuba, India, Mexico, Pakistan, and Spain) participated in IAEA VISION-CRT. A total of 199 patients were enrolled in the trial; 16 of them died before the follow-up or had greatly decreased ESV (ESV < 25 ml), which was an outlier caused by the low resolution of GMPS when measuring a small heart. For ML modeling, 30 patients without LVEF data were excluded leaving a total of 153 samples. For guideline modeling, another three patients without LBBB, ECG QRS, or NYHA were excluded leaving a total of 150 samples. Figure 2 presents a flow chart describing the sample and losses. The baseline clinical characteristics for the sample with 153 patients are presented in Tables 1 and 2. The testing sample used for validation is presented in Table 3.
Table 1.
Baseline clinical characteristics for responder/non-responder
| Variables | Non-responder (n = 83, 54.3%) |
Responder (n = 70, 45.7%) |
p |
|---|---|---|---|
| Age, years | 59.8 ± 11.1 | 61.5 ± 10.2 | 0.3361 |
| Females | 31 (37.3%) | 32 (45.7%) | 0.3775 |
| Race | |||
| African | 9 (10.8%) | 6 (8.6%) | 0.8431 |
| Asian | 5 (6.0%) | 1 (1.4%) | 0.2979 |
| Caucasian | 11 (13.3%) | 10 (14.3%) | 0.9999 |
| Hispanic | 47 (56.6%) | 34 (48.6%) | 0.4055 |
| Indian | 11 (13.3%) | 19 (27.1%) | 0.0510 |
| Smoking | 14 (16.9%) | 13 (18.6%) | 0.9501 |
| DM | 24 (28.9%) | 14 (20.0%) | 0.2784 |
| HTN | 53 (63.9%) | 38 (54.3%) | 0.3002 |
| MI | 24 (28.9%) | 6 (8.6%) | 0.0031* |
| CAD | 33 (39.8%) | 13 (18.6%) | 0.0076* |
| CABG | 1 (1.2%) | 2 (2.9%) | 0.8814 |
| NYHA | |||
| II | 17 (20.5%) | 27 (38.6%) | 0.0224* |
| III | 58 (69.9%) | 34 (48.6%) | 0.0119* |
| IV | 8 (9.6%) | 9 (12.9%) | 0.7092 |
| ACEI/ARB | 62 (74.7%) | 63 (90.0%) | 0.0258* |
| SPECT | |||
| ESV | 205.1 ± 103.3 | 171.9 ± 90.6 | 0.0357* |
| LVEF | 28.3 ± 10.6 | 27.6 ± 11.9 | 0.6949 |
| Mass | 223.8 ± 54 | 202.1 ± 50.2 | 0.0115* |
| SRS | 23.9 ± 11.8 | 19.0 ± 10.8 | 0.0081* |
| Stroke volume | 70.5 ± 22.7 | 55.7 ± 18.5 | < 0.0001* |
| WT | 13.9 ± 8.5 | 13.2 ± 8.3 | 0.5932 |
| Concordance | 20 (24.1%) | 16 (22.9%) | 1.0000 |
| Scar | 27.9 ± 15.8 | 20.1 ± 11.9 | 0.0007* |
| Diastolic | |||
| PBW | 171.3 ± 75.7 | 151.4 ± 80.2 | 0.1176 |
| PK | 9.0 ± 8.3 | 9.1 ± 7.9 | 0.9238 |
| PP | 226.2 ± 28.6 | 234.2 ± 32.7 | 0.1138 |
| PSD | 52.7 ± 18.9 | 48.0 ± 22.0 | 0.1163 |
| Systolic | |||
| PBW | 158.6 ± 73.8 | 139.6 ± 76.1 | 0.1200 |
| PK | 8.6 ± 7.9 | 8.4 ± 8.0 | 0.8842 |
| PP | 136.6 ± 26.2 | 145.8 ± 27.8 | 0.0370* |
| PSD | 50.3 ± 19.9 | 46.0 ± 21.1 | 0.1990 |
| EDE | 0.50 ± 0.20 | 0.55 ± 0.16 | 0.4253 |
| EDSI | 0.84 ± 0.11 | 0.81 ± 0.09 | 0.0431* |
| EDV | 275.5 ± 111.4 | 227.6 ± 96.0 | 0.0049* |
| ESE | 0.54 ± 0.20 | 0.62 ± 0.13 | 0.0017* |
| ESSI | 0.82 ± 0.12 | 0.77 ± 0.10 | 0.0057* |
| ECG QRSd | 160.1 ± 27.5 | 158.9 ± 29.0 | 0.8072 |
Values are n (%) or mean ± SD. Chi-squared test was used for categorical and t-test for continuous variables
DM diabetes mellitus, HTN hypertension, MI myocardial infarction, CAD coronary artery disease, CABG coronary artery bypass graft, NYHA New York Heart Association Functional Classification, ACEI angiotensin-converting enzyme inhibitor, ARB angiotensin receptor blockers, SPECT single photon emission computed tomography, ESV end-systolic volume, LVEF left ventricular ejection fraction, SRS summed rest score, WT wall thickening, Scar % non-viable LV, PBW phase bandwidth, PP phase peak, PK phase kurtosis, PSD phase standard deviation, EDE end-diastolic eccentricity, EDSI end-diastolic shape index, ESE end-systolic eccentricity, ESSI end-systolic shape index
p < 0.05
Table 2.
Baseline clinical characteristics for super-responder/non-super-responder
| Variables | Non-super- responder (n = 125, 81.7%) |
Super-responder (n = 28, 18.3%) |
p |
|---|---|---|---|
| Age, years | 60.9 ± 10.8 | 59.1 ± 10.4 | 0.3983 |
| Females | 47 (37.6%) | 16 (57.1%) | 0.0916 |
| Race | |||
| African | 14 (11.2%) | 1 (3.6%) | 0.3813 |
| Asian | 5 (4.0%) | 3 (10.7%) | 0.9999 |
| Caucasian | 18 (14.4%) | 3 (10.7%) | 0.8348 |
| Hispanic | 68 (54.4%) | 13 (46.4%) | 0.5793 |
| Indian | 20 (16.0%) | 10 (35.7%) | 0.0347* |
| Smoking | 21 (16.8%) | 6 (21.4%) | 0.7592 |
| DM | 36 (28.8%) | 2 (7.1%) | 0.0311* |
| HTN | 79 (63.2%) | 12 (42.9%) | 0.0769 |
| MI | 28 (22.4%) | 2 (7.1%) | 0.1153 |
| CAD | 43 (34.4%) | 3 (10.7%) | 0.0249* |
| CABG | 3 (2.4%) | 0 (0.0%) | 0.0311* |
| NYHA | |||
| II | 30 (24.0%) | 14 (50.0%) | 0.0119* |
| III | 79 (63.2%) | 13 (46.4%) | 0.1542 |
| IV | 16 (12.8%) | 1 (3.6%) | 0.2837 |
| ACEI/ARB | 101 (80.8%) | 24 (85.7%) | 0.7357 |
| SPECT | |||
| ESV | 199.4 ± 102.9 | 147.4 ± 63.2 | 0.0010* |
| LVEF | 28.1 ± 11.6 | 27.4 ± 9.2 | 0.6961 |
| Mass | 219.6 ± 54.7 | 188.2 ± 40.3 | 0.0010* |
| SRS | 22.4 ± 11.8 | 18.5 ± 10.8 | 0.0986 |
| Stroke volume | 66.8 ± 22.2 | 50.0 ± 13.4 | < 0.0001* |
| WT | 13.6 ± 8.6 | 13.7 ± 7.4 | 0.9289 |
| Concordance | 29 (23.2%) | 7 (25.0%) | 0.9999 |
| Scar | 25.3 ± 15.1 | 20.3 ± 11.6 | 0.0704 |
| Diastolic | |||
| PBW | 168.6 ± 78.3 | 133.6 ± 72.2 | 0.0278* |
| PK | 9.2 ± 8.6 | 8.4 ± 5.1 | 0.5572 |
| PP | 226.3 ± 30.0 | 245.9 ± 28.4 | 0.0021* |
| PSD | 52.2 ± 20.5 | 43.0 ± 18.8 | 0.0256 |
| Systolic | |||
| PBW | 155.3 ± 76.2 | 126.1 ± 66.8 | 0.0481* |
| PK | 8.8 ± 8.6 | 7.1 ± 3.0 | 0.0773 |
| PP | 137.9 ± 26.9 | 153.9 ± 25.6 | 0.0050* |
| PSD | 49.6 ± 21.2 | 42.7 ± 16.2 | 0.0634 |
| EDE | 0.52 ± 0.18 | 0.56 ± 0.18 | 0.2571 |
| EDSI | 0.83 ± 0.10 | 0.81 ± 0.09 | 0.1657 |
| EDV | 266.2 ± 110.2 | 197.4 ± 68.6 | < 0.0001* |
| ESE | 0.57 ± 0.18 | 0.62 ± 0.12 | 0.0381* |
| ESSI | 0.80 ± 0.12 | 0.77 ± 0.08 | 0.1096 |
| ECG QRSd | 159.3 ± 30.3 | 160.8 ± 14.8 | 0.7016 |
Values are n (%) or mean ± SD. Chi-squared test was used for categorical and t-test for continuous variables
DM diabetes mellitus, HTN hypertension, MI myocardial infarction, CAD coronary artery disease, CABG coronary artery bypass graft, NYHA New York Heart Association Functional Classification, ACEI angiotensin-converting enzyme inhibitor, ARB angiotensin receptor blockers, SPECT single photon emission computed tomography, ESV end-systolic volume, LVEF left ventricular ejection fraction, SRS summed rest score, WT wall thickening, Scar % non-viable LV, PBW phase bandwidth, PP phase peak, PK phase kurtosis, PSD phase standard deviation, EDE end-diastolic eccentricity, EDSI end-diastolic shape index, ESE end-systolic eccentricity, ESSI end-systolic shape index
p < 0.05
Table 3.
Testing sample distribution
| Model | Response | Non-response | Super- response |
Non- super- response |
|---|---|---|---|---|
| ML models | 14 | 16 | 5 | 25 |
| Guideline models | 13 | 16 | 5 | 24 |
ML machine learning
The first endpoint of this study was response to CRT, defined as a difference greater than 5 percentage points (i.e., 5% increase) in LVEF from baseline to follow-up. The obtained percentage of responders (45.3%) was in agreement with the original results of Vision CRT clinical trial (47.0%) [10]. The second endpoint was super-response to CRT, defined as a difference greater than 15 percentage points (i.e., 15% increase) in LVEF. The actual super-responders (18.0%) in this study were also similar to the original results (21.5%) [25].
The variable importance obtained for each model is presented in Figs. 3 and 4 for PAM and NB models, respectively.
Fig. 3.
Importance of variables for prediction of CRT response
Fig. 4.
Importance of variables for prediction of CRT super-response
The PAM model presented an AUC of 0.80 against 0.72 from the PLSDA-Guideline model as shown in Fig. 5 (p = 0.52); all other parameters are presented in Table 4. For PAM, the sensitivity (0.86) and specificity (0.75) were better than for guideline alone, sensitivity (0.75) and specificity (0.24).
Fig. 5.
Receiver-operating characteristic curves for response to CRT
Table 4.
Performance of response prediction models
| Model | Accuracy | Sensitivity | Specificity | AUC (95% CI) |
|---|---|---|---|---|
| Guideline | 0.47 | 0.75 | 0.24 | – |
| PAM | 0.80 | 0.86 | 0.75 | 0.80 (0.64–0.97) |
| PLSDA-Guideline | 0.66 | 0.70 | 0.62 | 0.72 (0.52–0.93) |
PAM prediction analysis of microarrays, PLSDA partial least squares-discriminant analysis
When comparing super-response prediction models, the NNet-Guideline presented better performance in relation to NB model (AUC = 0.93 vs. 0.87; p = 0.48) as shown in Fig. 6; all other parameters are presented in Table 5. Considering the sensitivity and specificity, NNet-Guideline (1.0 and 0.75, respectively) was better than guideline alone (0.78 and 0.25, respectively).
Fig. 6.
Receiver-operating characteristic curves for super-response to CRT
Table 5.
Performance of super-response prediction models
| Model | Accuracy | Sensitivity | Specificity | AUC (95% CI) |
|---|---|---|---|---|
| Guideline | 0.35 | 0.78 | 0.25 | – |
| NB | 0.80 | 0.80 | 0.80 | 0.87 (0.73–1.00) |
| NNet guideline | 0.79 | 1.00 | 0.75 | 0.93 (0.84–1.00) |
NB naïve Bayes, NNet neural network
Discussion
In this study, we developed and validated a method for cardiac resynchronization therapy decision support which presented better response prediction performance in all metrics when compared to guideline recommendation. In a direct comparison, the PAM model presented an accuracy of 0.80 vs. 0.47 of the guideline, sensitivity of 0.86 vs. 0.75, and specificity of 0.75 vs. 0.24. In AUC comparison, PAM model (0.80) was superior to partial least squares-discriminant analysis (0.72) using guideline parameters QRS, LBBB, LVEF, and NYHA; however, the difference was not statistically significant (p = 0.52).
This study follows the construction of a ML model built on clinical variables and imaging data from both ECG and GMPS of 153 patients from the VISION clinical trial, a non-randomized, multinational, multicenter prospective cohort study. The clinical trial design included patients from different countries making the sample better represent the heterogeneity observed in practice [1]. The use of GMPS was also important to face the multivariable behavior of heart failure in a single and reproducible procedure [9]. A total of 20 features were extracted from SPECT images: 6 from LV function (LVEF, ESV, mass, stroke volume, and wall thickening), 10 from mechanical dyssynchrony (diastolic and systolic PP, PSD, and PWB), and 4 from shape analysis.
Although early studies in favor of mechanical dyssynchrony and phase analysis have not been confirmed by clinical trials as independent predictors [5, 10, 26], our results showed their role in multivariable analysis. As presented in variable importance results (Figs. 3 and 4), phase peak and phase bandwidth had relevance and were included in the modeling process together with the usual variables.
In addition, the shape parameter EDE showed relevance. Recently, a post hoc study has shown a potential applicability of shape parameters and ESE was also shown to be an independent predictor of CRT response in a multivariate analysis [27].
Another model was proposed for prediction of “super-response” to CRT. There is no clear definition for super-responders [28-31]; however, they can be understood as patients that present rare and remarkable improvements in the LV function after CRT and, consequently, reduced risk of subsequent cardiac events [28, 31, 32]. The option for LVEF alone as the marker of super-response was done considering that it is an objective and reproducible parameter [9] with LV functional information, feasible by echo, magnetic resonance, or GMPS. In our study, a LVEF increase of 15% was the criteria for super-response.
Only 27 of 150 patients (18%) presented super-response in our sample, similarly to other studies based on echo data that ranged from 16 to 25% super-responders [28, 31-33]. Guideline criteria for CRT seem to be not effective for the selection of super-responders; six patients did not have any guideline class, and, as demonstrated in Table 5, the accuracy (0.35), sensitivity (0.78), and specificity (0.25) were low. However, modeling a neural network with the same input variables (NNet-Guideline) resulted in substantial improvement for accuracy (0.79), sensitivity (1.00), and specificity (0.75). The Naïve Bayes model (NB) built from selected variables (AUC = 0.87) had no significant difference (p = 0.48) in comparison to NNet-Guideline (AUC = 0.93). We hypothesized that the differences between inclusion criteria (NYHA class, LVEF, and QRS) and post hoc results based on the same baseline parameters are mainly caused by the variabilities of different diagnostic methods and of different operators (intra- and interobserver variation) [34].
It is well known that many complex problems do require complex solutions, which seems to be the case for the treatment of HF. In the present work, we demonstrated that a multivariable approach can surpass single parameters or conventional criteria selection in the prediction of response and of super-response to CRT. Even with a relatively small and heterogeneous sample set, the ML models were able to combine clinical, functional, shape, and phase data. In addition, Hung et al. (2021) demonstrated that LV dyssynchrony when evaluated in the viable myocardium, excluding scar, could be a better predictor than in the entire myocardium [35]. These promising early results suggest a need for continued research in these area as well as the controversial results about lead placement indication.
This study has several limitations. First of all, the study was a post hoc analysis of a prospective non-randomized trial. In addition, for the use of ML models, large sample sizes are desirable; in this study, the training set had 123 patients while the testing set had 30 patients. Further prospective trials and external validations are expected. Another limitation was that all centers used the same software (ECTb) for cardiac orientation and parameter estimation. Because the approaches of other software providers to quantify LV function and dyssynchrony are different, the results from one cannot be directly translated to others [36, 37]. Gated SPECT data was available to physicians before CRT implantation, but it was not mandatory to follow these results guiding the CRT LV lead implantation. These may reduce the capacity to evaluate the clinical impact of the technique. The information provided by the short follow-up period (6 ± 1 months after CRT) also represents a limitation. Finally, the prognostic value of the models needs further investigation.
Conclusions
Compared to guideline criteria, ML methods trended toward improved CRT response and super-response prediction, combining clinical, functional, shape, and phase data. GMPS had a central role in the acquisition of most parameters. Further studies are needed to validate the models.
Acknowledgements
The authors thank the Vision CRT researchers for sharing the data and collaborating: Amelia Jimenez-Heffernan, Sadaf Butt, Claudio T. Mesquita, Teresa Massardo, Amalia Peix, Alka Kumar, Chetan Patel, Erick Alexanderson, Luz M. Pabon, Ganesan Karthikeyan, Claudia Gutierrez, Ernest Garcia, and Diana Paez.
Funding
This study presents the results derived from the International Atomic Energy Agency (IAEA) multicenter trial: “Value of intraventricular synchronism assessment by gated-SPECT myocardial perfusion imaging in the management of heart failure patients submitted to cardiac resynchronization therapy” (IAEA VISION-CRT), Coordinated Research Protocol E1.30.34, and received funds from IAEA. CTM receives grants from CNPq and FAPERJ. It was supported in part by a grant from The American Heart Association (Project Number: 17AIREA33700016, PI: Weihua Zhou) and by Michigan Technological University Undergraduate Research Internship Program (PI: Kristoffer Larsen).
Abbreviations
- CABG
Coronary artery bypass graft
- CAD
Coronary artery disease
- CRT
Cardiac resynchronization therapy
- ECTb4
Emory Cardiac Toolbox Version 4.0
- ECG
Electrocardiogram
- ESV
Left ventricular end systolic volume
- GMPS
Gated myocardial perfusion SPECT
- HF
Heart failure
- IAEA
International Atomic Energy Agency
- LBBB
Left bundle branch block
- LV
Left ventricle
- LVEF
Left ventricular ejection fraction
- MI
Myocardial infarction
- ML
Machine learning
- NYHA class
New York Heart Association Class
- OSEM
Ordered subset expectation maximization
- PCI
Percutaneous coronary intervention
- PSD
Left ventricular phase histogram standard deviation
- SPECT
Single photon emission tomography
Footnotes
Ethical approval The study was approved by the participant countries’ scientific councils and complies with the Declaration of Helsinki. Written informed consent was obtained from all participants and patient anonymity was maintained during data analysis. In addition, we would like to state that (1) the paper is not under consideration elsewhere, (2) none of the paper’s contents have been previously published, (3) all authors have read and approved the manuscript, and that (4) the paper has been published as a preprint in arXiv and is available at http://arxiv.org/abs/2211.07472.
Conflict of interest The authors declare no competing interests.
References
- 1.Tsao CW, Aday AW, Almarzooq ZI, Alonso A, Beaton AZ, Bittencourt MS, et al. Heart Disease and Stroke Statistics-2022 Update: a report from the American Heart Association. Circulation. 2022;145:e153–639. http://www.ncbi.nlm.nih.gov/pubmed/35078371. [DOI] [PubMed] [Google Scholar]
- 2.Tracy CM, Epstein AE, Darbar D, DiMarco JP, Dunbar SB, Estes NAM, et al. 2012 ACCF/AHA/HRS focused update of the 2008 guidelines for device-based therapy of cardiac rhythm abnormalities: a report of the American College of Cardiology Foundation/American Heart Association Task Force on Practice Guidelines and the Heart Rhythm Society. Circulation. 2012;126:1784–800. http://www.ncbi.nlm.nih.gov/pubmed/22965336. [DOI] [PubMed] [Google Scholar]
- 3.Normand C, Linde C, Singh J, Dickstein K. Indications for cardiac resynchronization therapy: a comparison of the major international guidelines. JACC Heart Fail. Elsevier Inc. 2018;6:308–16. http://www.ncbi.nlm.nih.gov/pubmed/29598935. [DOI] [PubMed] [Google Scholar]
- 4.Bax JJ, Bleeker GB, Marwick TH, Molhoek SG, Boersma E, Steendijk P, et al. Left ventricular dyssynchrony predicts response and prognosis after cardiac resynchronization therapy. J Am Coll Cardiol. 2004;44:1834–40. [DOI] [PubMed] [Google Scholar]
- 5.Delgado V, Van Bommel RJ, Bertini M, Borleffs JCW, Marsan NA, et al. Imaging relative merits of left ventricular dyssynchrony, left ventricular lead position, and myocardial scar to predict long-term survival of ischemic heart failure patients undergoing cardiac resynchronization therapy. 2011. http://circ.ahajournals.org/cgi/content/full/CIRCULATIONAHA.110.945345/DC1. [DOI] [PubMed] [Google Scholar]
- 6.Shetty AK, Duckett SG, Ginks MR, Ma Y, Sohal M, Bostock J, et al. Cardiac magnetic resonance-derived anatomy, scar, and dyssynchrony fused with fluoroscopy to guide LV lead placement in cardiac resynchronization therapy: a comparison with acute haemodynamic measures and echocardiographic reverse remodelling. Eur Hear journal Cardiovasc Imaging. Eur Heart J Cardiovasc Imaging. 2013, 2022;14:692–9. https://pubmed.ncbi.nlm.nih.gov/23175695/. [DOI] [PubMed] [Google Scholar]
- 7.Khan FZ, Virdee MS, Palmer CR, Pugh PJ, O’Halloran D, Elsik M, et al. Targeted left ventricular lead placement to guide cardiac resynchronization therapy: the TARGET study: a randomized, controlled trial. J Am Coll Cardiol. 2012;59(17):1509–18. 10.1016/j.jacc.2011.12.030. [DOI] [PubMed] [Google Scholar]
- 8.Henneman MM, Chen J, Dibbets-Schneider P, Stokkel MP, Bleeker GB, Ypenburg C, et al. Can LV dyssynchrony as assessed with phase analysis on gated myocardial perfusion SPECT predict response to CRT? J Nucl Med. 2007;48:1104–11. http://www.ncbi.nlm.nih.gov/pubmed/17574987. [DOI] [PubMed] [Google Scholar]
- 9.de Amorim FF, Peix A, Giubbini R, Karthikeyan G, Massardo T, Patel C, et al. Reproducibility of global LV function and dyssynchrony parameters derived from phase analysis of gated myocardial perfusion SPECT: a multicenter comparison with core laboratory setting. J Nucl Cardiol. 2022;29:952–61. 10.1007/s12350-020-02397-6. (Springer International Publishing; ). [DOI] [PubMed] [Google Scholar]
- 10.Peix A, Karthikeyan G, Massardo T, Kalaivani M, Patel C, Pabon LM, et al. Value of intraventricular dyssynchrony assessment by gated-SPECT myocardial perfusion imaging in the management of heart failure patients undergoing cardiac resynchronization therapy (VISION-CRT). J Nucl Cardiol. 2021;28(1):55–64. 10.1007/s12350-018-01589-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Boogers MJ, Chen J, Van Bommel RJ, Borleffs CJW, Dibbets-Schneider P, Van Der Hiel B, et al. Optimal left ventricular lead position assessed with phase analysis on gated myocardial perfusion SPECT. Eur J Nucl Med Mol Imaging. 2011;38:230–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Udelson JE, Konstam MA. Ventricular remodeling fundamental to the progression (and regression) of heart failure. J Am Coll Cardiol. 2011, 2022;57:1477–9. https://pubmed.ncbi.nlm.nih.gov/21435517/. [DOI] [PubMed] [Google Scholar]
- 13.Cohn JN, Ferrari R, Sharpe N. Cardiac remodeling-concepts and clinical implications: a consensus paper from an International Forum on Cardiac Remodeling. J Am Coll Cardiol. 2000, 2022;35:569–82. https://pubmed.ncbi.nlm.nih.gov/10716457/. [DOI] [PubMed] [Google Scholar]
- 14.Gimelli A, Liga R, Giorgetti A, Casagranda M, Marzullo P. Stress-induced alteration of left ventricular eccentricity: an additional marker of multivessel CAD. J Nucl Cardiol. 2019, 2022;26:227–32. https://pubmed.ncbi.nlm.nih.gov/28353212/. [DOI] [PubMed] [Google Scholar]
- 15.Zou J, Hua W, Su Y, Xu G, Zhen L, Liang Y, et al. SPECT-guided LV lead placement for incremental CRT efficacy: validated by a prospective, randomized, controlled study. JACC Cardiovasc Imaging. 2019;12:2580–3. http://www.ncbi.nlm.nih.gov/pubmed/31542533. [DOI] [PubMed] [Google Scholar]
- 16.Juarez-Orozco LE, Knol RJJ, Sanchez-Catasus CA, Martinez-Manzanera O, van der Zant FM, Knuuti J. Machine learning in the integration of simple variables for identifying patients with myocardial ischemia. J Nucl Cardiol. 2020;27:147–55. [DOI] [PubMed] [Google Scholar]
- 17.Haro Alonso D, Wernick MN, Yang Y, Germano G, Berman DS, Slomka P. Prediction of cardiac death after adenosine myocardial perfusion SPECT based on machine learning. J Nucl Cardiol. 2019;26:1746–54. http://www.ncbi.nlm.nih.gov/pubmed/29542015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Cikes M, Sanchez-Martinez S, Claggett B, Duchateau N, Piella G, Butakoff C, et al. Machine learning-based phenogrouping in heart failure to identify responders to cardiac resynchronization therapy. Eur J Heart Fail. 2019;21:74–85. http://www.ncbi.nlm.nih.gov/pubmed/30328654. [DOI] [PubMed] [Google Scholar]
- 19.Kalscheur MM, Kipp RT, Tattersall MC, Mei C, Buhr KA, DeMets DL, et al. Machine learning algorithm predicts cardiac resynchronization therapy outcomes: lessons from the COMPANION Trial. Circ Arrhythm Electrophysiol. 2018;11:e005499. 10.1161/CIRCEP.117.005499. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Jimenez-Heffernan A, Butt S, Mesquita CT, Massardo T, Peix A, Kumar A, et al. Technical aspects of gated SPECT MPI assessment of left ventricular dyssynchrony used in the VISION-CRT study. J. Nucl. Cardiol 2021;28:1165–71. 10.1007/s12350-020-02122-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Henzlova MJ, Duvall WL, Einstein AJ, Travin MI, Verberne HJ. ASNC imaging guidelines for SPECT nuclear cardiology procedures: stress, protocols, and tracers. J Nucl Cardiol. 2016;23:606–39. http://www.ncbi.nlm.nih.gov/pubmed/26914678. [DOI] [PubMed] [Google Scholar]
- 22.Zhou W, Tao N, Hou X, Wang Y, Folks RD, Cooke DC, et al. Development and validation of an automatic method to detect the latest contracting viable left ventricular segments to assist guide CRT therapy from gated SPECT myocardial perfusion imaging. J Nucl Cardiol. 2017;1–10. http://www.ncbi.nlm.nih.gov/pubmed/28353213. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Tibshirani R, Hastie T, Narasimhan B, Chu G. Diagnosis of multiple cancer types by shrunken centroids of gene expression. Proc Natl Acad Sci USA. 2002;99(10):6567–72. 10.1073/pnas.082099299. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988;44:837–45. http://www.ncbi.nlm.nih.gov/pubmed/3203132. [PubMed] [Google Scholar]
- 25.Mesquita CT, Peix A, de Amorim FF, Giubbini R, Karthikeyan G, Massardo T, et al. Clinical and gated SPECT MPI parameters associated with super-response to cardiac resynchronization therapy. J Nucl Cardiol. 2022;29:1166–74. 10.1007/s12350-020-02414-8. (Springer International Publishing; ). [DOI] [PubMed] [Google Scholar]
- 26.Henneman MM, Chen J, Ypenburg C, Dibbets P, Bleeker GB, Boersma E, et al. Phase analysis of gated myocardial perfusion single-photon emission computed tomography compared with tissue Doppler imaging for the assessment of left ventricular dyssynchrony. J Am Coll Cardiol. 2007;49:1708–14. http://www.ncbi.nlm.nih.gov/pubmed/17448373. [DOI] [PubMed] [Google Scholar]
- 27.He Z, de Amorim Fernandes F, do Nascimento EA, Garcia EV, Mesquita CT, Zhou W. Incremental value of left ventricular shape parameters measured by gated SPECT MPI in predicting the super-response to CRT. J. Nucl. Cardiol. 2022;29:1537–46. 10.1007/s12350-020-02469-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Killu AM, Grupper A, Friedman PA, Powell BD, Asirvatham SJ, Espinosa RE, et al. Predictors and outcomes of “super-response” to cardiac resynchronization therapy. J Card Fail; 2014;20:379–86. https://pubmed.ncbi.nlm.nih.gov/24632340/. [DOI] [PubMed] [Google Scholar]
- 29.Ypenburg C, van Bommel RJ, Borleffs CJW, Bleeker GB, Boersma E, Schalij MJ, et al. Long-term prognosis after cardiac resynchronization therapy is related to the extent of left ventricular reverse remodeling at midterm follow-up. J Am Coll Cardiol. 2009;53:483–90. http://www.ncbi.nlm.nih.gov/pubmed/19195605. [DOI] [PubMed] [Google Scholar]
- 30.Rickard J, Kumbhani DJ, Popovic Z, Verhaert D, Manne M, Sraow D, et al. Characterization of super-response to cardiac resynchronization therapy. Hear Rhythm. 2010;7:885–9. http://www.ncbi.nlm.nih.gov/pubmed/20382271. [DOI] [PubMed] [Google Scholar]
- 31.Ghani A, Delnoy PPHM, Adiyaman A, Ottervanger JP, Ramdat Misier AR, Smit JJJ, et al. Predictors and long-term outcome of super-responders to cardiac resynchronization therapy. Clin Cardiol. 2017;40:292–9. http://www.ncbi.nlm.nih.gov/pubmed/28294364. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Zecchin M, Proclemer A, Magnani S, Vitali-Serdoz L, Facchin D, Muser D, et al. Long-term outcome of “super-responder” patients to cardiac resynchronization therapy. Europace; 2014;16:363–71. https://pubmed.ncbi.nlm.nih.gov/24189477/. [DOI] [PubMed] [Google Scholar]
- 33.Hsu JC, Solomon SD, Bourgoun M, McNitt S, Goldenberg I, Klein H, et al. Predictors of super-response to cardiac resynchronization therapy and associated improvement in clinical outcome: the MADIT-CRT (multicenter automatic defibrillator implantation trial with cardiac resynchronization therapy) study. J Am Coll Cardiol. 2012;59(25):2366–73. 10.1016/j.jacc.2012.01.065. [DOI] [PubMed] [Google Scholar]
- 34.Chung ES, Leon AR, Tavazzi L, Sun J, Nihoyannopoulos P, Merlino J, et al. Results of the Predictors of Response to CRT (PROSPECT) trial. Circulation. 2008;117:2608–16. http://www.ncbi.nlm.nih.gov/pubmed/18458170. [DOI] [PubMed] [Google Scholar]
- 35.Hung GU, Zou J, He Z, Zhang X, Tsai SC, Wang CY, et al. Left-ventricular dyssynchrony in viable myocardium by myocardial perfusion SPECT is predictive of mechanical response to CRT. Ann Nucl Med. 2021;35:947–54. 10.1007/s12149-021-01632-5. (Springer; Singapore: ). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.de Souza Filho EM, Mesquita CT. Phase analysis of gated myocardial perfusion SPECT and cardiac resynchronization therapy: the good, the bad, and the ugly. J Nucl Cardiol. 2021;28:1136–39. 10.1007/s12350-020-02056-w. [DOI] [PubMed] [Google Scholar]
- 37.Nichols K, Santana CA, Folks R, Krawczynska E, Cooke CD, Faber TL, et al. Comparison between ECTb and QGS for assessment of left ventricular function from gated myocardial perfusion SPECT. J Nucl Cardiol. 9:285–93. http://www.ncbi.nlm.nih.gov/pubmed/12032476. [DOI] [PubMed] [Google Scholar]