Contemporary Applications of Machine Learning for Device Therapy in Heart Failure
State-of-the-Art Review
Central Illustration
Abstract
Despite a better understanding of the underlying pathogenesis of heart failure (HF), pharmacotherapy, surgical, and percutaneous interventions do not prevent disease progression in all patients, and a significant proportion of patients end up requiring advanced therapies. Machine learning (ML) is gaining wider acceptance in cardiovascular medicine because of its ability to incorporate large, complex, and multidimensional data and to potentially facilitate the creation of predictive models not constrained by many of the limitations of traditional statistical approaches. With the coexistence of “big data” and novel advanced analytic techniques using ML, there is ever-increasing research into applying ML in the context of HF with the goal of improving patient outcomes. Through this review, the authors describe the basics of ML and summarize the existing published reports regarding contemporary applications of ML in device therapy for HF while highlighting the limitations to widespread implementation and its future promises.
Highlights
| • | While actuating exemplary changes, clinical gaps exist in the domain of device therapy in heart failure. | ||||
| • | Machine learning has been postulated to improve candidate selection and outcomes for device therapy. | ||||
| • | Integration of machine learning with current clinical practices can further the idea of precision medicine. | ||||
| • | Data integration and model interpretability are the major challenges preventing widespread assimilation into clinical practice. | ||||
Introduction
Heart failure (HF) is a global pandemic that affects more than 6 million individuals in the United States alone.1 Additionally, it contributes a significant financial burden, costing the United States health care system about $43 billion in 2020.1,2 Given its enormous impact on health care, an ever-increasing proportion of resources have been allocated to understanding the complex and multifaceted pathophysiology of HF and to designing accurate multimodal tools and therapies to improve patient outcomes.
Identification of risk factors and culprit causes via traditional statistical approaches has markedly improved our understanding of the underlying relative and absolute magnitudes of clinical risk and also the causes of adverse clinical events. However, such models place constraints on data utilization, with only a limited number of well-structured variables available for incorporation. Furthermore, a sizeable nonresponder rate and the limited predictive power of prognostic scores point toward the complex interrelationships among incorporated variables while emphasizing the need for continued improvement.3
By contrast, “big data” approaches can incorporate a vast amount of raw and unstructured data, conferring an advantage over traditional methods. Compared with statistical approaches that focus on inferring formal relationships between variables and outcomes (in the form of mathematical equations), machine learning (ML) aims to uncover the complex role and interrelationships between multiple variables while overcoming the constraints of mathematical equations, often in an agnostic manner, by examining trends and associations between variables and a specific outcome.4 For instance, to devise a risk score for predicting onset of a disease, a statistical computation will involve a careful curation of data, deducing a mathematical association between variables and an outcome (eg, odds ratios, relative risks), and using simple regression techniques to create a model. ML models use a different approach. By using a large amount of data, the model will aim to identify linear and nonlinear associations both within variables and between variables and an outcome. By limiting assumptions a priori, the focus of the ML model is to generate predictions rather than infer mathematical relationships between variables and outcomes.
Given the opportunity for advanced and novel analytic techniques to fill existing gaps, ML has received significant interest in recent times, with an exponential increase in research related to the application of artificial intelligence (AI) and ML for cardiovascular applications.5 This review aims to summarize the most recent applications of ML with respect to device therapy within the realm of HF (Central Illustration).
Applications of Machine Learning in Device Therapy for Heart Failure
A Primer on AI and ML for Clinicians
AI refers to computer systems that can perform tasks that would have otherwise required human input. ML is a subset of AI by which computer systems can “learn” from troves of data. Whereas “data” pertains to numeric, string, or visual variables, learning happens when these data along with a known output (ie, training data set) are fed into the ML algorithm. The ML algorithm can infer useful patterns from the implicit data (ie, learning) to generate programs that can predict outcomes on unseen data. This approach is in contrast to traditional programming, wherein a predefined program made by humans is fed into computers, which process the data provided to generate an outcome (therefore, no learning is involved). In cardiovascular medicine, ML has been applied in 2 broad ways: unsupervised and supervised ML (Figure 1).
Brief Overview of Machine Learning
Both supervised and unsupervised machine learning require structured data, the key difference being the need for data labeling. RF = random forest; SVM = support vector machines; PCA = principal component analysis. Created with Biorender.com.
Unsupervised ML is a subset of ML whereby raw, unlabeled data are fed into an algorithm, intending to learn new associations and patterns within the provided data.6 Dimensionality reduction is a method that can allow for the conversion of multidimensional data (eg, ECG, blood pressure, echocardiography) into lower-dimensional space, thus allowing for easier appreciation of the components accountable for variability in the data while allowing for the extraction of relevant features (ie, feature selection). The extracted variables can be fed into clustering algorithms, allowing subgroup identification with similar outcomes. This approach becomes especially useful in HF, where k-means clustering has been used to understand the complex interactions of variables to predict outcomes and prognosis.7-10 A detailed description of the different types of ML algorithms is provided in Table 1 and Supplemental Figures 1 to 5.
| Machine Learning | Description | Example |
|---|---|---|
| Unsupervised machine learning | ||
| Principal component analysis | Transforms raw data dispersed in multiple dimensions into fewer dimensions, helping identify “components” accounting for maximum variance among data. | 200 QRS complexes each from 12-lead ECG combined with QRS duration to form a total of 2,401 dimensions, out of which 2 principal components selected accounting for maximum variance.9 |
| Hierarchical clustering | Clusters groups based on feature similarity by either merging (agglomerative) or dividing (divisive) the clusters at each step. | After LVAD implantation, patients clustered into 5 groups based on adverse event sequences in a divisive approach. These clusters had different dominant adverse events, and patients differed in their clinical profiles.56 |
| k-means clustering | 'k' number of clusters randomly predefined, and data points assigned to nearest cluster centroid (ie, center of the cluster). Centroid of the newly formed cluster is then recalculated, and the process is repeated until the individual cluster centroids don’t change, hence decomposing data into similar groups. | 4 predefined clusters were randomly defined, and patient characteristics were grouped to their nearest cluster, allowing for identification of phenogroups to predict CRT outcomes.8 |
| Supervised machine learning | ||
| Lasso regression | Useful with small number of outcome data (because conventional linear regression risks overfitting data). It aims to reduce overfitting by penalizing a higher slope, thereby introducing some bias but minimizing variance. Can be used for feature selection because the variables contributing minimally to slope are excluded. | Among 2,803 patients, of whom only 44 had infection, variables predicting infections after ICD implantation were evaluated. Of 81 variables, 17 variables selected via Lasso regression based on how much the variables contributed to the slope.48 |
| Decision trees | Makes a prediction about an outcome by arranging variables in a tree/flowchart-based fashion. It starts with an input variable (ie, root), and goes down the hierarchy (ie, nodes) via a true/false approach, finally predicting an outcome (ie, leaf). Very simple to use but is prone to overfitting. | To predict RV failure after LVAD implantation, 8 features organized, with transpulmonary gradient as the root, going down the hierarchy via 7 nodes, to predict RV failure (leaf).62 |
| Random forest | Uses multiple decision trees, whereby a part of training data is fed into each tree in parallel, and the results from the trees are aggregated to give a combined output. Less prone to because since it combines multiple decision trees. | 45 features selected and fed through decision trees in parallel, which were then combined (ie, ensembled) to produce a random forest classifier to predict CRT outcomes.40 |
| Naïve Bayes | It assumes that the effect of 1 variable on the output is independent of other variables. Using probability estimates for individual variables and combining it with the prior probability of fed outcome, classifies a given set of variables to a particular outcome. | When fed with testing data set, a trained naïve Bayes algorithm computes the probability of finding a significant clinical variable (eg, presence of atrial fibrillation) given a patient’s CRT response, and combines the individual probabilities of all variables with the prior probability of CRT response (taught when the model was trained) to classify outcomes in the test data set.32 |
| Ensemble boosting (eg, AdaBoost, XgBoost) | Multiple weak models (decision trees) are aggregated in series to build a stronger prediction model. Higher weightage is given to the data misclassified by the first model to correctly classify that data via next model, hence yielding an overall better performance. | Preoperative variables fed through 1,600 decision trees in series, whereby the depth of the decision trees was decided, and the final XgBoost model to predict mortality after LVAD implantation.55 |
| Neural networks | Consists of nodes (ie, neurons) arranged in the form of interconnected layers. The nodes are trained to be activated when exposed to specific input variables (eg, pixels in images), which then activate nodes in successive layers to generate an output. “Deep” refers to using multiple layers of neurons. | 14,035 echocardiographic videos and images used to train a neural network composed of 16 layers, which can be used for automatic cardiac segmentation and disease detection.21 |
By contrast, supervised ML uses labeled data (eg, presence or absence of HF) to evaluate the association of features with predetermined outcomes (labels).6 A wide variety of labeled structured data can be fed to train the ML algorithm. For example, by using a host of features (eg, age, smoking, diabetes) and a known outcome (HF), models can be trained and used to predict whether individuals will experience HF over time. Supervised ML algorithms use the best fit of the variables to predict both categorical (classification) and continuous (regression) outcomes by ranking variables according to their importance/weightage. Supervised ML has been widely used for building risk prediction and prognostic models, which have performed reasonably well when compared with traditional logistic regression models.8,9 Semisupervised learning falls between the paradigms of unsupervised and supervised ML, whereby a small amount of labeled data is used along with a vast amount of unlabeled data to construct ML models.6 The advantage of a semisupervised approach is the ability to circumvent the significant cost and labor associated with creating fully labeled training sets, whereas the use of unlabeled data can be relatively inexpensive and can result in considerable improvement in model accuracy.
Deep learning (DL) is a subset of ML that can be used in both a supervised and an unsupervised fashion, allowing for models to perform segmentation or classification tasks. It uses multiple layers of nodes (ie, neurons) to extract higher-level features from the input data, which is usually in the form of medical images (Figure 2). The images can be considered a collection of pixels denoted via mathematical numbers. When fed into a DL model, the pixels stimulate a specific set of neurons in the input layer, transmitting signals to neurons in the successive layers (neural network). Whereas the input layer recognizes simpler features (eg, pixels), successive layers recognize complex features as the signal is passed in a hierarchy (eg, ventricular edges on echocardiography, left ventricle [LV] dilation). Thus, the power of DL lies in the fact that it can transform its input data into a slightly more abstract and composite representation. It has been applied vastly in such applications as cardiovascular imaging for automatic computation of ventricular volumes, image noise reduction, and automated atherosclerotic plaque analysis.11
Overview of Deep Learning
Deep learning can use both structured and unstructured data, and process it via multiple layers of neurons, which can be used in a supervised or unsupervised fashion to yield various results. Created with Biorender.com.
Medical Sensors and AI to Reduce Hospitalizations: Current Research and Future Applications
Implantable sensor devices, wearables, and digital health technology allow for continuous measurement of physiological data, thus engaging patients in the decision-making process and tailoring health care interventions to the individual patient. Wearables are increasingly being incorporated into HF clinical trials generating massive amounts of data, and the interconnectivity of these devices has created opportunities for pooling data from multiple sensors.12,13
Although novel sensor metrics have performed reasonably well in the reported studies, whether using ML approaches to correlate these parameters with the traditional gold standard remains an area that requires further exploration. The marriage of these novel sensor metrics and ML is a critical concept, and a few studies have tried to put it together (Supplemental Table 1).7,14,15 In a feasibility study of 20 patients, Shandhi et al14 used globalized regression models to track changes in pulmonary capillary wedge pressure (PCWP) and pulmonary artery pressure (PAP) via seismocardiographic signals emitted from a wearable patch. After vasodilator administration, acute changes in PCWP and PAP estimated during right heart catheterization correlated well with the seismocardiographic signals (root mean square error 2.7 mm Hg and 2.9 mm Hg for measuring PAP and PCWP, respectively). Whether such an approach can allow for noninvasive longitudinal monitoring of patients with HF is a question that needs to be validated in future trials. Nevertheless, it provides an initial breakthrough in using ML as a chaperone, allowing for a comparison of novel sensor metrics with traditional gold standards, thus paving the way for more accurate noninvasive hemodynamic assessment in HF.
A major concern with noninvasive sensors is the generation of large amounts of data, which, coupled with a scarcity of medical workforce personnel, limits the ability to harness these resources to their maximal potential. Automated processing and integration of remote monitoring data with pre-existing clinical variables have the potential to improve the predictive accuracy over the current models. For instance, data integration has been used to develop the HeartLogic index (a multisensor-based algorithm), which, by combining inputs from multiple sensors, can generate an alert signal predicting HF decompensation. Using sensor data on first and third heart sounds, thoracic impedance, respiratory rate, the ratio of respiration rate to tidal volume, heart rate, and patient activity, the HeartLogic index had a sensitivity of 70% (95% CI: 55.4%-82.1%) with a false positive alert rate of 1.47 (95% CI: 1.32-1.65) per patient-year.16
Whereas the HeartLogic index demonstrated good results via integration of a few remote monitoring variables, ML-based data modeling offers unprecedented potential. With the emergence of implantable cardiac hemodynamic sensors, and studies showing their efficacy in improving outcomes, it becomes all the more imperative to devise strategies to integrate these colossal amounts of data.17 Supervised ML algorithms combining a vast amount of remote monitoring data collected via medical sensors and implantables can be envisioned in the imminent future, thereby helping create a window during which an intervention might help abort an incipient clinical deterioration in high-risk patients.
However, certain challenges need to be addressed before the successful integration of ML and sensor devices into contemporary HF guidelines. Commercial and health care devices generate an enormous amount of complex data, and incorporating data into electronic medical records requires sophisticated storage options. Moreover, there is a need for novel advanced analytical tools to process these troves of data in an efficient manner. AI can manage this information overload, thus assisting in optimizing workflows. Once integrated, timely alerts about an impending decompensation (eg, increased PAP, decreased heart rate variability) can help physicians institute directed therapies to prevent hospitalizations and downstream adverse events (Figure 3). Nevertheless, a lingering issue will be the topic of data privacy and safety, because any remote monitoring system might be vulnerable to a data breach or hacking attempts.
Integrating Data From Digital Devices With ML
Integrating data from wearable devices and implantables with traditional variables and electronic health records can lead to a real-time prediction of adverse events, thereby allowing for early intervention and improved outcomes. Created with Biorender.com. ECG = electrocardiogram; ICD = implantable cardiac defibrillator; LVAD = left ventricular assist device; ML = machine learning.
Application of ML in Transcatheter Edge-to-Edge Repair
Transcatheter edge-to-edge repair (TEER) is a safe, well-tolerated procedure and is recommended for certain patients with severe symptomatic secondary mitral regurgitation (MR) who have progressive clinical deterioration despite the institution of goal-directed medical therapies.18 Although qualitative and semiqualitative echocardiographic parameters such as leaflet morphology, regurgitant jet severity, and left atrioventricular volumes are critical in evaluating the severity of MR, these parameters have shown poor reproducibility across multiple studies.18-20
Can ML help optimize perioperative TEER outcomes? A few approaches using ML with the potential to streamline TEER workflow in this regard deserve mention. By training a neural network on 14,035 echocardiograms, Zhang et al21 achieved automatic cardiac structure segmentation and fully automated left atrioventricular volume and global longitudinal strain estimation with an accuracy of 84%. Jin et al22 devised a novel ML model, Anatomical Intelligence in Ultrasound (AIUS), for automatic computation of a range of quantitative mitral valve parameters. Once fed with 3-D TEE images, the model identified mitral annular reference points, aorta, and the coaptation nadir on the end-systolic TEE frame selected automatically via ECG-guided correlation. This allowed for the creation of a topographic map, which was then color-coded to indicate the degree of leaflet displacement and prolapse if any (Figure 4). While exhibiting an excellent correlation with expert manual readings (r = 0.85-0.99 for all parameters; P < 0.05), the AIUS model significantly improved intraobserver and interobserver reproducibility. Also, it allowed for a fivefold reduction in computational time compared with manual quantification (770 seconds vs 144 seconds for manual quantification vs AIUS, respectively). Moreover, when applied among operators with different levels of experience, the model augmented the performance of nonexperts for localization of mitral valve prolapse (accuracy of 83% vs 89%, for manual vs AIUS-guided interpretation; P = 0.003).23 Although it didn’t improve the performance of expert readers, such models can be used as an adjunct to improve patient outcomes. Similarly, DL models created to identify mitral annular dimensions have been described, the performance of which is yet to be validated in large external cohorts (Supplemental Table 2).24
Anatomical Intelligence in Ultrasound Model
Once the end-systolic frame is identified (A), mitral annular, aortic, and coaptation nadir are marked (B); the coaptation line is drawn manually (C, D). The resulting anatomical intelligence ultrasound—guided topographic representation (F) correlates well with 3-dimensional transesophageal echocardiogram–illustrated (E) mitral valve prolapse (red arrows in E and F). Adapted with permission from Jin et al.23 Ao = aortic annulus on the anterior side of the valve; AL = anterolateral; P = posterior; PM = posteromedial.
On the patient level, Hernandez-Suarez et al25 developed an ML model to predict in-hospital mortality in patients undergoing TEER. Using a data set of 849 patients who underwent TEER from 2012 to 2015, a Naïve Bayes algorithm was able to better predict in-hospital mortality than the logistic regression (LR) model (AUC: 0.83 vs 0.77; P < 0.95 for naive Bayes and LR, respectively). Variable ranking identified coronary artery disease, chronic kidney disease, and smoking as the most important parameters in predicting in-hospital mortality.25 Although the score performed well internally, the lack of external validation currently restricts its incorporation into daily clinical practice. Moreover, the study was conducted on a cohort of patients who underwent TEER for primary MR. Whether these results can be extrapolated to patients undergoing TEER for secondary MR is a question that merits further research.
More recently, ML has been used to develop the MITRALITY (Transcatheter Mitral Valve Repair Mortality Prediction System) score to predict 1-year mortality after TEER. Comprising 6 variables, the score better predicted mortality than the existing prediction scores (AUC: 0.783, 0.721, and 0.657 for MITRALITY, GWTG [Get With The Guidelines], and GRASP [Getting Reduction of Mitral Insufficiency by Percutaneous Clip Implantation] scores, respectively).26
AI for Prediction of Cardiac Resynchronization Therapy Response and Outcomes
Current limitations and improvement of CRT response prediction using ML
Current limitations and improvement of CRT response prediction using ML
Cardiac resynchronization therapy (CRT), also known as biventricular pacing, entails the implantation of an additional left ventricular lead, via the coronary sinus, to mimic physiological ventricular conduction across both the right and left ventricles. Whereas CRT has been shown to improve outcomes by inducing reverse remodeling and decreasing intraventricular and atrioventricular dyssynchrony, approximately one third of patients with HF do not have an adequate clinical response to CRT, highlighting the limitations of the current selection criteria.3 For instance, current guidelines predominantly rely on QRS morphology and duration on electrocardiography, and on HF symptoms, for selecting candidates for CRT. Yet, there is a marked patient heterogeneity in terms of clinical characteristics, patterns of mechanical dyssynchrony, and LV lead placement—factors that are overlooked in the current guidelines and that influence outcomes after CRT.
Several investigations have demonstrated the correlation between individual electrocardiographic and imaging parameters and CRT outcomes, whereas ML has been specifically used to combine these variables so as to perform phenogrouping of individuals who might have a CRT response.8,9,27-32 For example, Cikes et al8 created an unsupervised ML algorithm to identify 4 phenogroups with varying outcomes after CRT. To incorporate the dynamic LV function during the cardiac cycle, LV strain and volumes were transcribed to a set of data points (1,632 data points), which were then combined with 50 clinical, electrocardiographic, and echocardiographic parameters. After performance of dimensionality reduction, k-means clustering was used to generate 4 phenogroups. Two phenogroups (phenogroups 1 and 3) with a high proportion of female subjects, nonischemic cardiomyopathy, and left bundle branch block were more likely to exhibit a response after CRT implantation when compared directly with their counterparts (HR: 0.35 and HR: 0.36; P = 0.001 and P = 0.0005, respectively) (Figure 5). Furthermore, multiple studies have used ML to incorporate a wide range of parameters to predict response and mortality after CRT implantation; these are summarized in Table 2.
Phenogrouping of Individuals Based on CRT Response
Individuals in phenogroups 1 and 3 had the maximal CRT response and had significant differences compared with phenogroups 2 and 4 in their clinical characteristics. Adapted with permission from Cikes et al.8 CRT = cardiac resynchronization therapy; DM = diabetes mellitus; FAC = fractional area change; GLS = global longitudinal strain; HF = heart failure; HR = heart rate; HTN = hypertension; ICD = implantable cardioverter-defibrillator; ICM = ischemic cardiomyopathy; LA = left atrium; LBBB = left bundle branch block; LV = left ventricle; LVEF = left ventricular ejection fraction; MRA = mineralocorticoid receptor antagonist; NYHA = New York Heart Association; RV = right ventricle; SBP = systolic blood pressure.
| First Author, Year | Sample Size (n) and Study Description | ML Technique Used/Best Performing Model | Results | Limitations | ||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Cai et al,27 2021 | n = 1,664; 487 preimplantation clinical, laboratory, ECG, and imaging variables used to devise CRT risk calculator. 1° endpoint: ΔLVEF >5% at 6 mo follow-up. | Ensemble of ensemble | Model performance better than traditional ML models (AUC 0.66, 0.64, and 0.73 for RF, decision tree, and EoE, respectively). |
| ||||||||||||||||||||||||||||||
| Cikes et al,8 2018 | n = 1,106; 50 clinical and ECG parameters combined with data from speckle tracking echocardiography to phenogroup patients according to CRT response. 1° endpoint: Death of any cause or a nonfatal heart failure event. | k-means clustering | Of 4 phenogroups, groups 1 and 3 were associated with better CRT response (HR: 0.35 and 0.36) when compared with other groups (Figure 5). |
| ||||||||||||||||||||||||||||||
| Feeny et al,32 2019 | n = 935; 9 clinical, ECG, and echocardiographic parameters incorporated to predict CRT response and survival. 1° endpoint: 1. ΔLVEF >10% after CRT. 2. Survival from composite endpoint of death, transplantation, or LVAD. | Naïve Bayes | ML model predicted mortality better than current guidelines (Class I, IIa, and IIb (AUC: 0.70 vs 0.65, respectively). |
| ||||||||||||||||||||||||||||||
| Feeny et al,9 2020 | n = 925; PCA of the QRS vectors on the ECG done; combined with k-means clustering to stratify patients into 2 groups. 1° endpoint: 1. Composite endpoint of death, transplantation, or LVAD 2. ΔLVEF after CRT | PCA with k-means clustering | As compared with QRS PCA group 2, QRS PCA group 1:
|
| ||||||||||||||||||||||||||||||
| Galli et al,28 2021 | n = 193; 28, clinical, ECG, and echocardiographic parameters used to phenogroup individuals and identify variable importance for CRT response. 1° endpoint: composite of LVAD, heart transplantation, or ACM. | RF |
|
| ||||||||||||||||||||||||||||||
| Howell et al,39 2021 | n = 741; 19 clinical, ECG, echocardiography, and LV lead position used to predict short-term mortality after CRT. 1° endpoint: Composite of freedom from death and hospitalization, and ΔLVEF >15% at 6 months after implantation. | Adaptive Lasso | High accuracy of ML model (AUC: 0.759); 19 variables identified, out of which 8 were modifiable (Figure 6). |
| ||||||||||||||||||||||||||||||
| Hu et al,30 2019 | n = 990; unstructured data from electronic health records used to identify CRT nonresponders. 1° endpoint: ΔLVEF <0% at 6-18 mo after CRT implantation. | Natural language processing and gradient boosting | Final model had an accuracy of 0.65, with a positive predictive value of 0.79 to predict CRT nonresponse. |
| ||||||||||||||||||||||||||||||
| Lei et al,31 2019 | n = 117; 31 preimplantation clinical, ECG, and echocardiographic variables to predict CRT response. 1° endpoint: Improvement in NYHA functional class or ΔLVEF ≥15% at 6 mo after CRT. | SVM |
|
| ||||||||||||||||||||||||||||||
| Liang et al,29 2021 | n = 752; 19 preimplantation variables (clinical, ECG, echo, and LV lead characteristics) to predict CRT response. 1° endpoint: ΔLVEF >10% at 1 y follow-up. | Boosting, LR, and Ridge model | LR, ensemble and ridge model attained highest predictive power (AUC: 0.77), with improvement over Class Ia and IIa guidelines (AUC: 0.54 and 0.53, respectively). |
| ||||||||||||||||||||||||||||||
| Kalscheur et al,40 2018 | n = 481; 45 clinical, ECG, and echocardiographic parameters used to predict CRT outcomes. 1° endpoint: composite ACM and HF hospitalization at 12 mo I after CRT. | RF |
|
| ||||||||||||||||||||||||||||||
| Tokodi et al,41 2020 | n = 1,510; 33 preimplantation variables used to develop a score to predict post-CRT outcomes. 1° endpoint: ACM at 1, 2, 3, 4, and 5 y. | RF | Semmelweis-CRT score performed better than existing scores (mean AUC 0.785 and 0.541 for ML model and Seattle heart failure model, respectively for prediction of 1- to 5-y mortality). |
| ||||||||||||||||||||||||||||||
In addition to biventricular dyssynchrony, the role of right ventricular (RV) dysfunction in predicting outcomes after CRT is still debatable.33,34 A recent ML approach by Galli et al28 combined clinical, electrocardiographic, LV, and RV echocardiography parameters to predict CRT response and outcomes. Out of the 16 variables that were predictive of CRT response, 8 pertained to RV dynamics. Of note, septal flash pattern and apical rocking (motion of apex perpendicular to the longitudinal LV axis in systole) were the most important parameters. QRS interval was the tenth most crucial parameter in the variable importance list, highlighting the complex interplay of factors determining CRT response.
Optimal lead positioning is vital to an adequate CRT response, with LV delay (Q-LV) or interventricular delay (RV-LV) guided lead placement being superior to a simple anatomical lead placement approach.35,36 Decision trees using RV-LV and Q-LV duration have been shown to delineate individuals with better CRT outcomes. Whether combining these variables and clinical variables yields better results remains an area that needs further exploration.37
Identifying high-risk candidates after CRT implantation
Immediately after CRT implantation, a multidisciplinary approach incorporating aggressive risk factor modification and dynamic atrioventricular optimization has been shown to improve long-term outcomes.38 A recent ML model devised by Howell et al39 used 19 variables to predict 6-month mortality with good discriminatory power (AUC: 0.759). Interestingly, half of the variables were modifiable, and atrioventricular optimization was associated with lower mortality risk in those subgroups (Figure 6). Moreover, after patients were stratified via the ML algorithm, participants in the fifth quintile had 14 times higher odds of achieving a CRT response at the 6-month follow-up visit. Therefore, the availability of such tools that can better predict CRT response can help further the notion of precision medicine.
Machine Learning in Improving Short-Term CRT Response
(A) Role of lifestyle and medical interventions, along with adaptive AV optimization, in improving short-term CRT response. (B) Moreover, ML model based on the characteristics fed could stratify patients with differential short-term outcomes after CRT implantation. Adapted with permission from Howell et al.39 AV = atrioventricular; BP = blood pressure; CM = cardiomyopathy; CRP = C-reactive protein; DS = disease; eGFR = estimated glomerular filtration rate; LVEDVI = left ventricular end-diastolic volume index; LVESDI = left ventricular end-systolic dimension index; NT-proBNP = N-terminal pro–B-type natriuretic peptide; OR = odds ratio; other abbreviations as in Figures 3 and 5.
ML has also been used to devise risk scores for mortality prediction after CRT.40,41 Using preimplantation variables, the random forest-based Semmelwies-CRT developed by Tokodi et al41 outperformed contemporary risk scores in predicting mortality at 1 to 5 years after CRT (AUC for ML-based and Seattle HF model 0.768 vs 0.537 and 0.803 vs 0.544 at 1 year and 5 years respectively; P < 0.05). Although predicting CRT response can lead to better patient selection, accurate prediction of mortality, both short-term and long-term, can help in resource allocation, earlier referral for higher-risk patients, and room for earlier intervention, overall improving patient outcomes (Figure 7).
Implications of Integrating ML Algorithms Into Workflow Optimization for CRT
ML-derived phenogrouping can be used to identify high-risk individuals and patients most likely to benefit and respond to CRT. Further, accurate prognostication can allow for earlier intervention and referral for advanced therapies if goal-directed medical therapies and CRT fail. Created with Biorender.com. 2-D = 2-dimensional; LVAD = left ventricular assist device; Q-LV = intrinsic left ventricular delay; RV-LV = interventricular delay; other abbreviations as in Figures 3, 5, and 6.
AI in Implantable Cardiac Defibrillators
Implantable cardiac defibrillators (ICDs) have been the standard of care to improve outcomes in patients with HF, and ML has been applied to improve patient mortality prediction and ICD outcomes (Supplemental Table 3).42-48 Using the concept of heart rate variability (HRV), ML has recently been used to predict ventricular arrhythmias up to an hour before they happen instead of ICD being used as a reactive mechanism to abort arrhythmias.42-45 Au-Yeung et al45 developed ML models to predict ventricular tachyarrhythmia up to 5 minutes before the event, with increasing accuracy closer to the arrhythmic event (mean AUC for prediction was 0.81 at 5 minutes and 0.87-0.88 at 10 seconds before the event, respectively; P < 0.05). A major challenge with implementing this technology is that the energy-intensive requirements the algorithms place on modern-day ICDs is prohibitive because of the limited battery life of ICDs. Provided the above-mentioned limitation is overcome, ML can be used as a tool in the future for early prediction of life-threatening ventricular arrhythmia, timely allocation of medical resources, and subsequent prevention of sudden cardiac deaths.
ML has also been used to determine the influence of variables in the prediction of appropriate ICD tachyarrhythmia therapies. Incorporating clinical, serum biomarkers and cardiac MRI features, Wu et al46 analyzed data from 328 patients to develop a random forest model to predict appropriate device therapies and mortality. The model performed well in predicting appropriate device therapy and outperformed the Seattle HF model in predicting mortality (AUC for random forest and Seattle HF model of 0.88 and 0.53, respectively). Interestingly, serial LV ejection fraction was not a significant predictor for appropriate ICD therapies.
ML has also been used to assist clinicians in issues that commonly arise during the care of patients with advanced HF and ICD therapy. By using a smartphone-based interface, ML has been used to identify the type and brand of ICD based on chest radiography, which can prevent patient care delays in emergency and inpatient settings.47,49 More recently, ML has been used to identify risk factors predicting infection after ICD insertion. Using Lasso regression, investigators identified 17 factors associated with risk of infection, of which 11 were potentially modifiable, with the length of procedure time, anticoagulant use, and device implant location being the most critical variables.48 Whereas the previously used risk scores emphasized the role of nonmodifiable factors, identification of modifiable risk factors via newer approaches can help institute preventative measures, further improving outcomes.50
AI to Improve Left Ventricular Assist Device Outcomes
Although LVADs have led to a paradigm shift in the management of advanced HF, a multipronged strategy composed of accurate risk prediction, preoperative optimization, and postoperative prognostication is lacking. ML has been increasingly applied toward improving prediction of RV failure, earlier recognition of adverse events, and prediction of mortality after LVAD implantation.51-62
Challenges associated with predicting postoperative RV failure and the emergence of ML
With an incidence estimated to be approximately 9% to 42%, RV failure after LVAD implantation is thought to be secondary to increased RV preload and the concomitant interventricular septal shift toward the LV, leading to impaired RV contractility.63 Despite a high incidence, prediction models developed in the past have shown a modest to poor performance, necessitating a need for further improvement.63
ML models to compute RV dynamics have been described using both 2-D and 3-D echocardiography with high reproducibility and good performance.51,52 A recent ML algorithm for computation of RV ejection fraction, RV end-systolic volume (RV ESV), and RV end-diastolic volume (RV EDV) from 3-D echocardiography was able to achieve an excellent correlation with values derived from cardiac magnetic resonance (MR) (r = 0.87-0.91 for all; P < 0.001) in a time-efficient manner (15 seconds to <2 minutes).52 Such models not only hold the promise of minimizing interobserver variability but also can augment the performance of predictive models for RV dysfunction.
Dating back to 2012, the Pittsburgh score was the first decision tree algorithm used to predict RV failure after LVAD implantation.62 Although it performed poorly on validation cohorts, newer ML models have been described.53,57 More recently, Bellavia et al53 used a supervised ML approach to identify parameters predictive of acute and chronic RV failure. They identified central venous pressure, Michigan risk score, and the RV apical longitudinal strain as the 3 most important independent predictors of acute RV failure (AUC for the model incorporating the above variables: 0.95, 95% CI: 0.91-1.00). Of note, tricuspid annular plane systolic excursion, right atrial strain, and RV free wall strain in the middle segment were found to be most predictive of chronic RV failure (AUC for the model incorporating the mentioned variables: 0.97, 95% CI: 0.91-1.00). A detailed description of studies using ML to improve the prediction of RV dysfunction after LVAD implantation and LVAD outcomes is provided in Table 3.
| First Author, Year | Sample Size (n) and Description | ML Model | Results | Limitations | ||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Bellavia et al,53 2020 | n = 74; clinical, hemodynamic, and imaging variables used to predict postoperative RV failure. 1° endpoint: Acute and chronic RV failure. | Naïve Bayes | MRS, CVP, and apical longitudinal systolic strain of RV predictive of acute RV failure (AUC: 0.95). RV free wall strain of the middle segment, right atrial strain, and tricuspid annular systolic plane excursion were predictive of chronic RV failure (AUC: 0.97). |
| ||||||||||||||||||
| Kanwar et al,54 2018 | n = 10,277; preimplantation clinical variables incorporated into ML model to predict prognosis. 1° endpoint: Mortality at 1, 3, and 12 mo after LVAD implantation. | Naïve Bayes | AUC for ML model 0.70, 0.71, and 0.70 at 1, 3, and 12 mo respectively. |
| ||||||||||||||||||
| Kilic et al,55 2021 | n = 16,120; outpatient and inpatient clinical variables before LVAD insertion and concomitant operative procedural variables assimilated into an ML model. 1° endpoint: 90-d and 1-y mortality after LVAD implantation. | XgBoost | ML model resulted in improved prediction as compared with LR for prediction of 90-d (C-index: 0.74 vs 0.70) and 1-y (C-index: 0.71 vs 0.69) mortality. |
| ||||||||||||||||||
| Kilic et al,56 2021 | n = 568; patients clustered in terms of adverse event profile, and exploratory analysis to predict the transition from 1 adverse event to other among patients. 1° endpoint: Occurrence of adverse event during 5-y follow-up. | Hierarchical clustering | Patients divided into 5 clusters based on adverse event profile. Bleeding and RV failure most common in the early phase. Infection and bleeding common in late phase after LVAD implantation. |
| ||||||||||||||||||
| Loghmanpur et al,57 2015 | n = 10,909; clinical, echocardiographic, and ECG variables used to predict acute (<48 h), early (<14 d) and late onset (>14 d) RV failure. 1° endpoint: RV failure after LVAD implantation. | Naïve Bayes | AUC 0.90, 0.84, and 0.88 for prediction of acute, early, and late RV failure via ML; outperforming Drakos and MRS (AUC: 0.55 and 0.50, respectively, for prediction of RV failure). |
| ||||||||||||||||||
| Misumi et al,58 2021 | n = 13; sound signals via an electronic stethoscope were recorded to extract 19 feature vectors, which were then incorporated into ML model. 1° endpoint: predicting AR at 1, 3, and 12 mo after LVAD implantation. | Ensemble classifier | Amplitude of the first harmonic, LVAD rotational speed during ILS, and the variation in the amplitude during normal rotation and ILS were the most useful features, which when combined via ensemble method, predicted AR with accuracy of 91%. |
| ||||||||||||||||||
| Shad et al,59 2022 | n = 723; DL used to train a model based on echocardiographic videos for the prediction of RV failure after LVAD implantation. | Convolutional neural networks | Deep learning model performed better than CRITT79 and Penn score (AUC for deep learning, Penn, and CRITT score: 0.73, 0.61, and 0.62 respectively). |
| ||||||||||||||||||
| Simkowski et al,60 2020 | n = 31; based on LGE patterns on cardiac MR, patients were grouped into clusters to predict the risk of RV failure after LVAD implantation. | Hierarchical clustering | 3 clusters were identified; cluster 2 (with extensive transmural LGE patterns indicative of ischemic CMP) at lower risk for RV failure than cluster 1 (no or atypical LGE enhancement; nonischemic CMP) or 3 (some subendocardial LGE but no extensive LGE, likely mixed). |
| ||||||||||||||||||
| Topkara et al,61 2021 | n = 20,270; 28 clinical variables were included to form an ML model to predict myocardial recovery. 1° endpoint: LVAD explantation for myocardial recovery. | Naïve Bayes | ML predicted myocardial recovery with highest accuracy, as compared to I-CARS and I-TOPS risk scores (AUC: 0.824, 0.748, and 0.744 for ML, I-CARS, and I-TOPS, respectively). |
| ||||||||||||||||||
| Wang et al,62 2012 | n = 183; 8 preoperative clinical variables used to predict postoperative RV failure. 1° endpoint: RVAD implantation after LVAD implantation. | Decision trees | ML model had a better performance than MRS (AUC: 0.87 and 0.54 for the ML and MRS, respectively). |
| ||||||||||||||||||
ML in the prediction of outcomes after LVAD implantation
With use of the INTERMACS (Interagency Registry for Mechanically Assisted Circulatory Support) registry, ML models have been developed to predict short-term and long-term mortality after LVAD implantation.54-56,64 A recent ML approach developed by Kilic et al55 achieved higher performance than the LR-based model to predict both short-term and long-term mortality (AUC 0.74 vs 0.70 for short-term and 0.71 vs 0.69 for long-term mortality for ML and the LR model, respectively).
A small proportion of patients experience remarkable clinical recovery after LVAD implantation, allowing for LVAD explantation. Identifying this stratum of patients can allow for directed resource utilization while maximizing patient outcomes.65 Using INTERMACS registry, Topkara et al61 used ML to identify 28 clinical variables associated with recovery. The ML model fared better than the established risk scores (AUC of 0.82, 0.74, and 0.75 for ML model, I-CARS, and INTERMACS recovery score, respectively; P < 0.05).
Although the models mentioned above have been devised and performed reasonably well when applied to the INTERMACS registry, they are limited by their derivation using older-generation LVAD populations. Additionally, the INTERMACS registry is composed of well-structured data (eg, clinical, laboratory, and hemodynamic parameters recorded in a “snapshot” (eg, preoperative, postoperative) fashion) and hence limits the incorporation of unstructured variables (eg, echocardiographic videos, longitudinal clinical parameters) for further improvement of ML algorithms. Future research incorporating raw unstructured data, especially the longitudinal clinical and laboratory data that are abundantly available in electronic health records (EHRs), is needed to capture the full potential of ML models to improve risk assessment and patient outcomes in such a challenging population of HF patients.
Current Challenges and Future Promises
Despite multiple promising applications, significant limitations need to be addressed before ML can be fully integrated into daily clinical practice. Most studies have been done on retrospective data and hence may suffer from data skewness and missing data, thereby yielding risk models that are not generalizable to the entire population. Moreover, ML algorithms have traditionally suffered from the black box conundrum, whereby models are inherently complex and might lack interpretability. Given that ML models aim to identify hidden patterns between variables and outcomes, not necessarily the causal relationship, model interpretation becomes even more tricky. Whereas imputation techniques such as MICE (multiple imputation by chained equations) can serve as interim solutions for missing data, models such as SHAP (Shapley additive explanation) and LIME (local interpretable model-agnostic explanations) have provided an initial breakthrough to help “open the black box of ML” and make the algorithms more explainable (eg, importance map plots for clinical variables and pixels for images).66-68
Another challenge to the widespread adoption of ML into routine clinical practice is the lack of reproducibility and external validity of constructed algorithms caused by restricted data sharing and the lack of large publicly available databases. Ethical considerations, lack of incentive to share data, or differing data formats can complicate data sharing across platforms and between investigators.69 In addition, databases generated with commercial partnerships, or for commercial purposes, are less likely to be released into the public domain.70 Publicly available large data sets, if available, can be used to test generalizability of the proposed models, which is critical to the implementation of ML models in clinical practice. BigData@Heart by the European Society of Cardiology is aimed at harmonizing data from worldwide registries for more than 5 million patients with cardiovascular disease, with an aim to develop algorithms to further patient care.71 A list of publicly available databases in the realm of HF that have been used to develop ML algorithms is shown in Supplemental Table 4. Furthermore, although public sharing of data will help test generalizability, validation of the proposed ML models in large-scale, well-designed prospective trials is imperative to their implementation. Standardized reporting is another challenge because there exists significant heterogeneity in terms of ML model development and testing, which makes comparisons between ML models challenging.72 Although multiple studies have shown that ML performs better than traditional statistics, a recent systematic review by Shin et al,73 which included studies that compared ML versus conventional statistics in terms of hospital readmissions and mortality, outlined concerns of bias in the majority of these studies, coupled with a lack of external validation of the ML models. The scientific community has taken multiple measures to facilitate standardized reporting in clinical trials involving AI, with the CONSORT-AI and SPIRIT-AI guidelines being recent examples (Figure 8).74,75
CONSORT-AI Guidelines for Clinical Trials Involving Artificial Intelligence
Adapted with permission from Liu et al.74 CONSORT-AI = Consolidated Standards of Reporting Trials–Artificial Intelligence.
Challenges specific to developing ML models for cardiovascular diseases also deserve a brief mention (Figure 9). With close to 80% of data in EHRs being unstructured (eg, clinical notes, imaging reports, there is a huge gap in data integration. Physician-documented notes and imaging reports can be a powerful tool that can significantly augment the power of the current ML algorithms. Whereas manual extraction of data from the reports can be too cumbersome, natural language processing algorithms can be applied to extract data, transforming these rich sources of data into structured variables and allowing for integration into ML models. With respect to imaging, Pandey et al76 demonstrated an excellent accuracy for retrieval of radiographic findings from the thoracoabdominal computed tomography (AUC 0.83-1.00 for 14 findings), with the Cox model derived from the extracted features having a good predictive power for 30-day all-cause mortality (AUC: 0.747). Data retrieval, when applied to clinical notes, can not only help augment the performance of ML models but also help move beyond the rigid and established definitions of a particular disease condition and allow for assimilation of all clinical data in a manner that does not take into account a prior assumption, which has been a major concern with contemporary research practices.77
Current Challenges and Prospects of Machine Learning in Cardiovascular Medicine
Unstructured data in EHRs, medical images, and the remote monitoring data can be combined and stored in innovative cloud platforms powered by blockchain technology. Bigger amounts of data, when fed to models, can lead to improved power, thus overall improving outcomes. Created with Biorender.com. EHR = electronic health records; SHAP = Shapley additive explanations; LIME = local interpretable model-agnostic explanation; ML = machine learning.
Moreover, image processing and reporting protocols vary across institutions (eg, different generations of scanners for computed tomography and MRI, different postprocessing protocols). A uniform standardization for image acquisition and reporting is imperative for better performance. DL-enabled automatic segmentation can allow for improved image quality. Moreover, DL-assisted image interpretation can be clubbed with other clinical parameters to enhance the performance of prediction models, furthering patient care. This will also require innovative cloud storage models, able to assimilate tons of gigabytes of data while simultaneously allowing for quick retrieval. Blockchain technology and smart contracts can mitigate this drawback by creating secure medical cloud storage and assisting in the real-time transfer of medical records among health care providers. AI coupled with blockchain technology has far-reaching applications in the future of medical data management, including but not limited to, enhanced encryption, data accountability, and fewer chances of data tampering.78
Finally, it is important to note that although the present review focuses on ML applications that could lead to improved outcomes within the realm of device therapies in HF patients, there is the potential for direct benefit to health care providers. Automation of repetitive and mundane processes such as the interpretation of normal study results in low-risk patients can help alleviate the burden on multitasking providers. This could help providers focus on more challenging situations and studies and, in the long run, could help prevent physician burnout.
Conclusions
The present review elucidates the role of AI in device therapies in HF. AI-powered data-driven outcomes research not only holds the potential to expand the current understanding of complex diseases such as HF but also promises to expand the approach to make it more patient-centric—a goal central to the theme of health care in the twenty-first century. More importantly, it is vital to consider current constraints that limit the widespread application of ML in day-to-day care, with emphasis on restriction in data sharing and the lack of standardized reporting.
Funding Support and Author Disclosures
Dr Greene has received research support from the Duke University Department of Medicine Chair’s Research Award, American Heart Association, Amgen, AstraZeneca, Bristol Myers Squibb, Cytokinetics, Merck, Novartis, Pfizer, and Sanofi; has served on advisory boards for Amgen, AstraZeneca, Bristol Myers Squibb, Cytokinetics, Roche Diagnostics, and Sanofi; has received speaker fees from Boehringer Ingelheim; and serves as a consultant for Amgen, Bayer, Bristol Myers Squibb, Merck, PharmaIN, Sanofi, Tricog Health, Urovant Pharmaceuticals, and Vifor. Dr Al’Aref is supported by National Institutes of Health 2R01 HL12766105 and 1R21 EB030654; and has received royalty fees from Elsevier. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.
Abbreviations and Acronyms
| AI | artificial intelligence |
| CRT | cardiac resynchronization therapy |
| ECG | electrocardiography |
| HF | heart failure |
| ICD | implantable cardiac defibrillator |
| LV | left ventricle |
| LVAD | left ventricular assist device |
| ML | machine learning |
| PAP | pulmonary artery pressure |
| PCWP | pulmonary capillary wedge pressure |
| RV | right ventricle |
| TEE | transesophageal echocardiogram |
| TEER | transcatheter edge-to-edge repair |
References
1. "Heart Disease and Stroke Statistics—2021 Update: a report from the American Heart Association". Circulation 2021;143:8: e254-e743.
2. "A systematic review of medical costs associated with heart failure in the USA (2014–2020)". PharmacoEconomics 2020;38:11: 1219-1236.
3. "The problem of non-response to cardiac resynchronization therapy". Curr Opin Cardiol 2006;21:1: 20-26.
4. "Statistics versus machine learning". Nat Methods 2018;15:4: 233-234.
5. "Machine learning and the future of cardiovascular care: JACC state-of-the-art review". J Am Coll Cardiol 2021;77:3: 300-313.
6. "Types of machine learning algorithms". New Advances in Machine Learning 2010;3:19-48.
7. "Smelling heart failure from human skin odor with an electronic nose". Annu Int Conf IEEE Eng Med Biol Soc 2012;2012:4034-4037.
8. "Machine learning-based phenogrouping in heart failure to identify responders to cardiac resynchronization therapy". Eur J Heart Fail 2019;21:1: 74-85.
9. "Machine learning of 12-lead QRS waveforms to identify cardiac resynchronization therapy patients with differential outcomes". Circ Arrhythm Electrophysiol 2020;13:7: e008210.
10. "Unsupervised machine learning for networking: Techniques, applications and research challenges". IEEE Access 2019;7:65579-65615.
11. "Deep learning in medical imaging: General overview". Korean J Radiol 2017;18:4: 570-584.
12. "Isosorbide mononitrate in heart failure with preserved ejection fraction". N Engl J Med 2015;373:24: 2314-2324.
13. "Continuous wearable monitoring analytics predict heart failure hospitalization". Circ Heart Fail 2020;13:3: e006513.
14. "Estimation of changes in intracardiac hemodynamics using wearable seismocardiography and machine learning in patients with heart failure: a feasibility study". IEEE Trans Biomed Eng 2022;69:8: 2443-2455. https://doi.org/10.1109/TBME.2022.3147066.
15. "Novel wearable seismocardiography and machine learning algorithms can assess clinical status of heart failure patients". Circ Heart Fail 2018;11:1: e004313.
16. "A multisensor algorithm predicts heart failure events in patients with implanted devices: results from the MultiSENSE study". J Am Coll Cardiol HF 2017;5:3: 216-225.
17. "Pulmonary artery pressure-guided therapy in ambulatory patients with symptomatic heart failure: the CardioMEMS European monitoring study for heart failure (MEMS-HF)". Eur J Heart Fail 2020;22:10: 1891-1901.
18. "2021 ESC/EACTS guidelines for the management of valvular heart disease". Eur J Cardiothorac Surg 2021;60:727-800.
19. "Intraobserver variability in grading severity of repeated identical cases of mitral regurgitation". Am Heart J 2008;156:6: 1089-1094.
20. "Reproducibility of proximal isovelocity surface area, vena contracta, and regurgitant jet area for assessment of mitral regurgitation severity". J Am Coll Cardiol Img 2010;3:3: 235-243.
21. "Fully automated echocardiogram interpretation in clinical practice". Circulation 2018;138:16: 1623-1635.
22. "Automated quantification of mitral valve anatomy using anatomical intelligence in three-dimensional echocardiography". Int J Cardiol 2015;199:232-238.
23. "Using anatomic intelligence to localize mitral valve prolapse on three-dimensional echocardiography". J Am Soc Echocardiogr 2016;29:10: 938-945.
24. "Mitral annulus segmentation using deep learning in 3-D transesophageal echocardiography". IEEE J Biomed Health Inform 2020;24:4: 994-1003.
25. "Machine-learning-based in-hospital mortality prediction for transcatheter mitral valve repair in the United States". Cardiovasc Revasc Med 2021;22:22-28.
26. "Machine learning identifies clinical parameters to predict mortality in patients undergoing transcatheter mitral valve repair". J Am Coll Cardiol Intv 2021;14:18: 2027-2036.
27. "Using ensemble of ensemble machine learning methods to predict outcomes of cardiac resynchronization". J Cardiovasc Electrophysiol 2021;32:9: 2504-2514.
28. "Importance of systematic right ventricular assessment in cardiac resynchronization therapy candidates: a machine learning approach". J Am Soc Echocardiogr 2021;34:5: 494-502.
29. "Prediction of response after cardiac resynchronization therapy with machine learning". Int J Cardiol 2021;344:120-126.
30. "Can machine learning improve patient selection for cardiac resynchronization therapy?". PloS One 2019;14:10: e0222397.
31. "Ventricular geometry-regularized QRSD predicts cardiac resynchronization therapy response: machine learning from crosstalk between electrocardiography and echocardiography". Int J Cardiovasc Imaging 2019;35:7: 1221-1229.
32. "Machine learning prediction of response to cardiac resynchronization therapy: improvement versus current guidelines". Circ Arrhythm Electrophysiol 2019;12:7: e007316.
33. "Interplay between right ventricular function and cardiac resynchronization therapy: an analysis of the CARE-HF Trial (cardiac resynchronization–heart failure)". J Am Coll Cardiol 2013;61:21: 2153-2160.
34. "Meta-analysis of the relation of baseline right ventricular function to response to cardiac resynchronization therapy". Am J Cardiol 2016;117:8: 1315-1321.
35. "The relationship between ventricular electrical delay and left ventricular remodelling with cardiac resynchronization therapy". Eur Heart J 2011;32:20: 2516-2524.
36. "Left ventricular lead electrical delay predicts response to cardiac resynchronization therapy". Heart Rhythm 2006;3:11: 1285-1292.
37. "Comparison of measures of ventricular delay on cardiac resynchronization therapy response". Heart Rhythm 2020;17:4: 615-620.
38. "Multidisciplinary care of patients receiving cardiac resynchronization therapy is associated with improved clinical outcomes". Eur Heart J 2012;33:17: 2181-2188.
39. "Using machine-learning for prediction of the response to cardiac resynchronization therapy: the SMART-AV study". J Am Coll Cardiol EP 2021;7:12: 1505-1515.
40. "Machine learning algorithm predicts cardiac resynchronization therapy outcomes: lessons from the COMPANION trial". Circ Arrhythm Electrophysiol 2018;11:1: e005499.
41. "Machine learning-based mortality prediction of patients undergoing cardiac resynchronization therapy: the SEMMELWEIS-CRT score". Eur Heart J 2020;41:18: 1747-1756.
42. "Machine learning approach to predict ventricular fibrillation based on QRS complex shape". Front Physiol 2019;10:1193.
43. "Prediction of ventricular tachycardia one hour before occurrence using artificial neural networks". Sci Rep 2016;6:1: 1-7.
44. "Prediction of spontaneous ventricular tachyarrhythmia by an artificial neural network using parameters gleaned from short-term heart rate variability". Expert Systems with Applications 2012;39:3: 3862-3866.
45. "Development and validation of warning system of ventricular tachyarrhythmia in patients with heart failure with heart rate variability data". PloS One 2018;13:11: e0207215-e.
46. "Baseline and dynamic risk predictors of appropriate implantable cardioverter defibrillator therapy". J Am Heart Assoc 2020;9:20: e017002.
47. "Deep learning-based algorithm for the detection and characterization of MRI safety of cardiac implantable electronic devices on chest radiographs". Korean J Radiol 2021;22:11: 1918-1928.
48. "Risk factors for CIED infection after secondary procedures: insights from the WRAP-IT trial". J Am Coll Cardiol EP 2022;8:1: 101-111.
49. "A head-to head comparison of machine learning algorithms for identification of implanted cardiac devices". Am J Cardiol 2021;144:77-82.
50. "Risk factors for infections involving cardiac implanted electronic devices". J Am Coll Cardiol 2019;74:23: 2845-2854.
51. "Development of novel machine learning model for right ventricular quantification on echocardiography: a multimodality validation study". Echocardiography 2020;37:5: 688-697.
52. "Machine learning–based three-dimensional echocardiographic quantification of right ventricular size and function: validation against cardiac magnetic resonance". J Am Soc Echocardiogr 2019;32:8: 969-977.
53. "Usefulness of regional right ventricular and right atrial strain for prediction of early and late right ventricular failure following a left ventricular assist device implant: a machine learning approach". Int J Artif Organs 2020;43:5: 297-314.
54. "A Bayesian model to predict survival after left ventricular assist device implantation". J Am Coll Cardiol HF 2018;6:9: 771-779.
55. "Using machine learning to improve risk prediction in durable left ventricular assist devices". PLoS One 2021;16:3: e0247866.
56. "Machine learning approaches to analyzing adverse events following durable LVAD implantation". Ann Thorac Surg 2021;112:3: 770-777.
57. "A Bayesian model to predict right ventricular failure following left ventricular assist device therapy". J Am Coll Cardiol HF 2016;4:9: 711-721.
58. "Prediction of aortic valve regurgitation after continuous-flow left ventricular assist device implantation using artificial intelligence trained on acoustic spectra". J Artif Organs 2021;24:2: 164-172.
59. "Predicting post-operative right ventricular failure using video-based deep learning". Nat Commun 2021;12:1: 5192.
60. "Unsupervised machine learning of LGE patterns on cardiac MRI identifies patients at risk for right ventricular failure after LVAD". J Card Fail 2020;26:10: S146.
61. "Machine learning-based prediction of myocardial recovery in patients with left ventricular assist device support". Circ Heart Fail 2022;15:1: e008711.
62. "Decision tree for adjuvant right ventricular support in patients receiving a left ventricular assist device". J Heart Lung Transplant 2012;31:2: 140-149.
63. "Predicting the risk of right ventricular failure in patients undergoing left ventricular assist device implantation". Circ Heart Fail 2020;13:10: e006994.
64. "Improving mortality predictions for patients with mechanical circulatory support using follow-up data and machine learning". Circ Genom Precis Med 2020;13:2: e002877.
65. "Prospective multicenter study of myocardial recovery using left ventricular assist devices (RESTAGE-HF [remission from stage D heart failure])". Circulation 2020;142:21: 2016-2028.
66. "A unified approach to interpreting model predictions". Adv Neural Inf Process Syst 2017;30:4768-4777.
67. "“Why should I trust you?” Explaining the predictions of any classifier". In: , eds. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 2016.
68. "Evaluating the state of the art in missing data imputation for clinical data". Brief Bioinform 2021;23:1: bbab489.
69. "The conundrum of sharing research data". J Am Soc Inf Sci Tech 2012;63:6: 1059-1078.
70. "The primary reasons behind data sharing, its wider benefits and how to cope with the realities of commercial data". BMC Genomics 2015;16:1: 626.
71. "Identification and mapping real-world data sources for heart failure, acute coronary syndrome, and atrial fibrillation". Cardiology 2022;147:1: 98-106.
72. "Addressing bias: artificial intelligence in cardiovascular medicine". Lancet Digit Health 2020;2:12: e635-e636.
73. "Machine learning vs. conventional statistical models for predicting heart failure readmission and mortality". ESC Heart Fail 2021;8:1: 106-115.
74. "Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension". Nat Med 2020;26:9: 1364-1374.
75. "Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension". Lancet Digit Health 2020;2:10: e549-e560.
76. "Extraction of radiographic findings from unstructured thoracoabdominal computed tomography reports using convolutional neural network based natural language processing". PLoS One 2020;15:7: e0236827.
77. "A natural language processing-based approach for identifying hospitalizations for worsening heart failure within an integrated health care delivery system". JAMA Netw Open 2021;4:11: e2135152.
78. "Blockchain technology in healthcare: a comprehensive review and directions for future research". Appl Sci 2019;9:9: 1736.
79. "Predicting right ventricular failure in the modern, continuous flow left ventricular assist device era". Ann Thorac Surg 2013;96:3: 857-863. https://doi.org/10.1016/j.athoracsur.2013.03.099.
Footnotes
The authors attest they are in compliance with human studies committees and animal welfare regulations of the authors’ institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the Author Center.
