Machine Learning in Non-Ischemic Cardiomyopathy: Phenotyping, Mechanism Discovery, and Clinical Applications

Abstract

Non-ischemic cardiomyopathies (NICMs) are a heterogeneous group of myocardial disorders whose shared phenotypes complicate diagnosis and risk stratification. This complexity has driven growing interest in computational approaches that can integrate high-dimensional data and capture patterns not evident through conventional analyses. In this review, we synthesize studies published between 2020 and 2026 that apply machine learning and deep learning models to NICMs across three interconnected domains: phenotype classification, mechanism discovery, and clinical decision support. We find that imaging and ECG-based models help improve subtype discrimination, arrhythmia detection, and early disease identification; genomic and multi-omics approaches advance variant interpretation and biomarker discovery; and emerging multimodal frameworks extend these efforts toward outcome prediction and individualized management. Challenges remain in generalizability, interpretability, and prospective validation, as well as in integrating modalities into patient-level risk models. Continued development of multimodal and longitudinal approaches will be essential for translating these advances into precision care for NICMs.

Introduction

Non-ischemic cardiomyopathy (NICM) refers to impaired left ventricular systolic function without obstructive coronary artery disease and represents 30%–40% of cardiomyopathy patients in heart failure trials.^[1] NICM is characterized by substantial patient heterogeneity and overlapping clinical determinants.^[2] Consequently, progress in clinical management is limited by incomplete mechanistic understanding and a reduced ability to predict individual disease trajectories. For instance, patients with similar left ventricular ejection fractions may experience markedly different outcomes, ranging from malignant arrhythmias to progressive heart failure.^[3] Traditional diagnostic and risk-stratification approaches, often based on categorical thresholds or single-modality assessments, struggle to capture non-linear interactions across variables.^[4]

Machine learning (ML) and deep learning (DL) have emerged as helpful approaches for integration of high-dimensional, multimodal data and modeling complex relationships. In NICM research, these approaches span a continuum of tasks: identifying phenotypes, uncovering mechanisms, and informing clinical decision-making. Their application has improved phenotype inference using imaging and ECG data, enabled genomic and multi-omics analyses to clarify disease mechanisms, and supported outcome prediction and therapy guidance in heterogeneous patient populations. These domains are inherently interconnected: Imaging phenotypes reflect molecular remodeling, electrophysiologic abnormalities may precede structural disease, and genetic variation shapes both phenotype and risk. ML offers a methodological bridge across these scales, enabling integration of structural, functional, and molecular information into unified disease models. However, challenges remain, including generalizability, interpretability, prospective validation, and translation into clinically deployable tools.

In this review, we summarize ML/DL studies published between 2020 and 2026, a period marked by rapid expansion in the application of ML/DL approaches to cardiomyopathy research. (Table 1). We focus on three interconnected domains: phenotype classification, mechanism modeling, and clinical decision support (Figure 1). We highlight how each addresses longstanding limitations in NICM research and care, and outline priorities for advancing toward precision, patient-centered care.

Table 1.

Summary of machine learning and deep learning studies in non-ischemic cardiomyopathies (2020–2026).

Study	Year	Task	ML/DL Models	Model Class
Van de Leur et al.	2021	Final adverse-event prediction	Support Vector Machine	Kernel method
Zhang et al.	2021	Variant pathogenicity prediction	AdaBoost (CardioBoost)	Ensemble method
Kanapeckaite et al.	2021	Biomarker/target prediction	Two-step machine learning predictor	Ensemble method
		Pathway discovery	Network-based modeling	Graph-based model
Al Hinai et al.	2021	ECG-based disease detection	Convolutional neural network	Neural network
		Baseline comparison	Logistic Regression	Linear model
		Baseline comparison	Support Vector Machine	Kernel method
		Classification	Dense neural network	Neural network
Chrispin et al.	2023	Arrhythmia substrate modeling	Mechanistic cardiac simulations	Physics-based model
		Integrative patient-level prediction	Multimodal AI frameworks	Neural network
Zhang et al.	2023	Cardiomyopathy classification	Support Vector Machine	Kernel method
		Pattern discrimination	K-nearest neighbor (K-NN)	Instance-based
		Non-linear feature modeling	Random Forest	Ensemble model
		Baseline classifier	Logistic Regression	Linear model
		Interpretable classification	Decision Tree	Tree-based model
		High-performance classifier	Gradient Boosting	Ensemble model
Lampert et al.	2023	Cardiomyopathy prediction from ECG	Deep neural network	Neural network
Kolk et al.	2023	Mortality prediction model	Extreme Gradient Boosting	Ensemble model
Kolk et al.	2024	Behavioral representation learning	ResNet-VAE	Neural network
Kolk et al.	2024	Image feature extraction	Convolutional neural network	Neural network
		Latent feature learning	Variational autoencoder	Neural network
		Final arrhythmia prediction	Multilayer perceptron	Neural network
Kolk et al.	2024	Ventricular arrhythmia risk prediction	Logistic Regression	Linear model
		Ventricular arrhythmia risk prediction	Deep neural network	Neural network
Zhan et al.	2025	Disease-specific adaptation	Siamese neural network (DYNA)	Neural network
Fu et al.	2025	Spatial feature learning from cine CMR	Convolutional neural network	Neural network
		Temporal dynamics modeling	Long short-term memory (LSTM)	Neural network
		Biomechanical time-series encoding	Feedforward neural network	Neural network
		Final subtype prediction	Dense classifier	Neural network
		Image classification	ResNet50	Neural network
		Feature extraction	ResNet50	Neural network
		Prediction on tabular data	Support Vector Machine	Kernel method
		Image classification	VGG16	Neural network
		Feature learning	VGG16	Neural network
Jiménez-Serrano et al.	2025	ACM classification	Logistic Regression	Linear model
Amyar et al.	2025	Radiomics pathological correlates	Hierarchical clustering	Unsupervised model
Zhang et al.	2025	Variant effect prediction	Siamese Genomic Language Model	Neural network
Carrick et al.	2025	ARVC detection	Convolutional neural network	Neural network
Quansah et al.	2026	ARVC detection	Gradient Boosted Trees	Ensemble model
		ARVC detection	Random Forest	Ensemble model
		ARVC detection	Decision Tree	Tree-based model
		ARVC detection	Naïve Bayes	Probabilistic model
		ARVC detection	TabNet	Neural network
		ARVC detection	Logistic Regression	Linear model
		ARVC detection	Lasso Logistic Regression	Linear model
		ARCC detection	Ordinary Least Squares	Linear model

Summary of studies included in this review detailing the machine learning and deep learning approaches applied to non-ischemic cardiomyopathy research. For each study, the table lists the publication year, primary task (e.g., cardiomyopathy classification, variant pathogenicity prediction, biomarker discovery, arrhythmia risk prediction, and clinical outcome modeling), specific ML/DL model(s) implemented, and corresponding model class (e.g., neural networks, ensemble methods, kernel methods, linear models, instance-based learning, graph-based models, physics-based models, and unsupervised models). The table highlights the breadth of methodological strategies, from classical supervised algorithms and radiomics pipelines to deep neural networks, multimodal architectures, genomic language models, physics-based cardiac simulations, and unsupervised clustering methods, applied to support phenotyping, mechanism discovery, and clinical prediction in NICM.

Abbreviations: ML, machine learning; DL, deep learning; NICM, non-ischemic cardiomyopathy; ECG, electrocardiogram; CMR, cardiac magnetic resonance; ARVC, arrhythmogenic right ventricular cardiomyopathy; ARCC, arrhythmogenic right ventricular cardiomyopathy with left-sided involvement; k-NN, k-nearest neighbors; LSTM, long short-term memory; SVM, support vector machine; VGG16, Visual Geometry Group 16-layer convolutional neural network; ResNet50, 50-layer residual neural network; ResNet-VAE, residual network-based variational autoencoder; DYNA, disease-specific genomic language model; ACM, arrhythmogenic cardiomyopathy; ICD, implantable cardioverter-defibrillator.

Figure 1. Mapping of data modalities to clinical tasks in machine learning applications for non-ischemic cardiomyopathies.

Illustration of the relationships between six data modalities applied in machine learning and deep learning research for non-ischemic cardiomyopathies (NICMs) and their corresponding clinical tasks. Data modalities (left) are grouped into three color-coded categories: imaging and electrophysiologic modalities (CMR imaging and ECG, green), genomic and multi-omics modalities (genomics and multi-omics, purple), and wearable and clinical data (wearables/accelerometry and clinical/EHR data, amber). Clinical tasks (right) span phenotype classification (subtype classification, arrhythmia detection), mechanism discovery (variant interpretation, biomarker discovery), and clinical decision support (ICD risk stratification, longitudinal monitoring). Solid arrows indicate primary modality-to-task mappings, while faint arrows denote secondary or multimodal contributions, reflecting the interconnected nature of data integration in NICM research.

Methods

This study was conducted as a narrative (non-systematic) review aimed at synthesizing recent advances in multi-omics and machine learning applications in non-ischemic cardiomyopathies. Relevant studies were identified through a targeted and iterative approach informed by domain expertise and supplemented by searches in academic databases, including Google Scholar and PubMed. Searches were focused on literature published between 2020 and early 2026.

Representative search terms included: “non-ischemic cardiomyopathy”, dilated cardiomyopathy”, “arrhythmogenic cardiomyopathy”, “multi-omics”, “single-cell RNA sequencing”, “proteomics”, “radiomics”, “machine learning”, and “deep learning”.

Studies were selected if they applied machine learning or deep learning methods to key thematic areas: (1) disease phenotype characterization, (2) mechanism modeling across imaging, electrophysiologic, genomic, or multi-omics, or (3) clinical prediction and decision support, including arrhythmic risk stratification and therapy guidance.

For each study, we extracted the clinical task, data modality, modeling approach, validation strategy, and reported performance. Findings were synthesized qualitatively and organized into three domains: phenotyping, mechanistic discovery, and clinical decision support, to highlight methodological trends, translational potential, and ongoing limitations related to interpretability, generalizability, and clinical implementation.

NICM Classification, Phenotyping and Mechanism Modeling Tasks

Phenotyping and classification challenges

Disease manifestation in NICM spans structural, electrophysiologic, and molecular domains, with signals distributed across imaging, ECG, genetics, and multi-omics data. Distinguishing cardiomyopathy subtypes, identifying arrhythmogenic substrates, linking imaging to tissue pathology, and connecting genotype to phenotype therefore require integration of high-dimensional, multimodal information. Conventional approaches based on handcrafted features, categorical interpretation, or single-modality analysis often fail to capture nonlinear, cross-scale relationships.

Data-driven ML models are well suited to these tasks. Radiomics-based approaches demonstrate that engineered cardiac MRI features can differentiate cardiomyopathy phenotypes, while deep-learning architectures applied to imaging and ECG data improve subtype classification, early disease detection, and electrophysiologic phenotyping. Multimodal frameworks further extend these capabilities by linking structural, functional, and molecular data. In this section, we review representative ML and DL models across imaging, electrophysiology, genomics, and multi-omics contexts, highlighting how each addresses specific limitations in NICM research.

Imaging-based phenotype classification and structural characterization

Cardiac magnetic resonance (CMR)–based modeling has become central to classifying NICM subtypes and characterizing myocardial structure, particularly as conventional visual interpretation often misses subtle morphological and tissue-level disease signals. Radiomics pipelines and deep-learning frameworks now extract high-dimensional imaging features to improve subtype discrimination while linking imaging phenotypes to myocardial biology.

Zhang et al. evaluated whether CMR-derived radiomics features could distinguish hypertrophic cardiomyopathy, dilated cardiomyopathy, and healthy controls using classical machine-learning pipelines.^[5] In a cohort of 283 patients (226 training, 57 validation), they implemented ten feature-selection strategies and nine classifiers for this three-class classification task. The final random-forest model achieved a validation accuracy of 0.912 with macro- and micro-AUCs of 0.947 and 0.934, respectively, and class-specific AUCs of 0.938 (controls), 0.966 (DCM), and 0.936 (HCM). Eight pipelines produced mean AUCs >0.910, and fivefold cross-validation yielded an average accuracy of 0.830. Performance exceeded prior methods by approximately 5 percentage points in accuracy and higher AUCs (0.947 vs. 0.924; 0.934 vs. 0.917), with sensitivity 80.4%, specificity 90.6%, precision 82.1%, and F1 score 81.1%. However, results depended on handcrafted radiomic features and segmentation quality, and were derived from a relatively small, imbalanced cohort without external validation, limiting generalizability.

Fu et al. advanced imaging-based classification by introducing a multimodal deep-learning framework integrating cine CMR with biomechanical data.^[6] Their dual-path architecture combined a Convolutional neural network and a Long short-term memory (CNN-LSTM) encoder for imaging with a Multilayer perceptron processing intraventricular pressure gradients, followed by feature fusion and classification. Trained on a multicenter cohort and externally validated, the model achieved strong performance (internal AUC 0.974; external AUC 0.962) and outperformed CNN baselines and radiomics pipelines. This work demonstrates the value of hierarchical feature learning and multimodal integration for capturing functional abnormalities beyond morphology alone, though reliance on specialized inputs and computational demands may limit scalability.

Amyar et al. examined whether CMR radiomics features reflect underlying myocardial tissue composition in non-ischemic dilated cardiomyopathy by linking imaging signatures to biopsies.^[7] The study implemented a radiomics-based machine-learning pipeline in which high-dimensional features extracted from native T1, extracellular volume, and late gadolinium enhancement sequences were reduced and modeled using logistic regression classifiers to predict histopathologic tissue characteristics. Radiomic feature clusters showed significant associations with fibrosis, inflammation, hypertrophy, vacuolization, and fat replacement, demonstrating that CMR-derived features can capture biologically meaningful remodeling processes and enabling noninvasive tissue-level phenotyping.

Taken together, these studies illustrate a progression in imaging-based NICM modeling. Radiomics approaches emphasize interpretability and biological linkage, whereas deep-learning frameworks enable hierarchical feature learning and multimodal integration with improved classification performance. There are some notable limitations to be mentioned. These include variability across imaging protocols, reliance on retrospective datasets, limited integration with clinical and genetic data, and the need to move from accurate classification toward mechanistically informative and clinically deployable models of disease progression.

ECG-based phenotyping, arrhythmia detection, and early cardiomyopathy prediction

Electrocardiography (ECG) remains a central modality for phenotyping NICMs, detecting arrhythmogenic substrates, and identifying early disease trajectories before structural changes are evident. Both classical ML and DL approaches have been applied to extract diagnostic and prognostic information from ECG waveforms, motivated by the limitations of rule-based interpretation and the need for earlier detection.

Carrick et al. developed a convolutional neural network–based ECG model to improve detection of arrhythmogenic right ventricular cardiomyopathy (ARVC), addressing the limited sensitivity and interobserver variability of the 2010 Task Force Criteria. The model demonstrated strong diagnostic performance (c-statistic 0.87 internally; 0.80 externally), with further improvement when integrated with clinical criteria, highlighting the additive value of ECG-based representation learning.^[8] However, interpretability remained limited, as predictions were driven by latent waveform features not easily mapped to established ECG markers.

Quansah et al. subsequently addressed this limitation using an interpretable gradient boosted tree model for ARVC detection, achieving strong discrimination (AUC ≈0.86–0.89 across validation cohorts).^[9] The model provided explicit feature attribution, identifying ECG predictors consistent with Task Force diagnostic markers while maintaining competitive performance. Together, these studies illustrate a shift from high-performing but opaque deep-learning models toward interpretable ML frameworks that better support clinical integration and mechanistic understanding in ARVC phenotyping.

Jiménez-Serrano et al. used engineered ECG biomarkers and logistic-regression–based classifiers to improve detection of arrhythmogenic cardiomyopathy, particularly left-dominant forms.^[10] Feature-selection pipelines identified both shared and sex-specific biomarkers, improving discrimination relative to conventional ECG criteria. Strengths include interpretability and suitability for smaller datasets; however, modest sample size and targeted cohorts limit generalizability.

Van de Leur et al. trained a deep neural network to identify disease-specific ECG signatures in carriers of phospholamban mutations, revealing subtle electrophysiologic features associated with genotype status.^[11] Lampert et al. extended this work toward early prediction, using deep learning to predict left ventricular dysfunction in patients with premature ventricular complexes from standard 12-lead ECG data linked to electronic health records.^[12] In a multicenter cohort, the model predicted LVEF decline with moderate performance (AUROC 0.79), suggesting utility for risk stratification and screening rather than standalone diagnosis.

Across these studies, classical ML models offer interpretability and robustness in smaller datasets, whereas DL approaches capture complex nonlinear waveform patterns and uncover subclinical disease signals. Performance is strongest when models are integrated with clinical or genetic data. Remaining challenges include generalizability across acquisition settings, prospective validation, and translation into clinically actionable tools.

Genomic variant interpretation and inherited cardiomyopathy risk modeling

Interpreting rare genetic variants remains a central challenge in inherited NICMs, where many variants lack clear clinical meaning. Recent work has shifted toward cardiac-specific modeling strategies incorporating gene function, disease mechanisms, and sequence context.

CardioBoost (Zhang et al.) and PathoPredictor (Evans et al.) exemplify disease-specific ML classifiers grounded in curated cardiomyopathy variant datasets.^[13,14] PathoPredictor, trained on clinical sequencing panel variants from 17,071 patients, achieved average precision >90% across disease panels and >95% on independent ClinVar evaluation sets, while reaching 95–96% accuracy for some high-variant-count genes (for example, KCNQ2). CardioBoost (AdaBoost; 76 features) was trained/tested on rare missense variants in inherited cardiomyopathies and arrhythmia syndromes and, using clinically motivated thresholds (Pr ≥0.9 pathogenic; Pr ≤0.1 benign), correctly classified 63.3% of cardiomyopathy test variants and 81.2% of arrhythmia test variants with high confidence; it achieved TPR/TNR = 69.5%/56.0% (cardiomyopathies) and 83.3%/78.6% (arrhythmias), with high-confidence classification accuracy of 90.2% and 91.9%, respectively. CardioBoost further linked variant classification to outcomes in HCM (SHaRe), with predicted disease-causing variants associated with worse event-free survival (HR 1.9 vs genotype-negative; 54% vs 33% composite outcome risk by age 60). However, reliance on curated datasets can introduce population bias and reduce coverage for rare or novel variants, and CardioBoost still assigned 29.8% of cardiomyopathy test variants to an indeterminate range (0.1 < Pr < 0.9).

In contrast, DYNA (Zhan et al.) applies representation learning using pretrained genomic language models fine-tuned for cardiac disease.^[15] In held-out rare-variant tests for cardiomyopathies (trained on 238 pathogenic/202 benign; tested on 118/100), DYNA achieved AUPR 0.910 (gene-sequence features) and the top AUC among evaluated baselines; for arrhythmias (trained on 168/158; tested on 84/79), DYNA reached AUC 0.94 and AUPR 0.95, including a 3.1% AUPR gain over the second-best method. Replication on ClinVar further improved discrimination versus the base model ESM1b, increasing AUC from 0.88→0.90 (ClinVar CM) and 0.93→0.96 (ClinVar ARM), with AUPR gains of +4% (CM) and +5% (ARM). While extending beyond curated annotations to sequence-level effects (including non-coding splicing models that achieved zero-shot AUC 0.85 on ClinVar Splicing and AUC 0.95/AUPR 0.87 after targeted fine-tuning), DYNA requires substantial computation and high-quality rare-variant training data.

Together, these approaches highlight complementary strategies: interpretable disease-specific ML models that integrate clinical genetics workflows, and DL-based sequence modeling that uncovers functional relationships beyond predefined features. Future progress will require integrating genomic predictions with imaging, electrophysiologic, and clinical data to enable patient-level risk modeling.

Transcriptomic, proteomic, and multi-omics biomarker discovery

Efforts to identify molecular biomarkers in NICM increasingly combine high-throughput omics data with ML to move beyond single-gene analyses toward systems-level disease characterization.

A study in Frontiers in Genetics by Zhang et al. applied a layered ML pipeline to RNA-seq data from two discovery cohorts (GSE230585: 5 HCM/3 controls; GSE249925: 97 HCM/23 controls) with external validation in GSE180313 (27 HCM/13 controls).^[16] Using limma (|log2FC| > 2, FDR < 0.05), they identified 271 differentially expressed genes and benchmarked 113 combinations of 12 machine-learning algorithms under 10-fold cross-validation, selecting a hybrid Lasso + StepGLM model that yielded a 12-gene diagnostic signature linked to fibrosis, inflammation, and extracellular matrix remodeling. While externally validated and biologically interpretable, reliance on retrospective public datasets, limited cohort size, and absence of formal overfitting diagnostics constrain clinical translation.

Kanapeckaite et al. extended this framework through integrative analysis of bulk RNA-seq (PRJNA477855), matched proteomics (PXD008934), and single-cell RNA-seq references, introducing an LFC_score that weights log fold change by gene–disease association metrics from Open Targets.^[17] They applied a two-step clustering approach (Gaussian mixture models followed by hierarchical clustering) incorporating expression magnitude, interaction-network degree, and association evidence, analyzing 3,521 gene scores in dilated cardiomyopathy (mean association 0.0705; max 1) versus 229 in ischemic cardiomyopathy (mean 0.00023; max 0.0196). Cross-modal overlap identified recurrent candidates such as MFAP4 across transcriptomic, proteomic, and single-cell datasets. While this strategy demonstrates structured cross-platform integration, incomplete proteome coverage and the need for experimental validation limit immediate translational impact.

Collectively, these studies establish transcriptomic and multi-omics modeling as key tools for NICM biomarker discovery, providing mechanistic insight while highlighting the need for independent validation and integration with clinical data.

ML/DL applied to Clinical Prediction and Decision Support in NICM Management

Clinical decision challenges in NICM

The clinical management of NICM requires the ability to anticipate which patients are at risk for sudden arrhythmic death, progressive heart failure, or stable disease. These outcomes differ not only in prognosis but in modifiability by interventions such as implantable cardioverter-defibrillators (ICDs). The DANISH trial highlighted this complexity: ICD therapy conferred no overall mortality benefit in unselected NICM patients because most deaths were non-arrhythmic. In this randomized trial (n=1116; 556 ICD vs 560 usual care; median follow-up 67.6 months), prophylactic ICD implantation did not reduce all-cause mortality (21.6% vs 23.4%; HR 0.87, 95% CI 0.68–1.12; P=0.28), although it significantly reduced sudden cardiac death (4.3% vs 8.2%; HR 0.50, 95% CI 0.31–0.82; P=0.005).^[18] Subsequent prespecified subgroup analysis demonstrated a linearly diminishing survival benefit with increasing age (HR per year increase 1.03; 95% CI 1.003–1.06; P=0.03), identifying an optimal cutoff of ≤70 years: ICD therapy reduced all-cause mortality in patients ≤70 years (HR 0.70, 95% CI 0.51–0.96; P=0.03) but not in those >70 years (HR 1.05, 95% CI 0.68–1.62; P=0.84).^[19] Mode-of-death analysis showed similar sudden death rates but substantially higher non sudden mortality in older patients (5.4 vs 2.7 events per 100 patient-years; P=0.01), underscoring how competing non-arrhythmic risk attenuates ICD benefit in NICM.

Despite this complexity, current frameworks rely on categorical stratification. Guidelines and risk scores cannot generate individualized, continuous risk estimates integrating multimodal data. Machine learning offers a path toward such individualized prediction, with capabilities including multimodal integration, cause-specific risk modeling, temporal prediction, and interpretability.

Multimodal AI for ICD benefit prediction

Machine learning approaches have tackled the challenge of selecting ICD candidates from two complementary perspectives: predicting which patients are at high risk for arrhythmias and will benefit from device therapy, and identifying those whose risk of mortality is primarily non-arrhythmic and unlikely to gain advantage, even if they meet guideline criteria.

The DEEP RISK ICD research program at Amsterdam UMC and collaborating institutions has developed complementary machine learning models to address these cause-specific prediction tasks through progressive multimodal data integration, representing one of the first comprehensive efforts to apply advanced computational methods to ICD decision-making in NICM populations. Initial work focused on predicting ICD non-benefit by identifying patients at high risk of non-arrhythmic mortality. In a 2023 study, investigators developed extreme gradient boosting models combining ECG time-series features with clinical variables to predict non-arrhythmic death in 1,010 primary-prevention ICD recipients with ischemic, dilated, or non-ischemic cardiomyopathy and LVEF ≤35%.^[20] The model achieved an AUROC of 0.90 during internal validation and 0.79 in external validation at a separate center, substantially outperforming the established MADIT-ICD clinical score (AUROC 0.67). ECG frequency-domain features proved particularly informative: lower-frequency Fourier coefficients predicted appropriate ICD therapy while higher-frequency components associated with non-arrhythmic mortality, suggesting distinct electrical signatures for different causes of death. Patients stratified to high non-arrhythmic mortality risk demonstrated a hazard ratio of 5.54 (95% CI 2.91–10.54) compared to low-risk groups. The model employed SHAP (SHapley Additive exPlanations) values to decompose each prediction into additive feature-level contributions, allowing clinicians to see which ECG and clinical variables increased or decreased an individual patient’s estimated risk, a critical requirement for decisions about withholding potentially life-saving therapy. However, this approach was limited by its reliance on ECG and clinical data without integration of cardiac imaging, which captures structural arrhythmic substrate.

The DEEP RISK ICD research program, at Amsterdam UMC, addressed the complementary question of arrhythmic risk prediction by incorporating late gadolinium enhancement cardiac MRI alongside ECG and clinical data.^[21] Focusing specifically on NICM patients, the model employed a variational autoencoder architecture to extract latent physiological features from both LGE-MRI and ECG data, then integrated these learned representations with clinical variables to predict malignant ventricular arrhythmia onset within one year of ICD implantation. In external validation across 103 patients, DEEP RISK achieved an AUROC of 0.84 (95% CI 0.71–0.96) with sensitivity of 0.98 and specificity of 0.73. Critically, multimodal integration outperformed single-modality approaches: LGE-MRI captured arrhythmic substrate burden including fibrosis extent and distribution patterns, while ECG provided complementary information about conduction abnormalities and repolarization dynamics that imaging alone cannot reveal. The model incorporated gradient-based explainability features to visualize anatomical regions on LGE-MRI and ECG intervals contributing to individual predictions, enabling clinicians to understand which specific features drove risk estimates. However, limitations remain: the development cohort was predominantly dilated cardiomyopathy (63.3%), and predictive performance across different NICM etiologies, in larger cohorts of cardiac resynchronization therapy patients, and among patients receiving novel heart failure medications requires further validation.

Longitudinal monitoring & wearable technology

Continuing work from the DEEP RISK ICD research program, investigators in the SafeHeart cohort used wearable accelerometry to monitor physical behavior in 277 patients with ICDs over six months, capturing 64,995 patient-days of continuous activity and sleep data.^[22] This approach enables longitudinal monitoring that captures dynamic changes and serves as a surrogate for disease progression, rather than relying on baseline measurements alone. Using logistic regression analyses, the investigators found that reduced numbers of daily physical activity bouts and low day-to-day variation in sleep characteristics were independently associated with increased risk of arrhythmic events. Critically, deep learning representations of behavioral time-series data improved predictive performance compared to simple statistical summary metrics, achieving an AUROC of 0.74±0.05 versus 0.67±0.14 for summary indices (p=0.05). This finding demonstrates that deep neural networks can extract prognostic patterns from continuous behavioral data that simple summary statistics, such as mean daily steps or average sleep duration, cannot capture.

Subsequent analysis employed deep representation learning to identify distinct behavioral phenotypes with differential arrhythmic risk in 272 patients across 37,478 days of monitoring (138±47 days per patient).^[23] Five behavioral profiles emerged with varying annual malignant ventricular arrhythmia rates: Cluster A (very low physical activity with disturbed sleep) demonstrated 30.4% annual risk and adjusted hazard ratio of 3.63 (95% CI 1.54–8.53, p<0.001) compared to low-risk profiles. Notably, high activity levels did not uniformly confer protection. Clusters B and C, despite substantially higher daily physical activity, exhibited ~17% annual event rates, while less active Clusters D and E demonstrated only 9.8% and 9.5% annual risk. This suggests that interplays between various behaviors rather than isolated activity metrics may explain differences in arrhythmic risk. Important limitations include the black-box nature of deep learning-derived latent representations, which lack direct clinical interpretability; the mixed cardiomyopathy etiologies (not exclusively NICM) with predominantly male representation (81%); uncertainty about generalizability to broader heart failure populations not meeting ICD criteria; and the absence of integrated imaging or electrocardiographic data at enrollment, which future studies should incorporate to explore interactions between structural disease, electrophysiological abnormalities, and behavioral phenotypes.

These behavioral phenotypes could inform several clinical decisions: patients in high-risk clusters may warrant more frequent remote monitoring, earlier consideration for ventricular tachycardia ablation, or adjustment of ICD programming parameters such as detection zones and therapy aggressiveness. Moreover, transitions between behavioral clusters over time, such as a patient moving from moderate-risk Cluster B to high-risk Cluster A, could trigger preemptive clinical intervention before structural deterioration or arrhythmic events occur, enabling truly dynamic risk-based management strategies that current static guideline frameworks cannot provide.

Implementation Considerations and Future Directions

Despite promising performance in development and validation cohorts, translating machine learning models for NICM risk stratification into routine clinical practice faces substantial implementation barriers.

First, interpretability remains paramount for decisions involving life-altering interventions. While current studies have incorporated gradient-based visualization and SHAP values, these explainability methods provide limited mechanistic insight. Clinicians can identify which features contributed to predictions but often cannot understand why those features confer risk or how they interact. As noted in a recent review of deep learning for ECG analysis, the gap between statistical performance and clinical interpretability remains a fundamental barrier to deployment.^[24] The SafeHeart behavioral phenotyping faces even greater challenges, as deep learning-derived latent representations lack direct clinical correspondence. For models recommending against ICD implantation in patients who subsequently experience sudden death, or recommending devices for patients who die of heart failure without arrhythmias, inability to explain reasoning erodes trust and creates medico-legal risk. Hybrid human-AI frameworks positioning model outputs as decision support rather than autonomous recommendations may mitigate these concerns.

Second, external validation and generalizability remain unmet needs. The concentration of reviewed work within a single research program, while demonstrating methodological rigor, underscores the field’s nascent state. Models trained on predominantly dilated cardiomyopathy cohorts in European tertiary centers may not generalize to different NICM etiologies, underrepresented populations, community hospitals, or healthcare systems with varying imaging access. As emphasized in recent analyses of machine learning model reproducibility in clinical cardiology, deployment environments often differ substantially from development cohorts in patient demographics, practice patterns, and data quality.^[25] Critically, none of these models have undergone prospective randomized trials comparing ML-guided selection to guideline-based care, which serves as a definitive test of whether AI improves clinical outcomes. Such trials must demonstrate not only discrimination but also appropriate calibration and ultimately reduction in inappropriate treatment decisions without increasing sudden death in untreated patients.

Third, clinical workflow integration poses practical challenges. Real-time EHR integration requires standardized data formats, automated variable extraction, seamless imaging transfer, and computational infrastructure for on-demand inference. SafeHeart’s continuous wearable monitoring raises additional questions: how frequently should risk scores update, what change magnitude warrants intervention, and who monitors continuously generated predictions? Key challenges for delivering clinical impact with artificial intelligence include technical performance, integration with existing clinical systems, physician trust and adoption, and demonstrated impact on patient-centered outcomes. Regulatory pathways for AI-based decision support remain evolving, with unclear requirements for prospective validation, post-market surveillance of performance drift, and liability frameworks when recommendations diverge from guidelines.

Future priorities include prospective multicenter registries enabling diverse training cohorts, integration of arrhythmic and non-arrhythmic models into unified platforms providing comprehensive competing-risk estimates, incorporation of longitudinal wearable data with cross-sectional assessments for dynamic risk updating, personalized ICD programming based on predicted trajectories, and cost-effectiveness analyses comparing AI-guided to guideline-based strategies. The goal is not replacing clinical judgment but augmenting it with data-driven insights enabling personalized decision-making for the heterogeneous NICM patient population.