Translating Mechanistic Insights Into Action and Revealing New Pathways: Machine Learning Approaches in Heart Failure With Preserved Ejection Fraction

Abstract

Heart failure with preserved ejection fraction is a prevalent condition that carries a high morbidity and mortality, with limited treatment options. Obtaining and integrating critical mechanistic insights are essential to development of optimal treatment strategies. Machine learning (ML) has played a transformative role in advancing clinical research and can be employed to understanding myocardial dysfunction in heart failure with preserved ejection fraction. ML techniques, including supervised, unsupervised, and reinforcement learning, can be applied to cardiac imaging to identify phenotypes and extract biomarkers. Mechanistic evaluation in heart failure with preserved ejection fraction integrating advanced imaging and ML can provide information on myocardial stiffness, steatosis, and energetics. Feature extraction and feature learning techniques build upon deep convolutional neural networks, and clustering algorithms can automate detection of myocardial fibrosis, energetics, and other mechanisms. Multimodal ML frameworks, such as multifidelity physics‐informed neural networks, can offer deeper insights into mechanisms, improving phenotype clustering and patient‐specific interventions. This review addresses how integrating ML approaches with advanced imaging can address traditional challenges and advance precision medicine for heart failure with preserved ejection fraction, guiding targeted therapies.

Nonstandard Abbreviations and Acronyms

³¹P‐MRS

phosphorus‐31 magnetic resonance spectroscopy

DTI

diffusion tensor imaging

ECV

extracellular volume

HFpEF

heart failure with preserved ejection fraction

ML

machine learning

MPINN

multifidelity physics‐informed neural network

PINN

physics‐informed neural network

Approximately 6 million individuals in the United States are diagnosed with heart failure (HF), with half of them developing HF with preserved ejection fraction (HFpEF). ¹ HFpEF is characterized by a left ventricular (LV) EF of ≥50%. Diagnosing HFpEF poses a challenge because the underlying pathophysiologic mechanisms of dysfunction are not well understood, and available treatment options are limited. HFpEF is a critical public health problem of rising prevalence that is associated with impaired functional tolerance, quality of life, and high morbidity and mortality. ² , ³ It is recognized as the greatest unmet need in cardiovascular medicine today according to the National Heart, Lung, and Blood Institute Working Group. ² It is a heterogeneous syndrome—a multiorgan, systemic disorder, that comprises multiple pathophysiologic abnormalities beyond LV diastolic dysfunction and is linked to the epidemics of obesity, hypertension, and diabetes. After HF hospitalization, the 5‐year survival of patients with HFpEF is a mere 35%, which portends a worse prognosis than many cancers. ² , ³ Quality of life is also diminished in HFpEF, leading to substantial physical limitations.

Therefore, there is a critical need to gain a better understanding of the underlying pathophysiologic mechanisms in HFpEF in order to identify novel therapeutic targets and to personalize therapy at an individual level. With advanced technology and deep phenotyping, we now have access to several data streams that can help with this. These include traditional clinical assessments, wearable monitoring technologies, multimodality imaging, laboratory data, and multiomic blood‐based biomarkers (eg proteomics, metabolomics, lipidomics). However, with these opportunities come unique challenges—integrating and analyzing these data to make meaningful inferences regarding underlying pathophysiology that can be used to target treatment in addition to diagnosing the condition and predicting its course.

Although traditional analytical methods can be used, given the heterogeneous nature of HFpEF, machine learning (ML) is emerging as a powerful tool that can integrate multimodal data and detect nuanced disease patterns to guide research and treatment. ML approaches can be leveraged to integrate mechanistic biomarkers, thereby paving the way for a precision medicine approach (Figure 1). In this review, we focus on available and potential ML approaches to evaluate the underlying mechanisms of dysfunction in HFpEF. We begin with an overview of supervised and unsupervised ML approaches, the application of deep learning applied in advanced cardiac imaging, and their role in subphenotyping HFpEF and predicting prognosis. Next, we examine mechanistic insights at the chamber level, focusing on myocardial dysfunction in HFpEF. We highlight how advanced imaging techniques, combined with ML, can help uncover tissue‐level energetic and mechanistic dysfunction such as myocardial fibrosis, fiber orientation, energetics, and steatosis. We then address multimodal integration, exploring how ML can combine different types and levels of data, along with challenges such as dimensionality reduction, and scarce clinical imaging data, and strategies to mitigate these limitations. Finally, we discuss adaptive ML approaches, including transfer learning, and continual learning, which enable models to incorporate multiple time points, continually learn and adapt to new information—capabilities that can be particularly useful in medical applications.

Figure 1. Machine learning approaches for integrating pathophysiologic mechanisms in HFpEF to advance precision medicine.

Left, key mechanisms of cardiac dysfunction in HFpEF; Middle, highlighting how machine learning can integrate these diverse data sources; and Right, multimodal integration with the goal of advancing precision medicine. HFpEF indicates heart failure with preserved ejection fraction.

MACHINE LEARNING APPROACHES IN MECHANISTIC EVALUATION OF HFpEF

In recent years, ML techniques, which focus on teaching computers to learn from data and make predictions, have become promising tools in clinical cardiovascular applications. ML has the potential to guide precision medicine approaches in HFpEF in particular due to heterogeneity of the condition and challenges in diagnosis and management. ML techniques can facilitate earlier diagnosis, enable subphenotyping of patients, predict response to specific therapies and risk of clinical outcomes such as hospitalization, and enable discovery of new biomarkers and therapeutic drugs. ⁴ There are 2 major types of ML—supervised and unsupervised. In supervised learning, outcomes or categories are defined a priori and used to develop models to predict future events; whereas in unsupervised learning, unlabeled data are used to understand patterns observed within the data. Deep learning is a subfield of ML that involves training deep neural networks (NNs) to learn and make predictions. ⁵ It uses multiple layers to extract complex features from raw inputs, as exemplified by convolutional NNs. Unlike traditional ML techniques, deep learning can enable interpretation and processing of high‐dimensional raw signal data, including images, which makes it particularly well suited to use with advanced imaging applications such as echocardiography and cardiac magnetic resonance (CMR) imaging. ⁴ , ⁶ Rather than relying on measured or annotated quantitative measurements, deep learning can use raw data (feature engineering) to identify underlying patterns. Feature engineering refers to the process of selecting variables or characteristics from raw data (such as images) to improve a ML model’s performance.

ML approaches offer several advantages as compared with traditional statistical techniques, especially with regards to large data sets and complex relationships in HFpEF. These include the ability to (1) analyze nonlinear relationships that are inherent in complex biological systems; (2) make accurate predictions along with uncertainty quantification; (3) include various tasks such as classification, regression, clustering, and reinforcement learning; (4) process large data sets for big data applications; (5) select relevant features (variables) in a data set; and (6) transfer learning from one data set to another, or even adapt to changes in the data set over time. ⁴ , ⁵ In contrast, traditional data sets often require a priori assumptions that may not apply in every setting, may prioritize parameter estimation (which may not translate to improved model performance), have computational challenges with large data sets, and may require prior expert knowledge in selecting important features in the data. However, the ultimate choice between ML or traditional statistical techniques depends on the context, data set features, and overall goals. In many contexts, a combination of both methods may provide the most effective strategy. In the following sections, we explore specific applications of ML in uncovering mechanistic insights at the chamber and tissue levels in HFpEF.

MECHANISTIC INSIGHTS AT THE CHAMBER LEVEL: ASSESSMENT OF DIASTOLIC DYSFUNCTION

The left ventricle’s efficiency relies on its capacity to alternate between 2 cycles: a relaxed chamber during diastole that permits filling and a firm chamber during systole that expels the stroke volume. Diastolic function is linked to the myocardial relaxation and passive properties of the left ventricle, which are affected by myocardial tone. Increased filling pressures are the main outcome of diastolic dysfunction, which can be evaluated using echocardiography. ⁷ The echocardiography‐based measures include diastolic filling velocities (E and A), and early diastolic myocardial velocity (e’), which are used to calculate E/A ratio, and E/e’ ratio used in the algorithm for assessing diastolic function that is associated with HF. ⁷ The E/e’ ratio in particular has been found to predict future HF events. ⁸ Diastolic dysfunction is prevalent in HFpEF and contemporary assessment of diastolic dysfunction is sometimes challenging and requires integration of multiple echocardiographic parameters. Several studies have used ML techniques to enhance assessment of diastolic dysfunction, which can aid clinical risk stratification. ⁹ ML techniques in classification and regression tree analysis have been applied to integrate atrial longitudinal strain with guideline‐recommended diastolic variables, thereby improving diastolic dysfunction classification and prognostic stratification in patients with HF, with better event prediction compared with traditional guideline‐based methods. ¹⁰ Using a limited data set of echocardiographic measurements of systolic and diastolic function, Pandey et al. employed an unsupervised clustering approach followed by a supervised classifier‐based deep learning model to determine the severity of diastolic dysfunction and identify patients with differential risks of adverse events and responses to treatment with spironolactone. ¹¹ Another study used least absolute shrinkage and selection operator regression and random forest to classify LV diastolic dysfunction with a focus on identifying second‐ and third‐degree diastolic dysfunction from single‐channel ECG. It achieved a sensitivity and specificity of 86% and 85%, respectively, indicating potential for ECG as a screening tool for diastolic dysfunction. ¹² Although available studies provide examples of how diastolic function can be identified using available tools, they fall short of identifying whether in a particular patient with HFpEF, presence of diastolic dysfunction is the mechanism of failure. Doing so would require using the clinical, imaging, and biochemical data to get a wholistic model of prediction and explainability analyses to identify a set of features contributing to diastolic dysfunction or HFpEF in a particular patient or groups of patients, as discussed in multimodal assessment.

MECHANISTIC INSIGHTS AT THE TISSUE LEVEL

In this section, we delve into ML approaches for mechanistic evaluation of HFpEF at the tissue‐specific level, focusing on myocardial stiffness, fiber orientation, fibrosis, steatosis, and energetics. We also highlight case studies demonstrating ML‐driven methods to provide insight into these mechanisms.

Evaluating Myocardial Stiffness

Myocardial stiffness is an important intrinsic property that affects both systolic and diastolic function of the heart. The level of myocardial stiffness indicates the degree of resistance of the tissue to deformation, which is influenced by the intracellular components of the cardiomyocyte (primarily the cytoskeleton) as well as extracellular components such as collagen fibers. ¹³ , ¹⁴ There are 2 main interrelated components of myocardial stiffness: chamber (structural) stiffness and intrinsic myocardial elasticity. Chamber stiffness results from myocardial material properties and ventricular chamber characteristics such as muscle mass and geometry. ¹⁴ It is quantified as the change in ventricular pressure relative to the change in ventricular volume, represented by the slope of the end‐diastolic pressure–volume relationship, known as the chamber stiffness constant, ß. ¹⁴ Intrinsic myocardial elasticity is the correlation between stress (σ, force per unit area) and strain (ε, segment length relative to reference length) locally applied to a small volume of myocardium. ¹⁴ Changes in myocardial stiffness are commonly associated with cardiac diseases and can be an important diagnostic tool in HFpEF. In healthy individuals, myocardial stiffness is typically in the range of 0.015 to 0.040 mm Hg/mL. ¹⁵

Challenges in Traditional Assessments

Assessment of myocardial stiffness in vivo is a complex task, requiring specialized techniques with limitations. Methods include ultrasound acoustic radiation force impulse imaging, ultrasound shear wave elastography, magnetic resonance elastography, pressure‐volume loops (which are the gold standard but invasive), and 3‐dimensional (3D) finite element models. ¹⁴ This 3D finite element modeling enables patient specific assessment of myocardial stiffness (shear stiffness modulus, μ) by integrating LV geometry and myocardial mechanics (Figure 2). ¹⁶ It uses data from diffusion tensor imaging (DTI)‐CMR advanced imaging technique and strain assessment to account for myofiber orientation. Key advantages include high spatial resolution, dynamic heart simulation and personalized data from CMR. Despite its prognostic value in HFpEF, 3D finite element modeling can be computationally intensive. ¹⁷ ML models using LV end‐diastolic pressure–volume relationship data offer a less time‐intensive alternative but still rely on invasive measurements. ¹⁸ Ultimate selection of the method chosen depends on the research‐specific goals and practical constraints.

Figure 2. Three‐dimensional finite element model reconstruction to estimate myocardial stiffness using cardiac magnetic resonance imaging data.

FE indicates finite element; MRI, magnetic resonance imaging

ML Case Study

Babaei et al. developed a ML model to predict passive myocardial stiffness by analyzing geometric, architectural, and functional features from cardiac imaging and patient demographics (Figure 3). ¹⁸ This addressed limitations of traditional techniques (such as finite element modeling) that can be computationally intensive. This model, trained on 2500 synthetic rodent heart data sets, used a multilayer feed forward NN and achieved high accuracy in predicting passive myocardial stiffness (R² of 96% with LV volume and endocardial area identified as important features). The model was validated using ex vivo rat myocardial stiffness measurements and patient‐specific finite element modeling, with both methods having excellent agreement with the ML model. This study demonstrates the potential of ML in improving evaluation of myocardial stiffness, which is otherwise technically challenging using traditional methods. It links organ level end‐diastolic pressure–volume relationship metrics to intrinsic myocardial tissue properties, providing patient specific prediction of myocardial stiffness, which can be particularly useful is conditions such as HFpEF.

Figure 3. Estimation of myocardial stiffness using a machine learning approach.

A, 3‐dimensional reconstruction of human cardiac anatomy at end‐diastole based on CMR images. This geometric information is then used in simulations to estimate mechanical behavior (motion) as in (B), which visualizes the displacement of the cardiac tissue. These data are then used by the ML model to predict myocardial stiffness. Figure adapted from A machine learning model to estimate myocardial stiffness from EDPVR by Babaei et al. ¹⁸ , published in Scientific Reports under a CC BY license. © [2022]. CMR indicates cardiac magnetic resonance; and ML, machine learning.

In addition to global myocardial stiffness, myocardial fiber orientation, which is associated with stiffness, plays a crucial role in cardiac function, impacting both systolic and diastolic function.

Myocardial Fiber Orientation

Myocardial fiber orientation, which is important in the regulation of systolic and diastolic function, is altered in HF in preclinical studies. ¹⁹ Cardiac fibers are normally arranged in layers of helices around the ventricles with a normal LV fiber orientation angle of 60 degrees at the endocardium to −60 degrees at the epicardium, which corresponds to an average slope of 1 to 1.2 degrees/% wall thickness. ²⁰ The helical arrangement of cardiac myofibers is important because it is responsible for equalizing myofiber strain and maximizing the EF in a normal heart. LV geometry and fiber structure are remodeled in several disease states, which contributes to cardiac dysfunction and impacts contractile ejection. Fiber orientation correlates with infarct size and LV systolic function after a myocardial infarction. ²⁰ Patients with HF have been found to have a shift in fiber orientation angle to a more oblique angle. ²⁰ Novel CMR sequences such as DTI allow for characterization of myocardial fiber orientation. Preclinical studies in small animal models showed a distinct correlation between myocardial fiber architecture and cardiac function, and it was determined that DTI‐CMR was capable of characterizing this association. ²⁰ However, assessment of myocardial fiber orientation (DTI) in humans had been extremely difficult in the past and therefore was rarely performed in vivo due to sensitivity of the technique to bulk motion and higher resting heart rates in humans. ²⁰ Prior studies using DTI‐CMR were performed in vitro or using small animal research magnetic resonance imaging (MRI) scanners. A recent technical advance that compensates for bulk motion and reduces sensitivity to velocity and acceleration has overcome this limitation and demonstrated feasibility of DTI‐CMR in human volunteers and patients with HF on a 3T clinical MRI scanner. ²¹ An example of DTI‐CMR short‐axis scans in 2 patients along with the 3D fiber orientation visualization is shown (Figure 4).

Figure 4. Diffusion tensor imaging scan and fiber angle visualization.

An example of a cardiac magnetic resonance diffusion tensor imaging short‐axis slice from a control (top left) and patient with heart failure with preserved ejection fraction (bottom left) with corresponding 3‐dimensional fiber orientation angle visualization (right). HFpEF indicates heart failure with preserved ejection fraction.

ML Case Study

Herrera et al. developed a method called Fibernet, which uses physics informed NNs (PINNs) to determine myocardial fiber orientation and cardiac conduction velocity tensor of the human atria from multiple electroanatomical maps. ²² PINNs are a class of ML models that can incorporate physical laws into the NN training to solve complex scientific and engineering problems. ²³ This model includes imaging data from CMR, synthetic 2‐dimensional/3D stimulations, and patient specific electroanatomical maps. ²² Although DTI‐CMR provides high resolution myofiber orientation data, it can have limited accessibility in clinical settings. Fibernet uses electroanatomical maps that may be more easily accessible and then uses PINNs to estimate myofiber orientation. Accurate assessment of myofiber orientation can improve understanding of the mechanical and electrical behavior of the heart in HFpEF. Thus, the Fibernet method may be a potential alternative to the evaluation of myofiber orientation, compared with traditional imaging techniques that have limited accessibility.

The association between myofiber orientation and cardiac function highlights the complexity of underlying HFpEF pathophysiology. Another key contributor to myocardial stiffness and dysfunction is myocardial fibrosis, which plays a vital role in disease progression and treatment response.

Myocardial Fibrosis

Prior studies have demonstrated that fibrosis, defined as excess deposition of extracellular matrix (including collagen accumulation and cross‐linking), is present to varying degrees in patients with HFpEF (as assessed using T1 mapping on CMR and autopsies) and is associated with a worse prognosis. ² Biomarkers of fibrosis have been found to correlate directly with lack of treatment response with spironolactone in HFpEF, implying that the presence of fibrosis may modulate the efficacy of certain treatments. ² Despite these observations, drugs targeting fibrotic pathways (eg, angiotensin‐converting enzyme inhibitors, angiotensin receptor blockers, etc) have not shown substantial benefit in clinical trials. This may be because myocardial fibrosis is present at varying degrees in patients with HFpEF and variably contributing to the pathophysiology at an individual level. Myocardial fibrosis is a vital component of myocardial stiffness and can be assessed noninvasively using CMR. It can be quantified using parametric mapping techniques on CMR (T1 mapping before and after gadolinium injection to calculate extracellular volume [ECV]) as part of advanced tissue characterization. ²⁴ , ²⁵ The ECV percentage is the proportion of myocardial tissue that represents the extracellular space and is derived from T1 mapping before and after the administration of gadolinium‐based contrast. The typical range for normal ECV values is 25% to 30%, with values >30% considered abnormal extracellular expansion due to fibrosis, edema, infiltration or other causes. ²⁶ , ²⁷

ML Case Studies

Although conventional CMR requires expert interpretation, convolutional NNs and deep learning models can enhance quantitative fibrosis analysis and segment‐specific fibrosis assessment. Using late gadolinium enhancement images, convolutional NNs can identify regions of the myocardium that are fibrotic, with abnormal extracellular matrix accumulations. ²⁸ Deep learning‐based feature extraction methods can extract differences in signal intensities between tissues to differentiate healthy myocardium from fibrotic tissue. Quantitative parametric mapping techniques, such as T1 mapping and ECV assessment, which delineate fibrotic areas, can also be automated using deep learning. ²⁹ By evaluating patterns across multiple cardiac slices, these techniques can classify stages of fibrosis, allowing for earlier diagnosis or monitoring of progression. ²⁹

Haidiri et al. developed a deep learning semiautomated tool that uses convolutional NNs to identify and localize myocardial fibrosis in CMR in a 17‐segment model. ³⁰ This framework had up to 0.86 accuracy in binary classification of fibrosis across myocardial segments, closely matching radiologist assessment. Traditional fibrosis assessment relies on ECV mapping using T1‐weighted MRI pre‐ and post gadolinium contrast, which requires manual or semiautomated segmentation and can be subject to some variability in interpretation. In contrast, this framework offers significant improvement upon the traditional manual and semiautomated method of fibrosis detection by performing fully automated segmentation with high accuracy in a faster time period, and reduced subjectivity.

Another example of using ML approaches to assess fibrosis is the development of FibrosisNet, a deep learning network that can automate myocardial fibrosis detection and classification from CMR. ²⁸ This model was built with 17 layered series to improve fibrosis detection, and deep transfer learning was applied to convolutional NNs, achieving a 96% accuracy and F1 score and enabling early disease detection. ²⁸ Compared with conventional methods using T1 mapping (which require manual or semiautomated segmentation subject to variability in interpretation), FibrosisNet improves the ability to detect and localize fibrosis with higher accuracy.

In parallel to the myocardial structural changes, metabolism‐related changes such as intramyocardial lipid accumulation and myocardial energetics also affect cardiac structure and function that may provide valuable input in phenotyping HFpEF.

Myocardial Steatosis

Hydrogen magnetic resonance spectroscopy (MRS) is a noninvasive technique that can quantify triglyceride content in the heart in vivo. Using hydrogen MRS, it has been shown that triglyceride droplets in the cytosol resonate at 1.4 ppm, triglycerides in the adipocytes resonate at 1.6ppm, and tissue water resonates at 4.8 ppm. ³¹ Clinical studies have demonstrated that lipid accumulation in the heart is associated with morphological changes and diastolic dysfunction. Myocardial triglyceride content is accompanied by increased LV mass in otherwise healthy individuals. ³² Furthermore, individuals with metabolic syndrome (constellation of metabolic abnormalities including hyperlipidemia, hypertension, insulin resistance and obesity) have been found to have greater myocardial lipid deposits and impaired myocardial performance. ³² Using hydrogen MRS, patients with HFpEF have been found to have 2‐fold increase in myocardial triglyceride content, which is associated with impaired diastolic strain rate. ³³ Triglyceride accumulation has also been found to contribute to diastolic dysfunction in women with coronary microvascular disease. ³⁴ Taken together, these data show that intramyocardial fat deposition is associated with LV hypertrophy and impaired diastolic function in humans. The deposition of lipids is associated with impaired myocyte metabolism, fatty acid uptake, and oxidative phosphorylation in the mitochondria, thereby also affecting myocardial energetics. To the best of our knowledge, ML has not been used to infer and assess myocardial steatosis in the heart using clinical, imaging, and blood‐based biomarkers and remains a promising area of development.

Myocardial Energetics

In the heart, adenosine triphosphate (ATP) generation occurs through the CK (creatine kinase) system, which catalyzes this reversible reaction:

$$\text{Phosphocreatine}+\mathrm{ADP}+\mathrm{H}+\leftrightarrow \text{creatine}+\mathrm{ATP}$$

The majority of ATP in the heart is produced by oxidative phosphorylation in the mitochondria using fatty acids as substrates. To adjust to variations in the workload, the heart uses the CK reserve system. ³⁵ The CK enzyme mediates the transfer of a phosphoryl group from ATP to phosphocreatine, which carries it to the cytosol, where the phosphoryl group is transferred back to ATP by the cytosolic CK. It can then be used by contractile proteins and the sarcoplasmic reticulum to maintain calcium homeostasis and for all processes requiring energy. ³⁶ The relative concentration of phosphocreatine to ATP is a marker of the myocardium’s ability to convert substrate into ATP for active processes and represents the myocardial energetic state. It is known that impaired mitochondrial function and altered energy metabolism results in an energy deficit and contributes to the severity of HF. ³⁵ Thus, homeostasis of myocardial energy metabolism is important and pharmacological targeting of the energy metabolic pathways has emerged as a novel therapeutic approach that can improve cardiac function in the failing heart.

Phosphorus‐31 MRS (³¹P‐MRS) is a powerful noninvasive and nonionizing modality that is uniquely able to provide an assessment of the myocardial energetic state in vivo, which has not been possible via other techniques. Although traditional biochemical analysis can provide the concentration of ³¹P metabolites, it does not provide information about metabolic changes. ³⁷ This traditional approach also involves performing an invasive tissue biopsy and due to the inherent instability of high‐energy phosphates that often breakdown rapidly, in particular phosphocreatine, makes these measurements very challenging. ³⁸ In contrast, ³¹P‐MRS allows for quantification of the myocardial phosphocreatine to ATP concentration ratio, which correlates with the cytosolic Gibbs free energy of ATP hydrolysis and represents the energy available to the cardiomyocytes to perform work. ³⁸ The typical normal values for phosphocreatine/ATP range from 1.6 to 1.9 in the literature (mean 1.7 + − 0.3). Many clinical studies cite a phosphocreatine/ATP ratio of <1.5 to 1.6 as reduced myocardial energetics. However, the precise normal values may vary depending on the MRI magnet field strength, the localization method, pulse sequence, the T1 and blood ATP corrections, and blood or skeletal‐muscle contamination. Thus, comparisons should be made across studies that have used similar methods. ³⁹ , ⁴⁰

Using ³¹P‐MRS, studies have shown decreased myocardial energetics in diabetic cardiomyopathy, inherited cardiomyopathies, and valvular heart disease. In one study, patients with HFpEF had reduced phosphocreatine/ATP (1.60±0.09 versus 2.00±0.10, P=0.005) compared with controls, which correlated with significantly reduced diastolic strain rate and maximal oxygen consumption. ³³ A recent randomized trial incorporated phosphocreatine/ATP as evaluated by ³¹P‐MRS as the main outcome. ⁴¹

ML Case Studies

In a study examining gene expression profiles from mice cardiac tissue, feature selection was used to identify key metabolism‐related genes that are potential diagnostic biomarkers of HFpEF and random forest, a supervised learning algorithm was employed to rank the importance of these genes. ⁴²

Han et al. used metabolomics, RNA sequencing and ML techniques to examine myocardial metabolism differences in HF. ⁴³ ML based agnostic clustering of myocardial metabolites revealed distinct metabolic profiles among these groups, with the HFpEF myocardium being characterized by reduced fatty acid oxidation, altered amino acid metabolism, and impaired use of alternative fuel sources. ⁴³ Specifically, in HFpEF, the myocardial metabolism was found to have impaired fatty acid oxidation (lower tissue long‐chain acylcarnities), and elevated amino acids (elevated branched‐chain amino acids with potential contribution to insulin resistance and altered mTOR [mammalian target of rapamycin] signaling). Additionally, there was dysregulation in the nitric oxide and nitrogen signaling pathways, which can affect vascular and cardiac function. ML integration of multimodal metabolomic data has identified subtle patterns linking specific metabolites such as asparagine, 2‐hydroxyglutarate with cardiac remodeling and structural abnormalities. It also identified altered markers of fibrosis (hydroxyproline), reduced second messengers (cyclic adenosine monophosphate and (cyclic guanosine monophosphate)) and lower levels of membrane associated lipids (ie, sphingomyelins), which also indicates abnormalities in myocardial signaling and structure. ML approaches allow for pattern recognition across hundreds of metabolites and can capture subtle and multipathway abnormalities, which are challenging to discern with traditional statistical methods. However, there is a lack of studies integrating and using hydrogen and ³¹P‐MRS data as a ground truth to train and refine ML models, presenting a promising area of future research.

INTEGRATION OF MULTIMODAL DATA USING MACHINE LEARNING

The preceding sections highlight the use of ML in specific case scenarios related to refining diagnosis, risk stratification, and prognosis. Increasingly, we are getting several multimodality data from patient encounters, imaging, and laboratory assessments. This provides a unique opportunity to use ML where traditional statistical approaches may fall short.

The availability of sufficient sample size to train NN architectures. remains a major challenge. This is especially true when the amount of data and number of variables far exceed the number of patients. As we deep phenotype patients with multimodal assessments, the inherent limitation is the ability to perform these comprehensive measurements on a large cohort due to cost and logistical challenges such as wide‐spread availability of advanced equipment and expertise. In such contexts, the ability to learn effectively from multimodal data derived from a limited sample size becomes crucial. To tackle this issue, the emerging field of scientific ML has made substantial contributions. One such innovation in this domain is the PINN. ²³ This approach integrates prior knowledge of physics or conservation laws with data to regularize the network. It represents a hybrid methodology that combines physics with machine learning. Physics‐informed and domain‐knowledge‐guided models enhance interpretability by aligning predictions with established principles, while reducing overfitting to noisy or sparse data. In clinical applications like HFpEF, they provide accurate and explainable insights, enabling better decision‐making. These hybrid models are more interpretable, less susceptible to overfitting, and require fewer training data. ⁴⁴ Multifidelity PINNs (MPINNs) are a type of physics informed NNs that incorporate mathematical models to solve complex multiscale problems. ⁴⁴ , ⁴⁵ They are particularly useful in cases in which traditional models are limited due to sparse or noisy data as they can incorporate multiple levels of data.

In the context of HFpEF, an MPINN can be structured as part of a system of interconnected subnetworks, with each network tailored to handle different types of data. For example, one network may be trained on approximating low‐fidelity data such as demographics, comorbidities, laboratory biomarkers, and vital signs. This network would include the population‐level data that are more readily available in clinical settings but lack physiological detail. A second network that connects the low‐ and high‐fidelity domains (these data with imaging data), would integrate intermediate representations learned from both types of data. This would allow for cross‐modal learning, integrating the clinical features with image‐derived parameters. Finally, a third subnetwork would focus on high‐fidelity data, incorporating detailed quantitative information from echocardiography and cardiac magnetic resonance imaging. This subnetwork would capture detailed structural and functional characteristics of the myocardium that reflect the underlying pathologic dysfunction. Taken together, these subnetworks using MPINNs combine mechanistic understanding with high‐fidelity data, enabling more robust modeling of disease mechanisms. A schematic representation of this multimodel MPINN approach and its data flow is shown in Figure 5. This process results in a multifidelity approximation, which enhances accuracy and efficiency of the model. Importantly, it can identify both linear and nonlinear relationships between the low‐fidelity and high‐fidelity data. ⁴⁵ MPINNs have the potential to reduce the high experimental cost for collecting high‐fidelity data (in this case CMR data) and can be employed for high‐dimensional problems with varying complexity of data. ⁴⁵

Figure 5. Multifidelity physics‐informed neural networks incorporating demographic, clinical, echocardiographic, and magnetic resonance imaging variables.

Layer 1 (left) depicts demographic and clinical variables; layers 2 and 3 (right) illustrate integration of advanced imaging variables. CMR indicates cardiac magnetic resonance.

ML Case Study

Mehdi et al. demonstrated a similar multifidelity model to quantify infarct regions in the left ventricle of patients with myocardial infarction. By combining rodent‐based in‐silico data with limited human data, the model accurately predicted areas on infarction similar to late gadolinium enhanced CMR results, highlighting the potential of multifidelity approaches to reduce the clinical data typically required by data‐intensive NNs. ⁴⁶ Thus, MPINNs also have the potential to reduce the need for collecting advanced imaging or invasive data and would be a promising approach to analyze multimodality data from patients with HFpEF.

DEEP OPERATOR NETWORK AND MULTIFIDELITY MODELING FRAMEWORKS

Another framework that is pertinent to study complex biological entities is the deep operator network, which is effective at incorporating data that span a wide range of spatial–temporal scales, and has the capability to uncover new operators when trained using multifidelity data. ⁴⁷ Although these frameworks offer powerful solutions for integrating diverse clinical and imaging data, their implementation comes with inherent challenges. As these ML models grow more complex and data intensive, it is becoming increasingly important to address issues such as overfitting, underfitting, dimensionality reduction, and bias to ensure accuracy and generalizability, especially for heterogeneous conditions such as HFpEF. In the next section, we discuss these challenges and strategies to mitigate them.

POTENTIAL CHALLENGES

Although the potential applications of ML in HFpEF are promising, several challenges remain. These include bias, confounding, overfitting, and lack of external validation. ⁴ A model may incorporate noise inherent in a particular data set and fit the data set well, leading to overfitting and lower generalizability. Conversely, a model may be too simplistic, leading to low variance. Thus, the goal is to minimize both bias and variance.

To prevent overfitting, a validation data set or development data set should be used during the model building process. In addition, the model should be validated in an external data set and rigorously developed and tested before implementation in health care settings. ⁴ , ⁵ Dimensionality reduction can further reduce model complexity and the potential for overfitting while enhancing interpretability, which is especially relevant in clinical data sets with large numbers of variables. This can be achieved through feature selection, which preserves data structure, or feature extraction, which uses linear (eg, principal component analysis) or nonlinear transformations (eg, autoencoders). ⁴⁸ Regularization techniques such as least absolute shrinkage and selection operator or ridge regression also help prevent models from fitting training data too closely. ⁵ In addition, particularly for clinical and imaging studies, addressing data or class imbalance is essential requiring rebalancing strategies such as resampling techniques, cost‐sensitive learning (in which higher misclassification costs are assigned to the minority class during training), and ensemble methods (which combine multiple models) to ensure accurate representation and identification of minority classes. ⁴⁹ In addition to preventing overfitting and reducing bias, the interpretability of ML models in medical applications is crucial to ensure that the clinicians can understand the reasoning behind model predictions. The ideal models would be transparent and provide insights into model decision making while allowing for biologically plausible patterns. ⁵⁰

Adaptive ML approaches, such as transfer learning and continual learning offer innovative solutions to overcome some of these challenges by using pretrained models and continuous model updates.

Transfer learning is an approach in deep learning in which a pretrained model on a large data set is used as a starting point for a new model. This is particularly useful when working with smaller sample sizes or limited data. Using transfer learning, the predefined model’s features can be transferred and it can then be optimized to the specific task and training on the new data set. ⁵¹ From a translational perspective, transfer learning may also enable the extrapolation of findings from studies conducted on murine samples to enhance understanding and predictions relevant to human health. In the TALE‐HFpEF (Transfer Learning for Echocardiographic Detection of Heart Failure With Preserved Ejection Fraction) study, investigators used a transfer learning approach by using a pretrained model (ResNet) that had learned to extract important spatial and temporal features including LV and right ventricular EFs from a large medical imaging data set and then fine‐tuned it to detect HFpEF from 4‐chamber echocardiographic videos, achieving high diagnostic accuracy (area under the curve of 0.95 and F1 score 0.93). ⁵²

Continual learning, also known as lifelong learning, enables models to learn and adapt to new data while retaining previously acquired knowledge. This can be particularly valuable in clinical applications such as in patients with HFpEF, where patient data evolve over time with changes in demographics, comorbidities, and treatments or with imaging modalities that incorporate multiple spatial and temporal data. By incorporating continual learning, models can dynamically update and refine their predictions, ensuring sustained accuracy and relevance in clinical decision‐making. ⁵³ In the CoReEcho (Continuous Representation Learning for 2D + Time Echocardiography Analysis) study, researchers developed a framework for continuous representations directly tailored for LVEF regression from multiple echocardiogram clips, resulting in improved performance. ⁵⁴

Explainable artificial intelligence further enhances transparency by clarifying why models make specific decisions. For example, researchers developed a model integrating clinical data with epicardial adipose tissue volume derived from CMR for diagnosis of HFpEF and prediction of outcomes. ⁵⁵ Using explainable artificial intelligence, they were able to determine which features were most significant in driving the classification. They found that the top features were epicardial adipose tissue volume, right atrioventricular groove measurements, tricuspid regurgitation velocity, and presence of metabolic syndrome. ⁵⁵

Thus far, most ML applications in HFpEF have focused on optimizing screening or identifying known pathophysiologic features, but preclinical literature is increasingly using multiomics approaches to uncover key signaling mechanisms and gene states associated with disease development and prognosis. Integrating multimodal preclinical data sets such as deep phenomics, histopathology, single‐cell or nuclear RNA sequencing, metabolomics, proteomics, and phosphoproteomics with clinical data, which are more heterogeneous and limited by tissue accessibility, could enable comprehensive mechanistic models at the patient level informed by explainability analyses. A key limitation to advancing this vision is the lack of standardized approaches for performing and reporting such measurements across studies, particularly in preclinical research. As funding agencies and journals promote open data repositories and common data elements, the hope is that curated, standardized data sets will support future ML models that extend beyond diagnosis and prognosis to provide insights into cell‐ and tissue‐specific molecular pathways and ultimately contribute to the development of novel and personalized therapies. These may include one targeting antifibrotic pathways with priferidone and sacubitril/valsartan; metabolism and steatosis with metformin and SGLT2 (sodium glucose transporter 2) inhibitors; or exploring use of calcium sensitizing agents, which could affect cardiac metabolism, and myocardial relaxation.

CONCLUSIONS AND FUTURE DIRECTIONS

The complex nature of HFpEF poses challenges for noninvasive diagnosis and management. However, an integrated approach including biomarkers and noninvasive advanced imaging techniques may enhance understanding of the pathophysiologic dysfunction in HFpEF. In addition to using advanced imaging biomarkers to determine underlying mechanisms such as myocardial fibrosis and energetics, multiomics approaches integrating genomics, metabolomics, and proteomics can provide a comprehensive understanding of underlying disease mechanisms in HFpEF. These methods can identify dysregulated pathways, such as inflammation, vascular abnormalities, and metabolic derangements, which are important in the pathophysiology of HFpEF and can inform targeted therapies. ² , ⁵⁶ By linking molecular pathways with imaging phenotypes, these approaches would enable more precise patient stratification. This approach can facilitate precise phenotyping, integrating the clinical, molecular and imaging data to determine HFpEF subphenotypes and tailor personalized treatments. ⁵⁷ Recent advancements can deepen our understanding of mechanistic dysfunction, refine risk stratification, and enhance therapeutic guidance in patients with HFpEF. Given the heterogeneous nature of HFpEF, ML can be employed to integrate mechanistic biomarkers at an individual level, leading the path towards a precision medicine approach and targeted therapies.

Sources of Funding

Research reported in this publication was supported by National Institutes of Health R01HL148727 (Gaurav Choudhary), R01HL169456 (Peng Zhang, Gaurav Choudhary), P30GM149398 (Gaurav Choudhary), R01HL168368 (Reza Avazmohammadi), R56HL172052 (Reza Avazmohammadi) and Veterans Affairs Veterans Health Administration I01BX006287 (Gaurav Choudhary). The views expressed in this article are those of the authors and do not necessarily reflect the position or policy of the Department of Veterans Affairs or the US government.

Disclosure

None.