HEART: Dynamic Cardiac Studies & Integrated Frameworks

• HEART is a term that describes both the biological heart as a dynamic, 4D, multimodal organ and a set of acronym-based frameworks used in diverse technical domains.
• Recent research in HEART includes conditional 4D anatomical modeling, precise segmentation with high Dice scores, and end-to-end mesh recovery for clinical intervention.
• The field emphasizes spatiotemporal coherence, multimodal alignment, and structured incompleteness to advance insights in cardiac imaging, electromechanics, and acoustic diagnostics.

Searching arXiv for the provided HEART-related papers to ground citations. In contemporary arXiv literature, “HEART” denotes both the biological heart and a proliferating family of acronymized methods, benchmarks, and systems. The biological heart appears as a multiscale object of study spanning 4D anatomy, electromechanics, hemodynamics, heart sounds, and image-guided intervention, while uppercase “HEART” and related spellings such as “HeaRT” and “CHeart” name domain-specific frameworks in cardiac imaging, language modeling, diffusion control, clinical NLP, and environmental forecasting (Qiao et al., 2023). This dual usage makes the term unusually broad: it can refer to the organ itself, to representations of the organ across time, or to technical systems whose names reuse the acronym independently of cardiology (Iyer et al., 9 Jan 2026).

1. Cardiac anatomy as a conditional and spatiotemporal object

A major contemporary line of research treats the heart not as a static structure but as a 4D anatomical process over the cardiac cycle. “CHeart: A Conditional Spatio-Temporal Generative Model for Cardiac Anatomy” models cardiac anatomy as sequences of 3D segmentation volumes with four labels—background, left ventricular cavity, myocardium, and right ventricular cavity—over $T=20$ time frames, aligned to a common template space and cropped to $128\times128\times64$ (Qiao et al., 2023). Clinical factors $c$ include age, gender, weight, height, and systolic blood pressure; these are embedded by an MLP into a condition latent vector $z_c\in\mathbb{R}^{32}$, while the end-diastolic anatomy is encoded into $z_0\in\mathbb{R}^{32}$ and propagated through a one-layer LSTMCell with shared weights across time (Qiao et al., 2023).

The resulting formulation supports two distinct tasks. In anatomical sequence completion, the model imputes missing frames $x_{1:T-1}$ from $x_0$ and $c$; in anatomical sequence generation, it synthesizes the full 4D anatomy from $c$ alone. On sequence completion, the reported average Dice is $0.874\pm0.024$, with Hausdorff distance $128\times128\times64$0 mm and ASSD $128\times128\times64$1 mm; per-structure Dice reaches $128\times128\times64$2 for LV, $128\times128\times64$3 for myocardium, and $128\times128\times64$4 for RV (Qiao et al., 2023). On sequence generation, the model produces synthetic 4D anatomies whose clinical distributions are close to real data, with age-conditioned KL divergences such as $128\times128\times64$5 for LVEDV and $128\times128\times64$6 for LVM (Qiao et al., 2023).

Whole-heart representation learning extends this idea from conditional generation to self-supervised latent modeling. “Whole Heart 3D+T Representation Learning Through Sparse 2D Cardiac MR Images” learns a compact latent space from routine short-axis and long-axis cine CMR by masking $128\times128\times64$7 of spatiotemporal patches of size $128\times128\times64$8 and reconstructing them with a 6-layer encoder and 2-layer decoder, pretrained on 14,000 unlabeled UK Biobank cases and evaluated on 1,000 annotated cases (Zhang et al., 2024). The learned representation improves phenotype prediction, for example reporting LVM MAE $128\times128\times64$9 g and RVEF MAE $c$0, and remains stable when 1–2 planes are missing, with cosine similarity approximately $c$1 and prediction shifts of $c$2 g for LVM and $c$3 for RVEF (Zhang et al., 2024). This suggests that, in current CMR research, “heart” increasingly denotes a latent 3D+T representation shared across views rather than a single image plane.

End-to-end mesh recovery pushes the same idea toward intervention. TetHeart reconstructs full 4D multi-structure meshes for myocardium, LV, and RV from either full cine stacks or sparse slices using deep deformable tetrahedra, slice-adaptive 2D–3D feature assembly, and weak supervision from ED and ES annotations alone (Chen et al., 15 Sep 2025). In the sparse-slice setting, even a single slice yields clinically oriented estimates such as LVESV MAE $c$4 ml, LVEF MAE $c$5, RVESV MAE $c$6 ml, and RVEF MAE $c$7 on M&Ms, while full-stack input improves these to $c$8 ml, $c$9, $z_c\in\mathbb{R}^{32}$0 ml, and $z_c\in\mathbb{R}^{32}$1 respectively (Chen et al., 15 Sep 2025).

2. Segmentation, view acquisition, and structural refinement

Another dominant usage of heart in the literature concerns the delineation of cardiac structures from imaging. In coronary CT angiography, “Whole Heart Anatomical Refinement from CCTA using Extrapolation and Parcellation” begins from a 6-label segmentation—LV, LV myocardium, RV, LA, RA, and aorta—and refines it to a 10-label map by adding pulmonary artery and subdividing the left atrium into LA body, left pulmonary veins, right pulmonary veins, and left atrial appendage (Xu et al., 2021). The method uses two label-to-label U-Nets: one for extrapolation to add PA and one for parcellation to split the LA. Manual correction was required for 80 extrapolation cases and 50 parcellation cases, compared with 260 cases for the initial labels; in the final 10-label image-to-label model, all original six labels achieved median Dice above 95%, with LA body improving from approximately 91% to approximately 97% and RV from approximately 92% to approximately 96% (Xu et al., 2021).

Long-axis cine CMR has its own segmentation challenges. “Transforming Heart Chamber Imaging: Self-Supervised Learning for Whole Heart Reconstruction and Segmentation” proposes 2D and 3D two-stage self-supervised hybrid transformer–CNN architectures for 4CH whole-heart segmentation, targeting LV cavity, LV myocardium, RV cavity, LA, RA, and, when visible, the aortic root (Qayyum et al., 2024). The reported 4CH performance reaches average Dice $z_c\in\mathbb{R}^{32}$2 with average Hausdorff distance $z_c\in\mathbb{R}^{32}$3 mm, compared with nnUNet at $z_c\in\mathbb{R}^{32}$4 and $z_c\in\mathbb{R}^{32}$5 mm; in SAX, the same framework reports average Dice $z_c\in\mathbb{R}^{32}$6 and average HD $z_c\in\mathbb{R}^{32}$7 mm (Qayyum et al., 2024). The reconstruction component then uses a label-completion 3D U-Net trained on 1,700 dense whole-heart CCTA segmentations to infer dense volumetric labels from sparse long-axis and short-axis inputs (Qayyum et al., 2024).

At the acquisition stage, the heart is also operationalized as an imaging target whose quality must be scored in real time. “Follow Your Heart: Landmark-Guided Transducer Pose Scoring for Point-of-Care Echocardiography” focuses on apical four-chamber transthoracic echocardiography and uses a dense sweep protocol with 430 sweeps and 88,524 frames from nine individuals to classify frames as green, yellow, or red according to a point-deduction rubric (Guo et al., 28 Mar 2026). An uncertainty-aware landmark detector predicts 47 landmark channels, while the pose-scoring module achieves mean test accuracy up to $z_c\in\mathbb{R}^{32}$8 on subject-level 5-fold cross-validation; the accompanying LVEF model, applied only to green clips, produced $z_c\in\mathbb{R}^{32}$9 versus anesthesiologist visual ground truth $z_0\in\mathbb{R}^{32}$0 in the reported POC cohort (Guo et al., 28 Mar 2026). The standard reference formula used throughout this context is

$z_0\in\mathbb{R}^{32}$1

Segmentation robustness is itself a research topic. “Adversarial Heart Attack: Neural Networks Fooled to Segment Heart Symbols in Chest X-Ray Images” shows that white-box PGD can force a chest X-ray segmentation network to predict the heart as a heart-shaped symbol, achieving target-overlap IoU$z_0\in\mathbb{R}^{32}$2 up to $z_0\in\mathbb{R}^{32}$3 for a small heart and $z_0\in\mathbb{R}^{32}$4 for a large heart, while targeted black-box attacks are substantially less effective (Bortsova et al., 2021). The study therefore documents both a vulnerability and a limit: white-box cardiac segmentation can be manipulated into anatomically implausible shapes, but transferability remains weak.

3. Electromechanics, hemodynamics, and latent physiological state

At the mechanistic end of the literature, the heart is treated as a coupled electrical, mechanical, and fluid dynamic system. “A comprehensive and biophysically detailed computational model of the whole human heart electromechanics” presents a four-chamber whole-heart model comprising anatomically accurate geometry, comprehensive myocardial fiber architecture, a biophysically detailed microscale active-force model, a 0D closed-loop circulation, and chamber-specific constitutive laws (Fedele et al., 2022). The model reproduces healthy function across all chambers, including atrial eight-shaped pressure–volume loops and ventricular indices within normal ranges; example values include LV EF approximately $z_0\in\mathbb{R}^{32}$5 and RV EF approximately $z_0\in\mathbb{R}^{32}$6, with semilunar peak flows around $z_0\in\mathbb{R}^{32}$7 mL/s (Fedele et al., 2022). Ablations show that removing atrial contraction suppresses the atrial A-loop and booster pump function, while removing fiber-stretch-rate feedback produces unphysiologically large aortic and pulmonary peak flows.

A more explicitly fluid-structural formulation appears in “Electromechanical human heart modeling for predicting endocardial heart motion,” which combines a realistic biventricular geometry, rule-based fibers, monodomain electrophysiology, voltage-dependent active stress, a two-way ALE FSI with actual 3D blood meshes, and a closed-loop 0D systemic–pulmonary circulation (Hasani et al., 4 Sep 2025). Validation against cine MRI feature tracking shows consistent rank-order agreement in RV regional displacement, with the largest motion in the basal and mid free walls and the smallest at the apex (Hasani et al., 4 Sep 2025). A practical implication stated in the paper is that the RV basal and mid free walls are ideal for implanting motion-driven energy harvesting devices (Hasani et al., 4 Sep 2025).

Physiology is also modeled at the level of hidden state estimation rather than full mechanics. “Bayesian at heart: Towards autonomic outflow estimation via generative state-space modelling of heart rate dynamics” replaces instantaneous heart rate point estimates with a latent stochastic process $z_0\in\mathbb{R}^{32}$8 that drives observed beat counts through a Poisson observation model and a Gamma Markov chain prior (Rosas et al., 2023). On a tilt-table dataset of 10 healthy subjects, the Bayesian formulation preserves linear HRV properties while improving nonlinear dynamical discrimination; for example, permutation entropy showed a self stand-up estimate of $z_0\in\mathbb{R}^{32}$9 with $x_{1:T-1}$0, and synthetic validation via IPFM yielded subject-wise Spearman $x_{1:T-1}$1 with $x_{1:T-1}$2 (Rosas et al., 2023). In this usage, the heart is not directly imaged at all; it is inferred as a hidden autonomic state.

A separate motion-focused line concerns energy harvesting. “Harnessing cardiac power: heart kinetic motion analysis for energy harvesters” reports in-vivo porcine epicardial measurements at nine sites using a Doppler laser displacement meter sampled at 2000 Hz (Khazaee et al., 2024). The right atrium showed the highest cardiac kinetic movement with amplitude $x_{1:T-1}$3 mm displacement and $x_{1:T-1}$4 m/s$x_{1:T-1}$5 acceleration; summing translational kinetic energy across locations yielded approximately $x_{1:T-1}$6 mJ per beat and mechanical power approximately $x_{1:T-1}$7 W (Khazaee et al., 2024). The simulated piezoelectric output was highest at the right atrium and RV outflow tract, reinforcing the location dependence of usable cardiac motion (Khazaee et al., 2024).

4. Heart sounds, ECG–PCG coupling, and automated screening

In acoustic diagnostics, heart refers to the organ as heard rather than imaged. “Heart Abnormality Detection from Heart Sound Signals using MFCC Feature and Dual Stream Attention Based Network” uses phonocardiogram cycles segmented into S1, systole, S2, and diastole, with fixed length 2500 samples at 1000 Hz and features from both raw waveform and MFCC representations (Rashid et al., 2022). The model combines a 1D CNN stream, a GRU stream with 128 units, and a feature-gating attention module over the concatenated 128-dimensional embedding. On PhysioNet/CinC 2016, patient-level averaged results were accuracy $x_{1:T-1}$8, sensitivity $x_{1:T-1}$9, specificity $x_0$0, and MACC $x_0$1 (Rashid et al., 2022). Ablations show that MFCCs improve the recurrent stream and that attention outperforms simple concatenation.

Synchronized multimodal screening broadens this setting. “A Novel Transfer Learning-Based Approach for Screening Pre-existing Heart Diseases Using Synchronized ECG Signals and Heart Sounds” aligns ECG and PCG windows by applying identical segment boundaries to simultaneous recordings from PhysioNet/CinC 2016 training-a, yielding 3,496 synchronized ECG–PCG samples from 407 recordings (Hettiarachchi et al., 2021). Modality-specific CNN feature extractors are pretrained separately on ECG and PCG corpora, then fused in a dual-CNN. In the transfer-learning, imbalanced, record-wise setting, the reported performance is sensitivity $x_0$2, specificity $x_0$3, accuracy $x_0$4, AUC $x_0$5, and G-Mean $x_0$6; in a no-transfer-learning, high-sensitivity setting, sensitivity rises to $x_0$7 at specificity $x_0$8 (Hettiarachchi et al., 2021). The paper’s core premise is that ECG supplies rhythm and conduction timing while PCG supplies mechanical events such as S1, S2, and murmurs.

“HeartFit: An Accurate Platform for Heart Murmur Diagnosis Utilizing Deep Learning” presents a deployment-oriented formulation in which a custom-designed stethoscope and mobile application upload audio to a database for classification by a deep recurrent convolutional neural network (Gupta et al., 2019). The abstract reports training on 300 prelabeled heartbeat audio samples, validation on 100 previously unseen samples, an $x_0$9 beta score of $c$0, and accuracy of $c$1 percent (Gupta et al., 2019). The corresponding detailed source is unavailable in the supplied material, so no further methodological claims can be made beyond those abstract-level figures.

5. Cross-domain reuse of “HEART” as an acronym

Uppercase HEART is also a recurrent acronym outside cardiology. In clinical NLP, “HeaRT: Health Record Timeliner to visualise patients’ medical history from health record text” transforms unstructured Japanese EHR text into Gantt-like timelines by combining JaMIE-based entity and relation extraction with temporal clustering and topological ordering (Yada et al., 2023). The reported NER F1 is approximately $c$2 to $c$3, RE F1 approximately $c$4 to $c$5, and physician-reviewed timeline placement accuracy reached $c$6 onset and $c$7 duration in a case report, with comparable robustness across independently written radiology reports sharing only $c$8 of word bigrams (Yada et al., 2023).

In evaluation of supportive dialogue, “HEART: A Unified Benchmark for Assessing Humans and LLMs in Emotional Support Dialogue” defines five axes—Human Alignment, Empathic Responsiveness, Attunement, Resonance, and Task-Following—over 300 multi-turn dialogue histories derived from ESConv and adversarial transformations (Iyer et al., 9 Jan 2026). Pairwise comparisons are aggregated with a Bradley–Terry model,

$c$9

and converted to Elo ratings (Iyer et al., 9 Jan 2026). The benchmark reports human–LLM agreement of $c$0, close to inter-human agreement of $c$1, and finds that several frontier models approach or surpass average human responses on regular dialogues while humans retain advantages in adversarial turns (Iyer et al., 9 Jan 2026).

In diffusion-based image editing, “HEART: Hyperspherical Embedding Alignment via Kent-Representation Traversal in Diffusion Models” argues that text encoder embeddings lie on a hypersphere and that concepts are better modeled by Kent distributions than by Euclidean offsets (Roy et al., 8 May 2026). The method is training-free, operates directly on token embeddings, and reports, for subject swap, HEART-inv with Acc $c$2, CLIP-T $c$3, and LPIPS $c$4, while HEART-A performs attribute control on SDXL in 32 s with approximately 0 GB additional memory and no training (Roy et al., 8 May 2026). Here the acronym has no relation to the biological heart.

Two additional language-model uses reinforce this pattern. “HEART: Emotionally-driven test-time scaling of LLMs” iterates over emotionally charged prompts based on Ekman’s six emotions, reporting strong oracle-guided improvements such as Gemini 2.5 Flash on HLE at $c$5 versus Self-Reflection at $c$6 and Wait at $c$7 (Pinto et al., 26 Sep 2025). “A HEART for the environment: Transformer-Based Spatiotemporal Modeling for Air Quality Prediction” defines HEART as “Hybrid Enhanced Autoregressive Transformer,” a pre-attention block for llull-environment, and reports MSE reductions of up to $c$8 and an average of $c$9 across tested cities and pollutants (Bodendorfer, 26 Feb 2025). The acronym is therefore best understood as an overloaded technical label whose meaning is fully domain-dependent.

6. Conceptual unity and recurring research themes

Despite the acronymic heterogeneity, several themes recur across HEART-related research. The first is spatiotemporal coherence: CHeart models a clinically conditioned 4D anatomical trajectory (Qiao et al., 2023), TetHeart recovers coherent 4D meshes from sparse cine slices (Chen et al., 15 Sep 2025), and electromechanical models link activation, stress, deformation, and circulation over the full cycle (Fedele et al., 2022). The second is structured incompleteness: representation learning from sparse 2D cine CMR tolerates missing planes with minimal latent drift (Zhang et al., 2024), label completion reconstructs dense whole-heart volumes from sparse views (Qayyum et al., 2024), and point-of-care TTE guidance scores whether an A4CH view is on target, close, or far from target without external trackers (Guo et al., 28 Mar 2026).

A third theme is multimodal alignment. In cardiology this alignment couples ECG with PCG (Hettiarachchi et al., 2021), image-derived landmarks with pose quality (Guo et al., 28 Mar 2026), or electrical activation with 3D blood flow and tissue deformation (Hasani et al., 4 Sep 2025). In non-cardiac acronymic uses, it couples dialogue histories with rubric-guided evaluation (Iyer et al., 9 Jan 2026), text embeddings with hyperspherical geometry (Roy et al., 8 May 2026), or pollutant histories with attention-preprocessed spatiotemporal context (Bodendorfer, 26 Feb 2025). This suggests that the persistence of the term “HEART” is not merely nominal: many of these systems are organized around central integration problems in which multiple signals, scales, or modalities must be brought into a coherent latent or physical state.

A common misconception is that research labeled HEART necessarily concerns cardiology. The current literature does not support that assumption. Some HEART papers are explicitly cardiac and organ-centered, such as whole-heart segmentation, motion, electromechanics, and auscultation (Xu et al., 2021); others use the acronym in emotional support dialogue, diffusion-model control, environmental forecasting, or clinical text timeline extraction (Iyer et al., 9 Jan 2026). Conversely, not all heart-organ research is titled HEART; many cardiac papers instead use variants such as CHeart, TetHeart, HeartFit, or “Bayesian at heart” (Rosas et al., 2023). The term therefore functions less as a stable ontology than as a convergence point between organ-level cardiology and acronym design in technical research.

In present usage, then, HEART names an unusually broad research space. At one pole it refers to the human heart as a dynamic, imageable, deformable, and audible organ. At the other, it serves as a reusable acronym for systems that prioritize integration, guidance, or evaluation across heterogeneous signals. The arXiv record shows both meanings coexisting, often within highly technical frameworks, and together they define the contemporary research vocabulary of HEART.