AI Reads Heart Muscle Layers to Tell Myocarditis Apart from Heart Attacks

When a young, otherwise healthy patient arrives at the emergency department with crushing chest pain, elevated cardiac enzymes, and changes on an electrocardiogram, clinicians face one of the most consequential diagnostic forks in acute cardiac care. Is this a heart attack, caused by a blocked coronary artery, or is it acute myocarditis, an inflammatory condition of the heart muscle that frequently mimics the presentation of infarction? The distinction matters enormously: the treatments diverge sharply, the prognoses differ, and misclassification can lead to inappropriate anticoagulation or delayed anti-inflammatory therapy. A new study published in BMC Medical Imaging by Yanping Su, Hongfei Hu, and colleagues from Lishui Hospital of Zhejiang University and affiliated institutions suggests that the answer may lie in how the individual microscopic layers of the heart muscle deform during each heartbeat, read by an interpretable machine-learning model built from routine cardiac magnetic resonance scans.

The research team set out to develop and evaluate a diagnostic strategy that integrates cardiac magnetic resonance (CMR)-derived myocardial strain with readily available clinical variables to separate acute myocarditis from acute myocardial infarction (AMI). Myocardial strain is a measure of how much the heart muscle deforms, shortens, or thickens during the cardiac cycle, and it can be extracted from standard cine CMR images using a technique known as tissue tracking. Unlike conventional visual assessment of wall motion, strain analysis quantifies contraction with sub-millimeter precision and, crucially, can be computed separately for the subendocardial (inner), mid-wall, and subepicardial (outer) layers of the left ventricle. This layer-specific capability is what gives the approach its diagnostic teeth, because inflammatory and ischemic injuries are known to affect the layers of the heart wall in characteristically different patterns.

The study was retrospective in design, drawing on 114 patients, of whom 52 had acute myocarditis and 62 had acute myocardial infarction. All patients underwent CMR with comprehensive left ventricular strain analysis, encompassing global, regional, and layer-specific strain parameters. The researchers were careful about how they validated their models, a point of growing importance in the medical machine-learning literature. Feature selection, model tuning, and model-size selection were all performed strictly within the training data of a nested cross-validation framework, while held-out outer folds were used exclusively for performance evaluation. This design guards against the subtle information leakage that can inflate reported accuracy when the same data are used both to build and to test a model, a pitfall that has undermined many published clinical prediction tools.

Multiple machine-learning algorithms were compared, including approaches such as random forests, support vector machines, k-nearest neighbors, and explainable boosting machines, alongside conventional logistic regression. Rather than simply chasing the highest raw performance number, the team applied a principle of parsimony: they selected the simplest model that achieved discrimination not significantly worse than the best-performing alternative. This philosophy reflects a practical reality of clinical deployment. A model with five interpretable inputs that a cardiologist can inspect, question, and reason about is far more likely to earn trust and survive real-world use than a black-box ensemble whose marginal performance gain cannot justify its opacity. The five-feature logistic regression strategy emerged as the preferred approach.

The performance figures, based on pooled predictions from the held-out outer folds, were strong for a tool of such modest complexity. The model achieved a receiver operating characteristic area under the curve (ROC-AUC) of 0.873, with a 95 percent confidence interval of 0.851 to 0.892, and a precision-recall area under the curve (PR-AUC) of 0.827, with a 95 percent confidence interval of 0.789 to 0.865. At the selected classification threshold, the model reached a sensitivity of 0.700 and a specificity of 0.868. In plain terms, the model correctly identified roughly seven in ten myocarditis cases while correctly clearing about 87 percent of heart attack patients, a balance that could meaningfully reduce unnecessary invasive procedures and misdirected treatments if validated prospectively.

Perhaps the most striking aspect of the final model is what it did not need. The classifier achieved its cross-validated discrimination without incorporating late gadolinium enhancement (LGE) information, the contrast-based CMR technique conventionally used to visualize scar and inflammation patterns in the heart muscle. LGE acquisition requires gadolinium contrast administration, adds scan time and cost, and its interpretation depends on reader expertise. By demonstrating that cine-derived strain features, captured from routine functional imaging, carry sufficient diagnostic signal on their own, the study opens the possibility of a faster, contrast-sparing diagnostic pathway, or at minimum a valuable adjunct when LGE findings are equivocal or unavailable.

The five predictors that survived feature selection form an intriguing portrait of the two diseases. The final fixed logistic regression model included patient age, end-systolic apical subendocardial circumferential strain (abbreviated ESACS_endo), apical longitudinal strain (ALS), lactate dehydrogenase (LDH), and creatine kinase-MB (CK-MB). The two strain parameters localize the diagnostic information to the apex of the left ventricle and, in the case of ESACS_endo, specifically to the innermost layer of the myocardium at that location. The two laboratory markers, LDH and CK-MB, are enzymes already measured in virtually every patient with suspected acute coronary syndrome, meaning the model requires no additional testing beyond what clinical practice already produces, combined with strain values computable from standard CMR cine loops.

Interpretability was not an afterthought but a core design goal. The researchers applied SHAP analysis, short for SHapley Additive exPlanations, a technique borrowed from cooperative game theory that assigns each feature a quantified contribution to every individual prediction. SHAP analysis demonstrated the relative contribution and direction of effect of the five predictors both across the cohort and for individual patients. This means a clinician examining a given patient’s classification can see exactly how much each strain value and each laboratory result pushed the prediction toward myocarditis or toward infarction, transforming the model’s output from an opaque probability into an auditable chain of evidence. Such transparency is increasingly viewed as a prerequisite for regulatory approval and clinical adoption of diagnostic algorithms.

The authors conclude that a parsimonious and interpretable model integrating cine CMR-derived myocardial strain with routinely available clinical variables showed good cross-validated discrimination between acute myocarditis and acute myocardial infarction, and that layer-specific and regional strain provided complementary diagnostic information beyond conventional clinical variables. They position the approach as a noninvasive adjunct for differentiating the two conditions, while candidly noting that further validation in independent cohorts is warranted before the tool enters clinical practice. The study was approved by the Institutional Review Board of The Fifth Affiliated Hospital of Wenzhou Medical University, with the informed consent requirement waived due to the retrospective use of de-identified data, and the work was supported by the National Key Research and Development Program of China and several Zhejiang provincial and municipal funding programs.

For the field of cardiovascular imaging, the study adds to a growing body of evidence that the heart’s mechanical fingerprint, measured layer by layer, encodes diagnostic information that visual inspection and conventional markers miss. Acute myocarditis remains a diagnostic challenge precisely because it so often masquerades as infarction, and the gold standard of endomyocardial biopsy is invasive and insensitive. If the findings hold up in larger, multicenter, prospective cohorts, a five-variable model computable from scans and blood tests already ordered in the emergency department could become a practical triage tool, helping clinicians decide which patients truly need the catheterization laboratory and which need anti-inflammatory care instead. The work also exemplifies a broader and welcome trend in medical artificial intelligence: smaller, explainable models built on mechanistically meaningful features, rather than sprawling black boxes, may be the fastest route from algorithm to bedside.

Subject of Research: Using CMR-derived layer-specific myocardial strain and explainable machine learning to differentiate acute myocarditis from acute myocardial infarction

Article Title: CMR-derived layer-specific strain to differentiate acute myocarditis from acute myocardial infarction with explainable machine learning

Article References: Su, Y., Hu, H., Jiang, L., Ye, R., Du, Z., Shen, D., Xia, S., Chen, M., Lu, C., & Ji, J. (2026). CMR-derived layer-specific strain to differentiate acute myocarditis from acute myocardial infarction with explainable machine learning. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02809-y

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02809-y

Keywords: cardiac magnetic resonance, myocardial strain, acute myocarditis, myocardial infarction, machine learning, SHAP, logistic regression, diagnosis, cardiology, medical imaging, predictive markers, nested cross-validation