Left Atrial Strain Model Flags Heart Risk in Uremia Patients

Patients with uremia face a sharply elevated danger of cardiovascular complications, and clinicians have long lacked reliable tools to spot who is most vulnerable before an event strikes. A new retrospective study from the Second Affiliated Hospital of Hainan Medical University in Haikou, China, reports that a machine learning model built on four-dimensional left atrial quantification, or 4D-LAQ, parameters can stratify that risk with an area under the curve of 0.840 in a held-out validation set. The work, published in Frontiers in Endocrinology, followed 223 patients with uremia between January 2021 and January 2026 and identified left atrial reservoir strain as the single most influential predictor of major adverse cardiovascular events in the cohort.

The study enrolled adults diagnosed with chronic kidney disease stage 5, defined by an estimated glomerular filtration rate below 15 mL/min/1.73 m², and excluded those with congenital heart disease, moderate-to-severe valvular disease, arrhythmias from other causes, prior kidney transplantation, or images too poor for analysis. Each participant was followed from the date of uremia diagnosis until March 2026, with major adverse cardiovascular events confirmed through electronic medical records and telephone follow-up. The composite endpoint included cardiac death, hospitalization for worsening heart failure, non-fatal stroke, and non-fatal myocardial infarction. By the end of follow-up, 94 patients, or 42.1 percent of the cohort, had experienced at least one such event. The individual components broke down as 54 hospitalizations for worsening heart failure, 26 non-fatal strokes, and 14 non-fatal myocardial infarctions, with no cardiac deaths recorded. Median follow-up reached 20 months, and no participants were lost to tracking.

Echocardiographic imaging formed the backbone of the data collection. Examinations were performed on a GE Vivid E95 ultrasound system, and dynamic images were transferred to an EchoPAC PC 203 workstation where software automatically traced the left atrial endocardium, with manual adjustment when needed. From these recordings the researchers derived twelve 4D-LAQ parameters spanning left atrial volumes at minimum, maximum, and pre-atrial-contraction phases, volume indices, emptying volumes, emptying fractions, and six strain measures covering the reservoir, conduit, and contractile phases of atrial function. A single ultrasonographer with more than five years of experience acquired all images, and two additional experienced physicians repeated the measurements three times to establish reproducibility. Intraclass correlation coefficients ranged from 0.791 to 0.970 across the parameter set, which the authors describe as satisfactory intra- and interobserver agreement.

Baseline comparisons revealed a consistent pattern of cardiac remodeling among patients who later suffered events. The MACE group showed significantly larger left atrial diameters and volumes, thicker interventricular septums and posterior walls, larger left ventricular end-diastolic diameters, and higher left ventricular mass index than the non-MACE group. Left atrial strain values told the same story from a functional angle: all six strain parameters, including reservoir strain, conduit strain, and contractile strain, were significantly reduced in those who experienced events. Median reservoir strain, for instance, stood at 13 percent in the MACE group versus 22 percent in the event-free group. B-type natriuretic peptide, a marker of wall stress and volume overload, was also markedly higher among event patients. Male sex predominated in the MACE group at 77.7 percent compared with 54.3 percent among those free of events.

To build the predictive model, the team first applied LASSO-penalized logistic regression to sift through the candidate variables. This regularization technique shrinks coefficients toward zero, effectively discarding features that add little predictive signal while guarding against multicollinearity and overfitting. Seven features survived with nonzero coefficients: height, natural-log-transformed BNP, LDL cholesterol, left ventricular mass index, left atrial reservoir strain, left atrial conduit strain, and left atrial reservoir strain corrected for body size. Pearson correlation analysis and Mantel’s test confirmed generally low correlations among the retained variables. Notably, every left atrial volume parameter was eliminated, while three strain parameters persisted, a pattern the authors interpret as evidence that strain captures left atrial dysfunction more sensitively than static volume measures.

With those seven features in hand, the researchers constructed eight candidate models using random forest, support vector machine, logistic regression, stochastic gradient descent, k-nearest neighbors, AdaBoost, naive Bayes, and multilayer perceptron algorithms. DeLong’s test, a statistical method for comparing correlated areas under receiver operating characteristic curves, guided the final selection. Logistic regression won out. AdaBoost actually posted a higher validation AUC of 0.867, but its training AUC of 0.955 signaled substantial overfitting, meaning the model had memorized idiosyncrasies of the training data rather than learning generalizable patterns. Naive Bayes performed comparably to logistic regression but rests on strict conditional independence assumptions that the authors judged unsuitable for their correlated dataset. Logistic regression, by contrast, makes no such assumptions and tends to generalize better at small and medium sample sizes.

The final LASSO-logistic model achieved an AUC of 0.890 in the training set and 0.840 in the validation set, the latter with a 95 percent confidence interval of 0.749 to 0.931. Multiple internal validation strategies reinforced the result: five-fold stratified cross-validation yielded an average AUC of 0.873, nested 5×5 cross-validation produced 0.856, and bootstrap internal validation with 1,000 resamples delivered an optimism-corrected AUC of 0.869. Calibration curves tracked the ideal diagonal closely, and the Hosmer-Lemeshow test returned a P value above 0.05, indicating no significant departure between predicted and observed event rates. Decision curve analysis showed a net benefit of up to 0.43 in the validation set, suggesting the model would add clinical value across a meaningful range of risk thresholds. The bootstrap-corrected Brier score, a measure of prediction accuracy, came in at 0.132.

The team also tested whether the 4D-LAQ parameters added anything beyond conventional information. Against a clinical-only model, which reached an AUC of 0.703 in the validation set, the full model improved discrimination by a delta AUC of 0.137. Against a conventional echocardiography model with an AUC of 0.750, the gain was 0.09. Both differences were statistically significant, supporting the claim that four-dimensional left atrial quantification carries incremental prognostic information. A nomogram was constructed from the seven selected features to translate the model into a bedside tool, and a risk threshold of 0.41, chosen by maximizing the Youden index on the training ROC curve, sorted patients into high- and low-risk categories. At that cutoff, sensitivity and specificity were 87.7 percent and 81.3 percent in the training set and 72.4 percent and 81.6 percent in the validation set.

Interpretability analysis via SHAP, or Shapley Additive Explanations, ranked the features by their marginal contribution to predictions. Left atrial reservoir strain topped the list, followed by height, left ventricular mass index, corrected reservoir strain, conduit strain, log-transformed BNP, and LDL cholesterol. The scatter distribution showed that lower reservoir strain values corresponded to higher positive SHAP contributions, meaning reduced atrial stretch capacity pushed predictions toward higher risk. The authors caution that height’s second-place ranking reflects mathematical coupling with left ventricular mass index rather than an independent biological effect, since left ventricular mass index is indexed to body size. LDL cholesterol occupied the lowest rank, and its inverse association with events, which the authors frame as a lipid paradox, likely reflects catabolic severity and protein-energy depletion rather than a protective lipid profile.

The authors acknowledge several constraints on the findings. The study was single-center and retrospective, with an events-per-variable ratio of approximately 1.71, below the traditional threshold, so external validation in larger multicenter cohorts remains necessary before clinical deployment. Data on the primary causes of kidney disease, renal replacement therapy modality, blood potassium, HbA1c, and electrocardiographic features were not systematically recorded. The composite MACE endpoint mixes subtypes that may respond to different predictors, and the model was designed to predict a first event rather than terminal outcomes such as death. Enrollment closed in January 2026, close to the unified follow-up deadline of March 2026, so patients enrolled later had shorter observation windows, a factor that may underestimate the cumulative event rate. The authors also note the model cannot guide treatment decisions and that clinical management should follow current guidelines combined with individualized assessment. Within those limits, the work offers a quantified, interpretable starting point for identifying which patients with uremia warrant intensified cardiovascular monitoring.

Subject of Research: A machine learning model using four-dimensional left atrial quantification parameters to predict major adverse cardiovascular events in patients with uremia.

Article Title: Four-dimensional left atrial quantification for predicting major adverse cardiovascular events in patients with uremia: an interpretable LASSO-logistic model

Article References: Ma, L., Jin, M., Lin, Z., Cui, R., Huang, X., & Chen, M. (2026). Four-dimensional left atrial quantification for predicting major adverse cardiovascular events in patients with uremia: an interpretable LASSO-logistic model. Frontiers in Endocrinology, 17, Article 1970519. https://doi.org/10.3389/fendo.2026.1970519

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1970519

Keywords: uremia, 4D-LAQ, left atrial strain, LASSO, logistic regression, machine learning, MACE, echocardiography, SHAP, nomogram, chronic kidney disease, cardiovascular risk