In a retrospective cohort study, researchers demonstrated that an electrocardiogram-based machine-learning model can detect left atrial appendage thrombus in patients with first-diagnosed non-valvular atrial fibrillation. The optimized logistic regression tool achieved an area under the receiver operating characteristic curve of 0.803 in internal validation, providing an interpretable, non-invasive method to evaluate thrombotic risk.

Prompt identification of left atrial appendage thrombus represents a critical clinical objective for stroke prevention in non-valvular atrial fibrillation.

The investigation evaluated a retrospective cohort of 486 patients with first-diagnosed non-valvular atrial fibrillation who underwent standard 12-lead electrocardiography within two days prior to receiving transesophageal echocardiography or cardiac computed tomography. Clinical evaluation confirmed the presence of left atrial appendage thrombus in 87 of the 486 participants, representing a baseline prevalence of 17.9 percent.

To establish a predictive pipeline from standard electrocardiographic data, the researchers utilized Lasso regression to screen variables, isolating eight predictive electrocardiogram features. Using these selected features, the team trained and evaluated five distinct machine-learning models to determine their capacity to predict left atrial appendage thrombus.

Among the evaluated algorithms, the logistic regression model demonstrated optimal diagnostic performance, recording a sensitivity of 0.77 and a specificity of 0.81 alongside its area under the curve of 0.803, with a 95 percent confidence interval ranging from 0.750 to 0.855. Model calibration proved robust according to Hosmer–Lemeshow goodness-of-fit testing, which yielded a chi-square value of 7.295 with eight degrees of freedom and a P value of 0.505. Furthermore, decision curve analysis revealed a net clinical benefit across decision threshold probabilities spanning 4 to 75 percent.

The electrocardiogram model significantly surpassed the conventional CHA2DS2–VASc scoring system, which registered an area under the curve of only 0.570 in this cohort. Reclassification metrics indicated marked enhancement, with the model delivering a continuous net reclassification improvement of 0.9228 and an integrated discrimination improvement of 0.1882 over the established score. Incorporating CHA2DS2–VASc variables into the electrocardiographic model produced no improvement in predictive performance.

To support real-time clinical evaluation, the investigators deployed the final model as both a nomogram and a freely accessible interactive web calculator. The resulting interface aims to facilitate point-of-care thrombotic risk assessment for individuals presenting with first-diagnosed non-valvular atrial fibrillation.

The authors reported that the tool provides rapid risk stratification capable of helping clinicians prioritize patients who warrant expedited definitive imaging such as transesophageal echocardiography. Nonetheless, the researchers emphasized that the findings derive from internal validation, noting that establishing true clinical utility and defining any appropriate role in guiding anticoagulation decisions will require prospective external validation.

Subject of Research: Machine learning prediction of left atrial appendage thrombus from electrocardiograms in non-valvular atrial fibrillation

Article Title: A noninvasive ECG-based machine-learning tool for early detection of left atrial appendage thrombus in first-diagnosed atrial fibrillation

Article References: Liao, P., Jiang, X., Tang, C., Chen, Y., Kong, H., Deng, F., Li, H., Wang, L., & Zhang, Z. (2026). A noninvasive ECG-based machine-learning tool for early detection of left atrial appendage thrombus in first-diagnosed atrial fibrillation. European Journal of Medical Research. https://doi.org/10.1186/s40001-026-05442-1

Image Credits: AI Generated

DOI: 10.1186/s40001-026-05442-1

Keywords: atrial fibrillation, non-valvular atrial fibrillation, left atrial appendage thrombus, electrocardiogram, 12-lead ECG, machine learning, logistic regression, Lasso regression, transesophageal echocardiography, cardiac CT, thrombotic risk, stroke prevention, CHA2DS2-VASc score, decision curve analysis, net reclassification improvement, integrated discrimination improvement, nomogram