Echocardiography-based multimodal models for thromboembolic risk prediction in atrial fibrillation: a narrative review

Abstract

Background and Objective

The congestive heart failure, hypertension, age ≥75 years, diabetes, stroke, vascular disease, age 65–74 years, sex category (CHA₂DS₂-VASc) score is fundamental for assessing thromboembolic risk in atrial fibrillation (AF). However, its limitations arise from relying solely on clinical parameters, without considering individual cardiac structural and functional changes. This review aims to explore the development and implementation of multimodal risk prediction models that integrate echocardiographic parameters with clinical data and biomarkers to achieve a more precise and individualized risk assessment.

Methods

A comprehensive literature search was conducted using relevant electronic databases (e.g., PubMed, Web of Science and Embase) to identify studies focusing on echocardiographic parameters, AF, and thromboembolic risk prediction models published up to May 2026.

Key Content and Findings

Echocardiographic parameters evaluating the left atrium and left atrial appendage (LAA; size, function, and morphology) have emerged as significant predictors of both prevalent thrombus formation and future thromboembolic events. This review details these key parameters and critically assesses various multimodal approaches, including those combining clinical-echocardiographic scores with artificial intelligence (AI) techniques. Importantly, these models demonstrate enhanced predictive accuracy for specific clinical endpoints over traditional clinical scores.

Conclusions

Utilizing a multimodal approach provides a promising supplementary tool for refined risk assessment, rather than replacing guideline-based anticoagulation decisions. Future directions should focus on standardizing echocardiographic measurements, expanding the use of AI for automated analysis, and conducting prospective interventional trials to confirm clinical benefit.

Introduction

Atrial fibrillation (AF) is the most common cardiac arrhythmia, posing a significant global public health challenge due to its strong association with elevated risks of stroke and other thromboembolic events (1,2). In accordance with contemporary clinical guidelines, the evidence, clinical trials, and predictive models discussed herein apply exclusively to patients with AF, explicitly excluding those with moderate-to-severe rheumatic mitral stenosis or mechanical prosthetic heart valves (3). Thrombogenesis in AF primarily involves complex pathological changes such as structural remodeling of the left atrium and left atrial appendage (LAA), hemodynamic alterations, and a prothrombotic state. These factors collectively contribute to blood stasis and thrombus formation, most often occurring within the LAA (4-7).

The congestive heart failure, hypertension, age ≥75 years, diabetes, stroke, vascular disease, age 65–74 years, sex category (CHA₂DS₂-VASc) score is crucial for clinical thromboembolic risk stratification in AF patients (8,9). However, its exclusive reliance on clinical parameters restricts its ability to account for individual variations in cardiac structure and function. This may lead to the misclassification of risk, especially in patients with low scores but significant atrial remodeling, implying high actual risk (5,10-12). This inadequacy highlights the need for more refined risk assessment tools.

Echocardiography, as a non-invasive and widely available imaging technique, offers critical insights into the determinants of thrombus formation (13,14). Quantifying parameters of left atrial/LAA size, morphology, and mechanical function consistently demonstrates strong predictive value for prevalent thrombosis, providing a direct view into the pathophysiology of atrial cardiomyopathy (5,10,11,15). Integrating these echocardiographic parameters with conventional clinical risk factors into multimodal predictive models shows promise in enhancing risk prediction accuracy for both prevalent left atrial appendage thrombus (LAAT)/spontaneous echo contrast (SEC) and future thromboembolic events, facilitating more individualized patient management (16,17) (see Figure 1).

Figure 1.

Schematic of multimodal data integration for thrombotic risk prediction. This diagram illustrates the integration of diverse data sources through a multimodal analysis engine to generate personalized risk assessments and clinical decision support for AF patients. The model combines traditional statistical approaches with advanced AI techniques to provide comprehensive risk stratification. AF, atrial fibrillation; AI, artificial intelligence; CHA₂DS₂-VASc, congestive heart failure, hypertension, age ≥75 years, diabetes, stroke, vascular disease, age 65–74 years, sex category; CLOTS-AF, creatinine, left ventricular ejection fraction, overload, tricuspid annular plane systolic excursion, stroke, and AF rhythm; CT, computed tomography; ECG, electrocardiogram; LA, left atrial; LAA, left atrial appendage; LAAeV, LAA emptying velocity; LAD, left atrial diameter; LAT, left atrial thrombus; LAVI, left atrial volume index; LVEF, left ventricular ejection fraction; LVMI, left ventricular mass index; LVMW, left ventricular myocardial work; MRI, magnetic resonance imaging; NT-proBNP, N-terminal pro-B-type natriuretic peptide.

This review summarizes current evidence and recent advancements in developing multimodal models that incorporate echocardiographic parameters for comprehensive risk prediction in AF. It presents the key echocardiographic parameters foundational to these models, evaluates the construction, validation, and clinical application of various multimodal approaches, explicitly distinguishing between models designed for procedural TEE avoidance and those predicting future stroke, and discusses future directions for improving risk stratification in this field. We present this article in accordance with the Narrative Review reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1700/rc).

Methods

This article is conducted as a narrative review. A comprehensive literature search was conducted using electronic databases (PubMed, Web of Science, and Embase) to identify relevant studies addressing thromboembolic risk prediction and echocardiographic parameters in AF. Given the narrative nature of this review, a formal systematic quality assessment tool was not utilized; however, the included studies were critically evaluated for methodological soundness and clinical relevance by the reviewing authors. The detailed search strategy, including the exact search date, full search terms, strict inclusion and exclusion criteria, and the independent screening process, is summarized in Table 1.

Table 1. The search strategy summary.

Items	Specification
Date of search	May, 2026
Databases and other sources searched	PubMed, Web of Science and Embase; reference lists of retrieved articles were also manually searched for relevant studies
Search terms used	(“atrial fibrillation” OR “AF”) AND (“echocardiography” OR “left atrial appendage” OR “Left Atrial Strain”) AND (“thromboembolism” OR “stroke” OR “risk prediction”)
Timeframe	January 1999 to May 2026
Inclusion and exclusion criteria	Inclusion criteria: peer-reviewed articles, clinical trials, and observational studies published in English
	Exclusion criteria: case reports, non-English literature, and studies involving moderate-to-severe rheumatic mitral stenosis or mechanical prosthetic heart valves
Selection process	Two reviewers (Lei Liu and A.C.) independently conducted the literature selection. Disagreements were resolved through consensus with a third reviewer (L.Z.)

Pathophysiological basis of thrombosis in AF

Structural remodeling, blood stasis, and atrial cardiomyopathy

In AF, the underlying substrate for thrombus formation is best conceptualized through the lens of “atrial cardiomyopathy”. This evolving concept extends far beyond simple anatomical remodeling, representing a complex spectrum of structural, hemodynamic, and electrical alterations affecting the atria (18,19). Within this cardiomyopathic process, structural remodeling sets the visible stage for thrombogenesis. The left atrium, particularly the LAA, undergoes significant enlargement and morphological changes (20-22). These alterations are markers of deeper pathological changes and result in profound hemodynamic disturbances, leading to a loss of effective contraction.

Functional abnormalities and myocardial pathology

Deterioration of atrial mechanical and electrical function significantly contributes to thrombogenesis, further illustrating the scope of atrial cardiomyopathy beyond mere anatomical enlargement (18,23). AF causes profound electrical and contractile dysfunction, often measured by parameters like left atrial strain, which reflect the reservoir, conduit, and pump functions of the atrium (24,25). Impaired atrial function, indicative of underlying atrial cardiomyopathy and interstitial fibrosis (26), exacerbates blood stasis regardless of absolute atrial size.

Hypercoagulable state

A systemic inflammatory response and hypercoagulable state frequently occur in AF patients (27). Inflammatory cytokines activate the coagulation system and platelets, interacting with endothelial damage to promote thrombus formation (28). While echocardiography cannot directly visualize these conditions, understanding them underscores the importance of combining echocardiographic parameters with biomarkers related to inflammation and coagulation activation, such as D-dimer, to develop a comprehensive risk assessment model (29). Table 2 systematically categorizes and compares key echocardiographic parameters used to assess thrombotic risk driven by underlying pathophysiological mechanisms. It details their imaging modalities, associated risk types, as well as principal advantages and limitations (4,5,10,16,30-35). These parameters offer foundational evidence and quantitative metrics crucial for developing the multimodal risk prediction models discussed in the subsequent section.

Table 2. Key echocardiographic parameters for thrombotic risk assessment in AF.

Parameter category	Parameter name	Modality	Predicts risk	Key advantages	Major limitations
Basic structure & function	LAD	2D-TTE	Future stroke/systemic embolism	Simple, rapid, widely available; correlated with stroke risk	Crude estimate; poor for asymmetric geometry; low reproducibility; superseded by LAVI
	LAVI	2D-TTE/3D-TTE	Future stroke/systemic embolism	More accurate than LAD for size and remodeling; ESC guideline recommended	2D relies on assumptions; 3D is accurate but time-consuming
	LVEF	2D-TTE/3D-TTE	Future stroke/systemic embolism	Core measure of global systolic function	Non-specific; does not reflect LA/LAA hemodynamics
	LVMI	2D-TTE	Future stroke/systemic embolism	Marker of long-term load and remodeling	Cumbersome calculation; indirect link to thrombosis
Myocardial mechanics & work	LVMW	2D-STE + blood pressure	Future stroke/systemic embolism (potential)	Load-independent; more sensitive than LVEF	Novel; limited data; complex acquisition
	LASr	2D-STE	Future stroke/prevalent LAAT/dense SEC	Highly sensitive for LA function; early fibrosis detection	High image quality needed; vendor-dependent; lacks cut-offs
	LA conduit strain	2D-STE	Investigational	Reflects passive LA emptying	Limited data; independent value unclear
	LA contractile strain	2D-STE	Future stroke/systemic embolism	Directly measures active LA contraction	Unreliable in AF
LAA specific parameters	LAA ostium size	TEE	Prevalent LAAT	Key for device sizing; larger size increases risk	Measurement variable; weak standalone value
	LAA strain	2D-STE/3D-STE (TEE)	Prevalent LAAT/dense SEC	Direct measure of LAA mechanical function	Requires TEE; technically challenging; needs validation
	LAA morphology	TEE/cardiac CT	Prevalent LAAT/procedural risk	Anatomical risk: “Chicken Wing” (low risk), others (high risk)	Subjective classification; mixed morphologies
	LAAeV	TEE	Prevalent LAAT/dense SEC	Gold standard for blood stasis; strong evidence	Semi-invasive TEE; values vary with rhythm

2D, two-dimensional; 3D, three-dimensional; AF, atrial fibrillation; CT, computed tomography; ESC, European Society of Cardiology; LA, left atrium; LAA, left atrial appendage; LAAeV, LAA emptying velocity; LAAT, left atrial appendage thrombus; LAD, left atrial diameter; LASr, left atrial reservoir strain; LAVI, left atrial volume index; LVEF, left ventricular ejection fraction; LVMI, left ventricular mass index; LVMW, left ventricular myocardial work; SEC, spontaneous echo contrast; STE, speckle-tracking echocardiography; TEE, transesophageal echocardiography; TTE, transthoracic echocardiography.

Beyond traditional scores: multimodal models incorporating echocardiographic parameters

Integrative framework for multimodal risk prediction

To synthesize the heterogeneous evidence base, ranging from externally validated models to mechanistic artificial intelligence (AI) research, we prioritized studies according to sample size, study design, external validation status, and direct clinical relevance. Exploratory or methodological AI studies not specifically designed for AF thromboembolic prediction were evaluated strictly as conceptual examples.

To maximize practical utility, we propose an integrative framework addressing critical clinical decision points. Multimodal risk assessment is most valuable in clinical scenarios where conventional risk scores exhibit limited discriminatory power, primarily including patients at intermediate stroke risk, such as those with a CHA₂DS₂-VA score of 1–2, for whom the decision to initiate long-term oral anticoagulation remains uncertain, as well as individuals scheduled for electrical cardioversion or catheter ablation, where ruling out LAA thrombus is essential. In these settings, integrating advanced functional parameters [e.g., left atrial reservoir strain (LASr)], circulating biomarkers [e.g., N-terminal pro-B-type natriuretic peptide (NT-proBNP)], and anatomical imaging refines individual risk profiling, guiding precise procedural triage and tailored management strategies.

Limitations and evolution of clinical scores: from CHA₂DS₂-VASc to CHA₂DS₂-VA

The CHA₂DS₂-VASc score, an advancement over the earlier CHADS₂ system, is fundamental for thromboembolic risk stratification in AF. It incorporates variables like female sex, age stratification, and vascular disease, and is widely used in clinical practice due to its enhanced ability to identify low-risk patients and its proven correlation with stroke risk across diverse populations (7,29,30,36-39).

Despite its utility, the CHA₂DS₂-VASc score has notable limitations. It solely relies on clinical parameters, neglecting cardiac physiological status, such as structural and functional changes, and biomarker levels (40,41). This oversight diminishes its predictive accuracy in certain subgroups, including younger patients, those with cardiac comorbidities like significant left atrial enlargement or valvular insufficiency, and those with metabolic disorders such as hyperuricemia, potentially leading to the undertreatment of high-risk individuals (42-46). The score is particularly inadequate for stratifying risk in patients with low to intermediate scores (CHA₂DS₂-VASc 0–1), where imaging and biochemical markers provide crucial supplemental information (47). To address these limitations, early studies provided crucial evidence regarding the incremental value of echocardiographic parameters (48). However, as summarized in the review by Providência et al. (49), efforts to evaluate whether adding simple, one-dimensional echocardiographic parameters (such as left atrial diameter) to clinical scores could increase event prediction generally failed to demonstrate a meaningful advantage, with the stroke prevention in atrial fibrillation (SPAF) trial being a notable exception (50). This highlighted that simple linear measurements are insufficient. In contrast, more integrated early models showed great promise. Predating many contemporary scores, pioneering tools successfully demonstrated the power of multimodal integration. The CHA₂DS₂-VASc + echo score revealed that combining left atrial area and left ventricular global systolic function with clinical parameters significantly improved the prediction of left atrial stasis and thrombus. Concurrently, the Comprehensive Atrial Thrombus Evaluation Score (CATES) integrated atrial volume with biomarkers (C-reactive protein and troponin) to safely identify patients who could be spared transesophageal echocardiography (TEE) [area under the curve (AUC) =0.816] (51). Building upon these foundational contributions, subsequent enhancements, such as the CHA₂DS₂-VASc-RAF and CHA₂DS₂-VASc-LAF scores, continued this trend by incorporating echocardiographic and renal function parameters (52). Ultimately, this historical context explains the evolution from simple linear measurements and highlights the absolute necessity of modern multimodal models that utilize advanced, multi-dimensional parameters [such as myocardial strain and three-dimensional (3D) morphology].

Building upon these foundational contributions, modern assessment tools have become significantly more refined. Initial enhancements, such as the CHA₂DS₂-VASc-renal dysfunction, atrial fibrillation type (RAF) and CHA₂DS₂-VASc-left atrial diameter and atrial fibrillation type (LAF) scores, began incorporating echocardiographic and renal function parameters (29,52,53). More comprehensive tools like the creatinine, left ventricular ejection fraction, overload, tricuspid annular plane systolic excursion, stroke, and AF rhythm (CLOTS-AF) and thrombolysis in myocardial infarction-atrial fibrillation (TIMI-AF) scores further this trend by systematically combining clinical history, echocardiographic parameters [e.g., left atrial volume index (LAVI), left ventricular ejection fraction (LVEF)], and biomarker data for a multidimensional risk assessment (54-56). The emergence of machine learning models underscores the potential of integrating complex, multimodal data to enhance predictive accuracy (57). These advances highlight the critical need to progress beyond the CHA₂DS₂-VASc score towards personalized, multimodal risk evaluation models that more accurately capture the pathophysiology of AF. Furthermore, the landscape of clinical risk scores is rapidly evolving. Recent epidemiological studies have demonstrated that the increased stroke risk historically associated with female sex is no longer detectable, prompting a re-evaluation of the sex category criterion (58-60). Consequently, the 2024 European Society of Cardiology (ESC) Guidelines have officially replaced the CHA₂DS₂-VASc score with the CHA₂DS₂-VA score, removing the female sex parameter to prevent overtreatment in low-risk women (61,62). However, the CHA₂DS₂-VASc score currently remains recommended by the American College of Cardiology/American Heart Association (ACC/AHA) guidelines. Although the recent ESC update removed female sex as a stand-alone risk factor, this does not imply that female sex is no longer clinically relevant. Instead, female sex is more accurately described as a critical clinical risk modifier. When present alongside other established cardiovascular comorbidities, it synergistically amplifies the overall thromboembolic risk, highlighting the need for comprehensive risk assessment. This divergence underscores a critical transition period in AF management. To maintain a global perspective, future multimodal tools must be designed to integrate advanced echocardiographic and biomarker data with both the CHA₂DS₂-VASc and the CHA₂DS₂-VA frameworks, accurately capturing the pathophysiology of AF regardless of the foundational clinical scoring system used. The following sections describe these advanced multimodal approaches, with their key features summarized for comparative analysis in Table 3.

Table 3. Classic studies of multimodal models based on echocardiographic parameters.

Authors/model	Target cohort (sample size)	Predicted endpoint	Imaging modality	Key predictors	Validation & performance
Providência et al., 2011 (CHA2DS2-VASc + Echo)	AF patients (N/A)	Current LAAT or dense SEC	2D-TTE	Echo: LA area, LVEF; clinical: CHA2DS2-VASc	Internal: improved prediction over CHA2DS2-VASc alone
Providência et al., 2012 (51) (CATES score)	AF patients before TEE (N/A)	Current LAAT	2D-TTE	Echo: atrial volume; clinical: CRP, troponin, stroke history	Internal: AUC =0.816
Yamamoto et al., 2014 (63) (3D-morphological model)	AF patients (N=564)	Current LAAT	3D-TEE	Echo: LAA lobes (≥3); clinical: N/A	Internal: independent risk factor, surpassed CHADS2
Pagola et al., 2021 (64) (LASr/NT-proBNP)	PAF with cryptogenic stroke (N/A)	High-risk embolic PAF	2D-STE	Echo: LASr; clinical: NT-proBNP	Prospective multicenter: AUC =0.82
Segan et al., 2023 (33) (CLOTS-AF score)	AF patients before TEE (N=1,001)	Current LAAT or dense SEC	2D-TTE	Echo: LAVI, LVEF, TAPSE; clinical: stroke history, renal insufficiency	Derivation: AUC =0.78
Pieszko et al., 2024 (65) (LAT-AI model)	AF patients (training: 2,827; validation: 1,284)	Current LAAT	2D-TTE/AI	Echo: LAVI; clinical: heart failure, rhythm	External: AUC =0.85 (99% NPV at specific threshold)
Kalenderoglu et al., 2025 (66) (CLOTS-AF external validation)	AF patients before TEE (N=668)	Current LAAT	2D-TTE	Echo: LAVI, LVEF, TAPSE; clinical: stroke history, renal insufficiency	External validation: AUC =0.70

2D, two-dimensional; 3D, three-dimensional; AF, atrial fibrillation; AI, artificial intelligence; AUC, area under the curve; CATES, Comprehensive Atrial Thrombus Evaluation Score; CHA₂DS₂-VASc, congestive heart failure, hypertension, age ≥75 years, diabetes, stroke, vascular disease, age 65–74 years, sex category; CLOTS-AF, creatinine, left ventricular ejection fraction, overload, tricuspid annular plane systolic excursion, stroke, and AF rhythm; CRP, C-reactive protein; LA, left atrial; LAA, left atrial appendage; LAAT, left atrial appendage thrombus; LASr, left atrial reservoir strain; LAT, left atrial thrombus; LAVI, left atrial volume index; LVEF, left ventricular ejection fraction; N/A, not applicable; NPV, negative predictive value; NT-proBNP, N-terminal pro-B-type natriuretic peptide; PAF, paroxysmal atrial fibrillation; SEC, spontaneous echo contrast; STE, speckle-tracking echocardiography; TAPSE, tricuspid annular plane systolic excursion; TEE, transesophageal echocardiography; TTE, transthoracic echocardiography.

Paradigms of multimodal model development and integration strategies

Clinical-echocardiographic integration

Integrating clinical and echocardiographic parameters is a foundational approach to developing multimodal models. The CLOTS-AF score exemplifies this integration. Developed from a study of 1,001 patients undergoing TEE, it identified key independent predictors of LAAT or dense SEC prior to cardioversion. These predictors include history of stroke, renal insufficiency (creatinine >1.5 mg/dL), LVEF <50%, LAVI >34 mL/m², tricuspid annular plane systolic excursion (TAPSE) <17 mm, and AF rhythm at the time of TEE (33). Each point in the CLOTS-AF score (range, 0–12) significantly increased prevalent LAAT/dense SEC risk [odds ratio (OR) =1.63]. This model demonstrated superior predictive performance (AUC =0.78) compared to the CHA₂DS₂-VASc score (AUC =0.62), with higher identification accuracy (72% vs. 51%) (33). External validation in 668 patients confirmed its robustness (AUC =0.70 vs. 0.61 for CHA₂DS₂-VASc, P<0.001), establishing it as a powerful tool for identifying high-risk patients who may benefit from standardized anticoagulation over immediate TEE-guided cardioversion. Crucially, this score optimizes procedural resource utilization by predicting prevalent LAAT/dense SEC, rather than predicting long-term future stroke risk (66,67). Therefore, its clinical utility is currently limited to triage before cardioversion, and it must not be generalized to routine, long-term anticoagulation decisions without further prospective outcome data [see Table 3 and Figure 2 (68-70)].

Figure 2.

Key Echocardiographic parameters for thrombotic risk assessment in atrial fibrillation. This schematic summarizes the essential echocardiographic parameters utilized in the evaluation of thrombotic risk in patients with AF. The parameters are systematically categorized into those assessing cardiac structure [e.g., left atrial volume index (68); left ventricular ejection fraction], myocardial mechanics (e.g., left ventricular myocardial work), and LAA specific morphology (69) and function (70) (e.g., LAA emptying velocity). These measurements, obtainable via transthoracic and transesophageal echocardiography, provide direct, quantifiable insights into the pathophysiological substrate of atrial cardiomyopathy, such as stasis, dilation, and impaired contractility, which are the primary drivers of intracardiac thrombus formation. The integration of these parameters surpasses the limitations of clinical risk scores alone, enabling a more precise and individualized assessment of thromboembolic risk. The panel “Left ventricular myocardial work” was reused from Open Access article from Liang Z et al. (68) under the terms of the Creative Commons Attribution 4.0 International License (CC BY-NC-ND 4.0) (https://creativecommons.org/licenses/by-nc-nd/4.0/). The panel “Left atrial appendage morphology” was reused from Open Access article from He J et al. (69) under the terms of the Creative Commons Attribution 4.0 International License (CC BY-NC-ND 4.0) (https://creativecommons.org/licenses/by/4.0/). The panel “Left atrial appendage strain” was reused from Open Access article from Skrebelyte-Strøm L et al. (70) under the terms of the Creative Commons Attribution 4.0 International License (CC BY-NC-ND 4.0) (https://creativecommons.org/licenses/by-nc/4.0/). LAA, left atrial appendage.

Fusion of functional parameters and biomarkers

Combining sensitive imaging indices with biomarkers offers deeper pathophysiological insight into the atrial substrate. This strategy is illustrated by combining LASr, a measure of atrial mechanical function obtained via speckle-tracking echocardiography, with NT-proBNP, a biomarker reflecting atrial pressure and load. A multicenter prospective study showed that this combination could accurately identify cryptogenic stroke patients with paroxysmal AF at high risk for future embolic events, surpassing models based solely on atrial size or clinical characteristics (64,71). LASr detects early atrial myopathy and fibrosis, while NT-proBNP quantifies hemodynamic consequences. Together, they reveal the pathological state of “atrial cardiomyopathy”, predisposing patients to thrombosis (15,64,72). This strategy provides a comprehensive risk assessment, guiding decisions for prolonged rhythm monitoring and early anticoagulation to prevent stroke recurrence (64) (see Table 3). Furthermore, integrating speckle-tracking echocardiography, specifically LASr, with circulating biomarkers significantly advances non-invasive risk stratification. As a reliable indicator of atrial fibrosis and mechanical remodeling, impaired LASr correlates independently with LAAT, even after adjustment for clinical scores and left atrial size (73,74). Consequently, combining LASr with biomarkers such as NT-proBNP refines individualized risk assessment to identify patients with an exceptionally low probability of thrombus formation. This integrated strategy can potentially minimize the need for routine TEE prior to cardioversion or ablation, establishing myocardial deformation imaging as a crucial component of precision cardiology.

Data-driven integration: AI model

AI and machine learning offer transformative potential for integrating complex, high-dimensional multimodal data. The left atrial thrombus (LAT)-AI model, developed using the extreme gradient boosting (XGBoost) algorithm on a large registry (n=2,827 training, n=1,284 validation), integrates diverse data points such as LAVI, heart failure status, and cardiac rhythm to non-invasively predict LAAT (65). It achieved exceptional predictive performance (AUC =0.85), outperforming models based on LVEF or the CHA₂DS₂-VASc score. A specific decision threshold yielded a 99% negative predictive value, suggesting its application could allow approximately 40% of patients to safely avoid invasive TEE prior to cardioversion or ablation (65). It is important to note that this model is currently designed for procedural TEE avoidance rather than guiding routine, long-term anticoagulation decisions. Similar to the CLOTS-AF score, the LAT-AI model requires rigorous validation through further prospective outcome data before its application can be generalized to broader clinical scenarios, such as determining long-term stroke prevention strategies (see Table 3). In addition, the Hierarchical Multimodal Self-Attention-Based Graph Neural Network (HMSA-DTI) model exploited intrinsic connections among various modalities to achieve deeper feature fusion, significantly improving drug-target prediction accuracy for drug-target interaction prediction (75). While these specific models were initially developed for different clinical or pharmacological endpoints (75), their underlying architectural frameworks demonstrate AI’s ability to uncover complex, non-linear interactions between variables that traditional statistics might miss, paving the way for personalized, data-driven risk assessment in AF (75).

Expanding the multimodal framework

The role of LAA anatomy

Unique risk stratification is achieved through anatomical assessment of the LAA. A seminal study using 3D-TEE evaluated LAA morphology in 564 patients, finding that possessing ≥3 LAA lobes was an independent thrombus formation risk factor, surpassing the predictive value of the CHADS₂ score, LVEF, left atrial volume, and spontaneous echo contrast (63). This morphological characteristic is a stable anatomical risk marker, aiding anticoagulation decisions, especially in patients with low clinical scores (63) (see Table 3).

Histopathological correlation

While most current research focuses on predicting thrombus formation in AF patients, the underlying mechanisms are seldom examined. A Hiroshima University research team provided the first direct evidence of the correlation between LAA fibrosis, endothelium damage, LAAT, and ischemic stroke occurrence through histological analysis of 96 AF patients who underwent LAA resection. Results indicated that greater LAA fibrosis and endothelial damage were correlated with increased thrombus and stroke risk. Endothelial damage predicted these risks independently of AF type (paroxysmal or persistent). Furthermore, fibrosis was highly correlated with endothelial damage, and CHADS₂ scores showed a positive correlation with both. This supports the joint role of fibrosis and endothelial damage in promoting thrombus formation amidst the pathological progression of atrial cardiomyopathy related to AF (15,76). These histological findings deepen our understanding of AF-related thrombogenesis. They emphasize the importance of anticoagulation therapy guided by the pathological state of the atrial myocardium, rather than merely the presence of AF (15,23).

The complementary value of biomarkers

Biomarkers quantify systemic prothrombotic states. Growth differentiation factor-15 (GDF-15), an inflammation marker, has levels associated with LA/LAA thrombus risk, with values >935.0 pg/mL serving as an independent predictor (77,78). Similarly, D-dimer, a marker of fibrin degradation and active thrombosis, correlates with structural changes like left atrial enlargement and is predictive, especially in patients with stroke history (79). Integrating these biomarkers with echocardiographic parameters creates a holistic risk profile, reflecting both atrial substrate and systemic clot formation tendency (80). Furthermore, combining biomarkers like interleukin (IL)-6 and IL-4 with echocardiographic parameters has been shown to more accurately assess thrombotic risk in elderly AF patients. Elevated IL-6 levels are associated with thrombotic events (81,82). Echocardiography can evaluate changes in atrial function, enabling a more comprehensive risk assessment. This multimodal risk assessment strategy enhances the capability to identify high-risk patients and provides a basis for individualized treatment.

Comparative analysis and synthesis of multimodal models

The multimodal models discussed and synthesized in Table 2 demonstrate an evolution from reliance on clinical checklists to comprehensive, pathophysiology-based assessment. The CLOTS-AF score integrates clinical and echocardiographic parameters into a validated risk score. The strain-NT-proBNP approach uses advanced imaging and biomarkers to detect subclinical atrial myopathy. The LAT-AI model utilizes machine learning to analyze complex patterns within multimodal data for improved prediction. Anatomical (LAA morphology) and histopathological evidence provide biological plausibility for these models. Collectively, they offer a versatile toolkit for clinicians, enabling precise, individualized risk stratification beyond the CHA₂DS₂-VASc score alone.

Practical limitations of echocardiographic assessment

Despite the predictive value of multimodal models, several practical limitations constrain the clinical application of advanced echocardiographic parameters. First, significant inter-vendor and inter-observer variability remains a major hurdle, particularly in speckle-tracking strain analysis. This variability is further compounded by the lack of universally accepted cut-offs and standardized software platforms. Second, echocardiographic indices are inherently rhythm-dependent. Acquiring reliable functional data, such as LAA flow velocity or strain, during irregular AF rhythms is technically challenging. This process typically requires averaging multiple cardiac cycles, thereby introducing potential measurement bias. Furthermore, structural and functional assessments are highly sensitive to dynamic loading conditions and the patient’s volume status during the examination. Finally, diagnostic accuracy relies heavily on operator expertise and is frequently compromised by poor acoustic windows or suboptimal image quality. Although current retrospective analyses are compelling, large-scale prospective validation studies are urgently needed. Establishing standardized acquisition protocols and robust threshold values is essential before integrating these echocardiographic indices into routine clinical practice.

Intended clinical applications of multimodal models

The clinical utility of these multimodal models encompasses multiple clinical scenarios based on their design and endpoints. For TEE avoidance before cardioversion or catheter ablation, models like the CLOTS-AF score (33,66) and LAT-AI predict prevalent left atrial thrombi to safely identify low-risk patients who can bypass pre-procedural TEE (65), though ablation-specific validation remains necessary. Regarding long-term anticoagulation decision support, models predicting incident thromboembolism help guide oral anticoagulation strategies, particularly for “gray zone” patients with low clinical scores but significant subclinical remodeling (47). Finally, beyond immediate thrombotic risk stratification, these imaging and biomarker tools facilitate the early identification of underlying atrial cardiomyopathy by detecting subclinical mechanical deterioration, thereby potentially guiding broader rhythm management and upstream therapies (18,24).

Future directions in multimodal risk prediction

Standardization and technical advancement

The widespread clinical adoption of multimodal models hinges on resolving measurement variability. Key echocardiographic parameters, such as LAVI and left atrial strain, suffer from inter-observer and inter-vendor variability due to the absence of standardized acquisition and analysis protocols (83,84). Future efforts must focus on developing international consensus guidelines to ensure reproducibility across centers and platforms. Concurrently, AI offers a powerful solution for automation and standardization. AI algorithms can be trained to automatically segment cardiac structures and quantify functional parameters, minimizing human-dependent error and enhancing measurement consistency (85,86). Integrating AI-based analytical tools into echocardiography machines represents a key technological advancement that will support the reliability of future risk prediction models (see Figure 3).

Figure 3.

Clinical decision pathway for thromboembolic risk assessment in AF patients. This flowchart outlines a forward-looking, hierarchical framework for thromboembolic risk stratification in AF patients, guiding clinicians through multimodal assessment strategies. Crucially, this pathway is designed to complement rather than override guideline-based anticoagulation decisions, serving as a supplementary decision-support tool tailored for clinical patients. AF, atrial fibrillation; AI, artificial intelligence; CHA₂DS₂-VASc, congestive heart failure, hypertension, age ≥75 years, diabetes, stroke, vascular disease, age 65–74 years, sex category; CLOTS-AF, creatinine, left ventricular ejection fraction, overload, tricuspid annular plane systolic excursion, stroke, and AF rhythm; LAA, left atrial appendage; LAT, left atrial thrombus; NT-proBNP, N-terminal pro-B-type natriuretic peptide; TEE, transesophageal echocardiography; TTE, Transthoracic Echocardiography.

Clinical validation and implementation pathways

For multimodal models to transition from research tools to clinical assets, robust prospective validation in large, diverse, real-world cohorts is essential. Multicenter studies are needed to confirm the generalizability of models like CLOTS-AF and LAT-AI across different populations and healthcare settings (33,65,66). Additionally, research must focus on demonstrating clinical utility, how the use of these models improves patient outcomes (e.g., reducing stroke without increasing major bleeding) and optimizes healthcare resource allocation. This involves developing clear clinical decision pathways that incorporate model outputs, such as using high negative predictive value to safely avoid TEE in low-risk patients (65). Successful implementation will require embedding these models into electronic health records and clinical decision support systems, providing clinicians with actionable risk assessments at the point of care.

Towards dynamic and deeply phenotyped risk assessment

Future risk prediction will evolve from static, cross-sectional assessments to dynamic, multi-omics-informed profiling. Beyond imaging and rhythms, incorporating genetic polymorphisms (e.g., related to anticoagulant drug metabolism), proteomic profiles, and other biomarkers can create a “deep phenotype” of the individual patient’s pathophysiology and therapeutic response (87). While these omics-based strategies are currently not applicable as clinical tools, this holistic approach, powered by advanced machine learning capable of fusing diverse data types, holds the long-term potential for truly personalized, precision anticoagulation management in AF, where treatment could one day be tailored to both the risk of clot formation and the individual’s specific biological characteristics.

Harmonizing simple clinical scores and precise multimodal models

While this review has predominantly focused on the development of highly precise multimodal models utilizing specialized equipment and blood sampling, it is crucial to recognize that the evolution of risk prediction encompasses two distinct but complementary directions. The first direction focuses on refining simple, purely clinical scores that rely solely on general patient information (e.g., medical history, demographics, and basic physical signs). These streamlined models are paramount for general physicians in primary care settings, providing a rapid, equipment-free method to determine the initial indication for anticoagulation therapy. The second direction, which is the focus of this review, involves the development of precise multimodal models designed to accurately estimate thromboembolic risk by incorporating detailed echocardiographic and biomarker data.

These two directions are not mutually exclusive; rather, they form a tiered approach to patient management. Simple clinical scores act as the accessible first line of defense for broad population screening and initial clinical decision-making. In contrast, precise multimodal models serve as advanced tools for specialists, particularly vital for patients falling into clinical “grey zones” or those requiring a deeper pathophysiological assessment before proceeding with interventions like cardioversion or ablation. Future research should not only advance each direction independently but also establish clear clinical pathways that seamlessly integrate both, ensuring that every patient receives the appropriate level of risk stratification.

Conclusions

This review synthesizes the evidence supporting the development and application of multimodal models for thromboembolic risk prediction in AF. While the traditional CHA₂DS₂-VASc score has been foundational, and the newly introduced CHA₂DS₂-VA score reflects critical epidemiological updates by removing the sex criterion, both scoring systems fundamentally rely on clinical parameters. They inherently face limitations in assessing the individual’s cardiac substrate, specifically the structural, hemodynamic, and electrical manifestations of atrial cardiomyopathy. Echocardiographic parameters, from basic measures like LAVI and LVEF to advanced indices like left atrial strain, provide quantifiable insights into this pathophysiology. Integrating these parameters with clinical data and biomarkers into multimodal models, such as the CLOTS-AF score, the strain-NT-proBNP approach, and the LAT-AI model, has consistently demonstrated superior predictive accuracy over traditional scores. However, it is crucial to emphasize that to date, thromboembolic risk cannot be evaluated by relying solely on echocardiographic parameters. Because they cannot entirely capture the systemic complexity of thrombogenesis, these imaging metrics are not standalone replacements for traditional clinical evaluation. Rather, they serve as powerful "add-on" tools designed to complement and refine baseline clinical risk stratification. Furthermore, a clear distinction must be made regarding the predicted outcomes: while models like strain-NT-proBNP aim to identify patients at high risk for future stroke, models such as CLOTS-AF and LAT-AI are primarily validated for detecting prevalent LAAT/dense SEC to safely avoid procedural TEE. These models represent a shift from a one-size-fits-all approach to a precision medicine paradigm, allowing for more accurate identification of patients who will benefit from anticoagulation. Crucially, to maintain global applicability, future multimodal models should be adaptable, designed to potentiate risk prediction on top of both the CHA₂DS₂-VASc (ACC/AHA) and CHA₂DS₂-VA (ESC) frameworks. Crucially, a clear distinction must be made between statistical predictive performance and verified clinical utility or therapeutic benefit. Although multimodality and AI-driven models efficiently identify low-risk patients, primarily serving to avoid unnecessary TEE prior to cardioversion or ablation, prospective interventional evidence confirming that these tools directly reduce stroke or systemic embolism rates is currently lacking. Future clinical trials are essential to bridge the gap between predictive accuracy and outcome-improving precision medicine.

Routine clinical use of these advanced tools requires concerted efforts across multiple fronts. While AI and automated analysis hold substantial promise for standardizing cardiovascular imaging by mitigating operator-dependent variability, their clinical translation faces critical technical barriers. These include training dataset selection biases, vendor- and platform-specific image characteristics, suboptimal image acquisition quality, cross-institutional protocol discrepancies, and domain shifts in external populations. Furthermore, physiological variations, such as differences between sinus rhythm and active AF, along with demographic inconsistencies across ethnicity, body size, sex, and underlying comorbidities, can severely impair model generalizability.

Consequently, successful clinical implementation requires rigorous multi-center external validation, standardized acquisition protocols, multi-vendor reproducibility testing, seamless electronic health record integration, enhanced model interpretability, and robust post-deployment monitoring. Overcoming these hurdles is essential to transition AI from an exploratory methodology into clinically actionable decision-support tools. Combined with dynamic monitoring and genomic information, these advancements will pave the way for adaptive, deeply personalized risk assessments. Through interdisciplinary collaboration among clinicians, imagers, data scientists, and clinical trialists, translating these multimodal models into clinical practice can fundamentally transform AF management, leading to more effective stroke prevention and improved patient outcomes.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.