Statistical shape analysis of the left atrial appendage predicts stroke in atrial fibrillation

Abstract

Purpose:

The shape of the left atrium (LA) and left atrial appendage (LAA) have been shown to predict stroke in patients with atrial fibrillation (AF). Prior studies rely on qualitative assessment of shape, which limits reproducibility and clinical utility. Statistical shape analysis (SSA) allows for quantitative assessment of shape. We use this method to assess the shape of the LA and LAA and predict stroke in patients with AF.

Methods:

From a database of AF patients who had previously undergone MRI of the LA, we identified 43 patients with AF who subsequently had an ischemic stroke. We also identified a cohort of 201 controls with AF who did not have a stroke after the MRI. We performed SSA of the LA and LAA shape to quantify the shape of these structures.

Results:

We found three of the candidate LAA shape parameters to be predictive of stroke, while none of the LA shape parameters predicted stroke. When the three predictive LAA shape parameters were added to a logistic regression model that included the CHA₂DS₂-VASc score, the area under the ROC curve increased from 0.640 to 0.778 (p = .003).

Conclusion:

The shape of the LA and LAA can be assessed quantitatively using SSA. LAA shape predicts stroke in AF patients, while LA shape does not. Additionally, LAA shape predicts stroke independent of CHA₂DS₂-VASc score. SSA for assessment of LAA shape may improve stroke risk stratification and clinical decision making for AF patients.

Background

Morphology, or shape, of the left atrium (LA) and left atrial appendage (LAA) have been shown to be important markers of remodeling in atrial fibrillation (AF) (1–4). Shape changes and other markers of LA remodeling can predict stroke in AF patients, although currently are rarely used clinically(3, 5–8). Previous studies of LAA morphology have relied on qualitative assessments of shape because the complex geometry of atrial shape is difficult to measure quantitatively(3). However, qualitative assessments of shape are subject to interobserver variation and may omit important morphologic details. Quantitative methods offer more potential for reproducibility in clinical use. Quantitative measures of LA shape have been studied(7, 8), but those that have been applied to date rely on a-priori assumptions about which measures of shape are relevant to clinical outcome.

Statistical shape analysis (SSA) is a computational approach to assessment of shape in medical imaging, and has been widely used in cardiac imaging(9) and other imaging fields(10, 11). In our study we use a specific SSA approach called Particle Based Modeling,(12) an image analysis tool that measures anatomical shape quantitatively from structural MRI or CT without imposing assumptions about the relevant geometry. This approach models LA and LAA morphology by automatically computing a set of corresponding surface landmark points on the surface of the atrium. We have previously demonstrated that LA shape predicts recurrence of AF after ablation using this method(1). In this study, we use SSA to assess shape of the LA and LAA quantitatively and determine whether the shape of these structures predicts stroke in patients with AF.

Methods

Patient Selection and Imaging

In this retrospective case-control study, we used a database of AF patients who underwent magnetic resonance imaging (MRI) of the left atrium in preparation for AF ablation to identify 43 patients who had subsequent ischemic stroke after the initial MRI. Details of the cohort from which cases were selected has been previously published(5). Scans were obtained between 2006 and 2015. All patients were required to have adequate imaging of the LA and LAA for segmentation to be included. Hemorrhagic strokes and transient ischemic attacks were excluded. A subset control group was selected from the non-stroke patients with a target ratio of greater than four controls per case. After eliminating patients with inadequate imaging of the left atrium, the control group consisted of 201 patients (Supplementary Figure 1). We collected data on multiple clinical comorbidities and patient characteristics, and then computed the CHA₂DS₂-VASc score for all patients. Imaging protocols have been previously published(13). Imaging was performed on 1.5T and 3T scanners (Siemens, Erlangen, Germany). DE-MRI was acquired approximately 15 minutes after the contrast agent injection using 3-dimensional (3D) inversion-recovery-prepared, free breathing with respiratory navigating, ECG-gated, gradient-echo pulse sequence with fat saturation. Acquisition parameters were as follows: a transverse imaging volume with true voxel size of 1.25×1.25×2.5 mm, flip angle of 22°, repetition time/echo time of 6.1/2.4 ms, inversion time of 230 to 320 ms, and parallel imaging with GRAPPA technique with R=2 and 42 reference lines. Typical scan time was 10 minutes for the sequence which varies based on heart rate and respiratory rate.

Imaging Segmentation and Shape Modeling

Steps for the segmentation and shape modeling process are shown in Figure 1. Trained observers manually segmented the LA and LAA in all patients using Corview image processing software (MARREK, Inc., Sandy, UT) by tracing the endocardial surface of each structure. Observers were blind to the clinical outcomes. We created three dimensional reconstructions of the LA and LAA from the segmentations. The details of our methods for SSA have been previously published(2, 12, 14). SSA using Particle-Based Modeling was performed both on the LA and LAA separately to define a set of 512 and 64 surface correspondence points, respectively. The process inherently controls for LA and LAA size by scaling. Thus, measures of size are removed from the analysis and shape parameters quantify shape independent of size.

Figure 1:

Steps of the LA and LAA shape modeling process are shown. First, MRI of the LA and LAA is obtained (A). Second, the LA and LAA are manually segmented (B). Third, a three-dimensional surface of the LA and LAA is created using the segmented images (C). Finally, SSA with Particle-Based Modeling is applied to define a set of surface correspondence points that define the overall shape of each structure (D).

Statistical Analysis

We used R: A Language and Environment for Statistical Computing (R Foundation for Statistical Computing, Vienna, Austria) for all statistical analysis. From each set of correspondence points, we used principal component analysis (PCA) to compute PCA modes, which we have defined as “shape parameters”. This resulted in two sets of shape parameters, one for the LA and one for the LAA. Based on PCA, the ten shape parameters from each set that described the largest amount of shape variance in the cohort were considered as candidates for inclusion in the statistical model for prediction of stroke. We used the LASSO logistic method to select the shape parameters most predictive of stroke(15). LASSO is a regularized regression model which penalizes model complexity. Using this method, shape parameters not predictive of stroke have coefficients set to zero.

The result is a simpler model that includes only shape parameters predictive of stroke. Next, we created a second logistic regression model using CHA₂DS₂-VASc score. Finally, we added shape parameters to the CHA₂DS₂-VASc score to create a third model. The second and third models were compared using the DeLong method(16) to determine if shape significantly improved stroke prediction independent of CHA₂DS2-VASc score. A p-value < 0.05 was considered statistically significant for all statistical comparisons performed.

Results

Clinical characteristics of the patients included in the stroke and no stroke groups are shown in Table 1. Mean patient age was 69 and a minority of patients were female. Clinical characteristics were similar between groups with the exception of vascular disease and history of prior stroke. The CHA₂DS₂-VASc score was higher in the stroke group than the no stroke group (median 4 vs. 3, p = .006). Patients in the stroke group had a median time of 420 days between the incident MRI and the stroke. Patients in the control group had a median follow up of 551 days.

Table 1:

Baseline characteristics of the patients included in the study in both the stroke and non-stroke groups are shown. Comorbidities are displayed as number of subjects and percentage of the cohort. Age is displayed as mean and standard deviation (SD). CHA₂DS₂-VASc score, days from MRI to stroke, and days of follow up are displayed as median and interquartile range (IQR).

	Stroke (N = 43)	No Stroke (N = 201)	p
Age	69 (SD, 14)	69 (SD, 11)	p = .77
Female sex	18 (42%)	70 (35%)	p = .41
Hypertension	26 (60%)	139 (70%)	p = .96
Diabetes	8 (19%)	30 (15%)	p = .21
Heart failure	7 (16%)	32 (16%)	p = .61
Vascular disease	14 (32%)	41 (21%)	p = .02
Prior stroke	20 (47%)	18 (9%)	p < .001
CHA2DS2-VASc score	4 (IQR, 3)	3 (IQR, 3)	p = .006
Anticoagulation	25 (58%)	127 (63%)	p = .53
Days from MRI to stroke	420 (IQR, 817)	NA	NA
Days follow up	551 (IQR, 824)	663 (IQR, 684)	p = .71

The 10 most important shape parameters from LA and the 10 most important shape parameters from the LAA were considered as candidates for inclusion in the model. These 10 parameters comprised 91.7% of the overall variance in shape within the cohort. Using the LASSO method, none of the LA shape parameters significantly predicted stroke and thus none were included in the final model. However, three LAA shape parameters were found to be predictive of stroke (all p < .05). LAA shape parameters associated with increased risk of stroke demonstrated a broader, shorter and less angulated LAA body, where the LAA tip was superior to the takeoff point from the LA. Examples of patients with low, average, and high risk LAA shapes are shown in Figure 3. These three LAA shape parameters were included in the logistic regression model for prediction of stroke. Coefficients and odds ratios from the model are shown in Table 2. The shape parameters remained significant after they were added to a model including CHA₂DS₂-VASc score (all p < .01). Area under the ROC curve for the model increased from 0.640 without shape parameters to 0.778 with shape parameters (p = .003). Examples of low, average, and high-risk appendage shapes are shown in Figure 2. LAA shape variation by combined shape parameters (Supplemental Figure 2) and by each individual shape parameter (Supplemental Figure 3) and are shown in the supplemental figures.

Figure 3:

ROC curve for prediction of stroke based on LAA shape parameters alone is shown in green (Model 1). ROC curve for prediction of stroke based on CHA₂DS₂-VASc score alone is shown in red (Model 2). ROC curve for prediction of stroke based on CHA₂DS₂-VASc score and LAA shape parameters are shown (Model 3). Shape of the LAA improved prediction of stroke and therefore predicts stroke independent of CHA₂DS₂-VASc score (p = .003).

Table 2:

Results from the three multivariate logistic regression models for prediction of stroke. Model 1 includes LAA shape alone. Model 2 includes CHA₂DS₂-VASc score alone. Model 3 includes both LAA shape and CHA2DS2-VASc score.

Model 1	Coeff	SE	OR	95% CI	P
LAA Shape Parameter 1	0.027	0.007	1.028	1.013–1.042	p < .001
LAA Shape Parameter 2	0.039	0.017	1.040	1.008–1.076	p = .018
LAA Shape Parameter 3	0.061	0.024	1.062	1.015–1.114	p = .011
Model 2
CHA2DS2-VASc score	0.290	0.095	1.34	1.113–1.615	p = .002
Model 3
LAA Shape Parameter 1	0.028	0.007	1.028	1.014–1.044	p < .001
LAA Shape Parameter 2	0.045	0.017	1.046	1.012–1.083	p = .01
LAA Shape Parameter 3	0.065	0.024	1.067	1.018–1.120	p = .008
CHA2DS2-VASc score	0.308	0.098	1.361	1.127–1.660	p = .002

Abbreviations: Coeff = coefficient; SE = standard error; OR = odds ratio; 95% CI = 95 percent confidence interval.

Figure 2:

Individual patient examples of low, average, and high-risk shapes of the left atrium (grey) and left atrial appendage (green) are shown from superior, left and anterior views. The low risk left atrial appendage has a is longer and more angulated, with the tip inferior to the takeoff point from the LA. The high-risk appendage has a broader, shorter body with the tip superior to the LAA ostium.

Discussion

In this study we demonstrate that the shape of the LAA is a strong predictor of stroke in patients with atrial fibrillation when measured quantitatively. While others(3, 17–22) have demonstrated this previously, we are the first to do so using a quantitative measure of shape. Our approach uses SSA, an emerging image analysis technique in medical imaging(1, 2, 11). This technique involves defining a set of surface points to define the surface of each structure. Importantly, these points correspond to a specific anatomic location, consistent from one subject to the next, and are thus termed “correspondence points”. For example, the tip of the LAA is defined by the same correspondence point in each patient. This feature allows for quantitative comparisons between individual patients and groups of patients. The result is a large dataset defining the shape of each patient’s LA and LAA. We then apply statistical methods designed for large datasets to understand the relationship between shape and a clinical outcome. First, we apply PCA to define the shape variability in the cohort. Second, we use the LASSO method to determine which modes of shape variability, which we call “shape parameters”, predict stroke.

We did not demonstrate that LA shape predicts stroke, which contradicts other studies(6, 8). Bisbal et al.(6) demonstrated that patients with prior thromboembolic events had differences in LA sphericity. Although they also qualitatively assessed LAA shape, in our study we assessed shape of the LA and LAA both quantitatively. Changes in LA and LAA shape may be associated. Because both LA and LAA shape parameters were assessed together, this may explain our negative finding for LA shape and stroke. Additionally, the stroke cases in our study included only individuals who had a documented stroke after the index MRI, rather than those with a history of thromboembolism. LA and LAA imaging was obtained prior to the incident stroke in all cases. Shape of the LA may change over time, and thus our study better assesses which shape changes can be used to predict future strokes.

Previous studies of the LAA shape and stroke have been limited to qualitative assessments of shape. Di Biase et al.(3) first proposed categorizing shape of the LAA into four categories: “chicken wing”, “cactus”, “cauliflower”, and “windsock”. They went on to demonstrate the association of the “chicken wing” LAA shape with stroke. Many others have applied this categorization in subsequent studies(7, 17, 19, 23, 24), but some have failed reproduce the results(18, 22). There are three major weaknesses of this qualitative approach. First, the shape categorization proposed may not be the most relevant to the clinical outcome. In fact, they demonstrated that only “chicken wing” morphology was relevant to stroke, and distinguishing between “cactus”, “cauliflower”, and “windsock” did not significantly change stroke risk. Second, the chosen categorization may miss important features of shape that are relevant to stroke risk, as others have suggested(25). SSA assesses shape without making any assumptions about what shape may be relevant in predicting the clinical outcome of interest. Third, despite detailed qualitative descriptions of shape, the categorization process is inherently subjective. This limits reproducibility, which in turn limits clinical utility. Our quantitative method addresses these limitations.

We also demonstrate that the shape of the LAA is a strong predictor of stroke in patients with atrial fibrillation that compares favorably to CHA₂DS₂-VASc score. The area under the receiver operator curve, or C-statistic, in our study was 0.753 using LAA shape alone. This indicates that LAA shape is a strong predictor of stroke. LAA shape compares favorably with the CHA₂DS₂-VASc score which has been found to have a median reported C-statistic of 0.673 in other studies(26). We also demonstrate that LAA shape predicts stroke independent of CHA₂DS₂-VASc score. CHA₂DS₂-VASc score is currently the standard clinical tool for assessment of stroke risk in patients with AF(27). Components of the CHA₂DS₂-VASc score have previously been shown to associate with changes in LA and LAA shape(28). However, studies of qualitative shape assessment of the LAA have either not addressed this question or have failed to show that that these are independent of CHA₂DS₂-VASc score(23). By demonstrating that LAA shape parameters improve prediction of stroke when added to a model with CHA₂DS₂-VASc score, we demonstrate LAA shape could be incorporated to the CHA₂DS₂-VASc score to improve stroke prediction. This could potentially identify patients with low CHA₂DS₂-VASc score who would benefit from anticoagulation or patients with higher CHA₂DS₂-VASc score in whom anticoagulation could be safely withheld. Our study, with its case-control design, was not designed to answer these questions specifically. However, this could be addressed in a future prospective cohort study.

The statistical approach we use to justify our conclusion is complex and differs from those used for qualitative assessments of shape. We use two approaches designed for use in large datasets: PCA and LASSO, which have been used previously in clinical imaging studies by ourselves(1) and others(10). PCA helps us understand how shape varies in the population. We define the PCA modes as “shape parameters”, each of which describes a mode of shape variation in the cohort. These shape parameters may or may not predict stroke, the relevant clinical outcome in our study. LASSO is a method of statistical model building that can be used for variable selection when there are a large number of candidate predictor variables. Using this method shape parameters that predict stroke are selected for inclusion in the model. These steps fall under the broad concept of machine learning. Using our method, we can statistically learn how shape of the LA and LAA varies and how it relates to stroke risk in AF. While more complex, our approach allows us to assess LA and LAA shape without making assumptions about which types of shape will be related to our outcome.

Our study does not explain the mechanism by which shape of the LAA changes stroke risk. Others have hypothesized that changes in LAA shape alter fluid dynamics in the LAA. Altered flow patterns in the left atrium have been directly measured using 4D flow MRI in patients with atrial fibrillation(29). Changes in flow pattern due to shape changes have been simulated in studies using computational fluid dynamics(30, 31). While a relationship between fluid dynamics and shape likely exists, characterization of this relationship is beyond the scope of this study. LA fibrosis, another marker of LA remodeling, has been shown to predict stroke in our population (5). While we have not previously identified a close relationship between fibrosis and shape of the LA or LAA, it’s possible that some relationship exists. Combining LAA shape with other imaging and clinical factors could further improve stroke prediction in AF patients.

The conclusions of our study are limited by its case control design. We demonstrate that LAA shape predicts stroke, but we cannot quantify stroke risk based on LAA shape. Our findings lay the groundwork for a larger study using a prospective cohort design, which would allow for estimation of stroke risk. Our cohort included patients who were referred for MRI prior to AF ablation procedure in most cases. This therefore represents a select cohort that may not be generalizable to all AF patients. Our study uses MRI of the left atrium, which has high spatial resolution. However, other studies have used computed tomography (CT)(3), which has the highest spatial resolution of 3-D medical imaging modalities. The LAA in particular is a complex structure, often with many lobes and trabeculations. MRI may not capture these finer elements of shape. However, given that prior studies have focused on the overall shape of the LAA rather than fine features, the spatial resolution of MRI is more than adequate to capture the relevant shape. One potential advantage of SSA is that it could be applied to any 3D imaging modality. Finally, our process relies on manual image segmentation, which is a subjective and time intensive process. Trained observers can process a segmentation in approximately 10–15 minutes, which may be reasonable for clinical utility. Interobserver variability related to this segmentation process has been previously published(13). Newer, automated image segmentation methods are a promising area of study(32) that could eliminate this subjectivity while minimizing the cost associated with manual segmentation.

Conclusions

We propose a quantitative, SSA approach to quantify shape of the LA and LAA, and demonstrate that shape of the LAA predicts stroke in AF patients. While statistically complex, our approach makes no assumptions about the most relevant shape to stroke and instead learns what LAA shape is associated with increased stroke risk. This improves upon stroke prediction beyond CHA₂DS₂-VASc score alone. This approach may lead to a more reproducible and thus more clinically useful way to use LAA shape to assess stroke risk in AF patients.

List of Abbreviations

LA

left atrium

LAA

left atrial appendage

AF

atrial fibrillation

SSA

statistical shape analysis

MRI

magnetic resonance imaging

PCI

principle component analysis

CT

computed tomography