# Non-invasive prediction of atrial cardiomyopathy characterized by multipolar high-density contact mapping

## TL;DR

This study uses non-invasive echocardiography and machine learning to predict the severity of atrial cardiomyopathy in patients with atrial fibrillation.

## Contribution

The novel use of a multipolar mapping catheter and machine learning to identify non-invasive markers for atrial cardiomyopathy severity.

## Key findings

- Machine learning accurately predicted mild and severe atrial cardiomyopathy subgroups with an AUC of 0.9.
- Echocardiographic parameters like LA reservoir strain and LAVI/a′ ratio differed significantly between mild and severe AC groups.
- Severe AC was associated with higher AF recurrence rates (40.9% vs. 10.7%) at 12 months.

## Abstract

Atrial cardiomyopathy (AC) establishes links between atrial fibrillation (AF), left atrial (LA) mechanical dysfunction, structural remodeling, and thromboembolic events. Early diagnosis of AC may impact AF treatment and stroke risk prevention. Modern endocardial contact-mapping provides high-resolution electro-anatomical (EA) maps of the LA, thus allowing to display the myocardial substrate based on impaired signal amplitude and to characterize AC. Correlation of invasively assessed AC using a novel, multipolar mapping catheter (OCTARAY™, Biosense Webster, limited market release) and LA echocardiographic parameters could form the basis for a set of echo parameters for non-invasive prediction of AC.

We retrospectively identified 50 adult patients who underwent primary pulmonary vein isolation (PVI) for paroxysmal or persistent AF between 08/22 and 05/23 fulfilling the selection criteria: (i) EA mapping with a novel multipolar mapping catheter (Octaray®); (ii) acquisition of voltage maps in sinus rhythm (SR) with ≥ 5000 points/map; and (iii) transthoracic echocardiography acquired in SR ≤ 48 h before PVI. Exclusion criterion was previous LA ablation. We generated EA maps with two sets of upper voltage thresholds (0.2–0.5 mV and 0.2–1.0 mV) and assessed total LA low voltage area (LVA). As LVA thresholds for the classification of AC are not yet established, an unsupervised machine learning cluster analysis was performed using a Gaussian mixture model (GMM), and two groups of patients with mild and severe AC were identified. Based on these two groups, we selected echo parameters for further analysis by applying the Boruta algorithm. The predictive capacity of the selected parameters was evaluated using a support vector machine.

The mean age of the studied sample (n = 50) was 63 ± 11 years, 62% were men, 64% showed persistent AF, median CHA2DS2-VASc score was 2 (quartiles 1, 3), and NT-proBNP was 190 (71, 391) pg/ml. A median of 5771 (5217, 6988) points/map were acquired. GMM yielded clusters of mild AC (n = 28) and severe AC (n = 22). Median LVA was 0.6 cm2 (< 0.5 mV) resp. 4.1 cm2 (< 1.0 mV) in group mild AC and 6.9 cm2 (< 0.5 mV) resp. 27.2 cm2 (< 1.0 mV) in group severe AC. Several echocardiographic parameters differed between the groups of mild and severe AC: dynamic LA parameters (end diastolic LA reservoir strain: 24.5% (22, 29) vs 15% (12, 19), p < 0.001; LA reservoir strain at atrial contraction: 22% (19, 25) vs 15% (11, 18), p < 0.001, end diastolic LA contraction strain: 13% (8, 15) vs 7.5% (3, 13), p < 0.01) as well as LA end-systolic volume index to a´ ratio (LAVI/a′: 297 (231,365) vs 510 (326,781), p < 0.01). Consistent distribution of NT-proBNP (mild AC: 125 (48,189) pg/ml, severe AC: 408 (254,557) pg/ml, p < 0.0001) and CHA2DS2-VASc score (mild AC: 1 (1–2), severe AC: 3 (3–4), p < 0.0001) served as proof of concept. Applying the selected echocardiographic parameters, the machine learning algorithm correctly identified both subgroups with a mean AUC of 0.9 (95% CI 0.83–0.99). At 12 months, AF recurrence rate was 10.7% in mild AC and 40.9% in severe AC (p < 0.05).

Among patients qualifying for PVI, machine learning analysis of high-resolution LA maps allowed to identify subgroups with mild and severe AC avoiding the use of arbitrary LVA thresholds. The subgroups were predicted non-invasively with good accuracy using a machine learning approach that incorporated a set of echocardiographic markers. This data could advance the clinical triage of patients with AF.

The online version contains supplementary material available at 10.1007/s10840-025-02001-2.