Identification of new atrial fibrillation phenotypes from insertable cardiac monitor recordings using unsupervised machine learning
E Flo Lenz, A Ostvik, A B Sellevold, B M Nes, J P Loennechen, J M LetnesAbstract
Introduction
Current classification of atrial fibrillation (AF) (paroxysmal/persistent) fails to capture clinically important differences. Continuous rhythm monitoring enables detailed assessment of individual AF behaviour, creating an opportunity for data-driven characterisation of AF and linking patterns to clinical profiles and outcomes.
Purpose
To derive data-driven AF phenotypes from longitudinal rhythm monitoring using machine learning and relate them to patient characteristics, cardiac structure/function and clinical outcomes.
Methods
We used data from a 12-month randomized (exercise intervention or usual care) multicentre trial, where all participants received an insertable cardiac monitor (ICM) ≥4 weeks prior to baseline. Baseline and 12-month data comprised AF-related outcomes, patient characteristics and echocardiography. Patient-level AF behaviour derived from ICM episodes (burden, ventricular rate, temporal distribution, and longitudinal episode dynamics) were aggregated into 14 curated non-redundant features and subjected to unsupervised clustering (K-means) to identify phenotypic subgroups. Cluster stability and optimal number were assessed using established cluster quality indices across multiple candidate solutions with varying cluster numbers and clustering methods.
Results
K-means clustering of ICM-derived AF features across 246 participants identified four phenotypes (P0–P3) with distinct AF behaviour. P0 had high AF burden, not explained by a higher episode count alone. P1 had low burden and few episodes, characterised by rapid ventricular response, mean >110 beats per minute. P2 displayed low-to-moderate burden driven by frequent, short episodes, whereas P3 had very low burden despite slightly longer mean episode duration. Clinically, phenotypes differed significantly in cardiorespiratory fitness, NT-proBNP, indexed left atrial end-systolic volume, the Atrial Fibrillation Effect on Quality of Life (AFEQT) symptom score and AF-related hospitalisations. P0 had the highest hospitalisation rate, poorest fitness and largest left atria, consistent with a more advanced substrate. P1 had the highest fitness and smallest atria yet reported a high symptom score despite low AF burden. P2 resembled P0 in clinical profile, but experienced higher hospitalisation rate and symptoms relative to its moderate burden, indicating that episode frequency may be clinically important beyond burden alone. P3 reported few symptoms and had low hospitalisation rates. Notably, phenotypes stratified AF-related hospitalisations even within paroxysmal AF.
Conclusion
Unsupervised machine learning applied to continuous ICM recordings identifies distinct, clinically meaningful AF-phenotypes that differ in AF behaviour and are associated with symptoms, clinical characteristics and AF-related hospitalisations. These findings support further phenomapping by AF episode characteristics as a step towards improved AF classification.Figure 1