AI-based derivation of atrial fibrillation phenotypes in the general and critical care populations

Abstract

Background: Atrial fibrillation (AF) is the most common heart arrhythmia worldwide and is linked to a higher risk of mortality and morbidity. To predict AF and AF-related complications, clinical risk scores are commonly employed, but their predictive accuracy is generally limited, given the inherent complexity and heterogeneity of patients with AF. By classifying different presentations of AF into coherent and manageable clinical phenotypes, the development of tailored prevention and treatment strategies can be facilitated. In this study, we propose an artificial intelligence (AI)-based methodology to derive meaningful clinical phenotypes of AF in the general and critical care populations.

Methods: Our approach employs generative topographic mapping, a probabilistic machine learning method, to identify micro-clusters of patients with similar characteristics. It then identifies macro-cluster regions (clinical phenotypes) in the latent space using Ward's minimum variance method. We applied it to two large cohort databases (UK-Biobank and MIMIC-IV) representing general and critical care populations.

Findings: The proposed methodology showed its ability to derive meaningful clinical phenotypes of AF. Because of its probabilistic foundations, it can enhance the robustness of patient stratification. It also produced interpretable visualisation of complex high-dimensional data, enhancing understanding of the derived phenotypes and their key characteristics. Using our methodology, we identified and characterised clinical phenotypes of AF across diverse patient populations.

Interpretation: Our methodology is robust to noise, can uncover hidden patterns and subgroups, and can elucidate more specific patient profiles, contributing to more robust patient stratification, which could facilitate the tailoring of prevention and treatment programs specific to each phenotype. It can also be applied to other datasets to derive clinically meaningful phenotypes of other conditions.

Funding: This study was funded by the DECIPHER project (LJMU QR-PSF) and the EU project TARGET (10113624).

Keywords: Atrial fibrillation; Clinical phenotypes; Clustering; Generative topographic mapping; MIMIC-IV; Machine learning; Probabilistic modelling; Stratification; UK-Biobank.

Fig. 1

Proposed AI-based methodology to generate reliable phenotypes. Data is modelled by the GTM algorithm, which projects the data into a 2-dimensional latent space, visualised in the membership map. The GTM also produces reference maps, which are used to indicate the influence of a variable over a micro-cluster. Hierarchical clustering is then applied to the reference vectors to group similar micro-clusters together into larger macro-clusters, which in turn are used to derive the phenotypes.

Fig. 2

Reference vector visualisations demonstrating how each biological sample variable affects the cluster distribution in the latent space for both, the UK-Biobank and the MIMIC-IV AF cohorts.

Fig. 3

Membership maps showing how a selection of investigative variables data are distributed within the latent space for the UK-Biobank and the MIMIC-IV cohorts. AF: Atrial Fibrillation. AKI: Acute Kidney Injury. ARDS: Acute Respiratory Distress Syndrome. GCS: Glasgow Coma Scale.

Fig. 4

Derived phenotypes of AF in the general population using UK-Biobank data. a) Dendrogram produced using Ward’s minimum variance method. The graph shows the 5 clusters that are used to define the 5 AF phenotypes for the general population. b) Membership map with a uniform size for the micro-clusters to show the distribution of the macro-cluster regions. c) The size of the micro-clusters in the membership map dictated by the number of participants assigned to it. d) Main characterising features for each of the phenotypes.

Fig. 5

Derived phenotypes of AF in the general population using MIMIC-IV data. a) Dendrogram produced using Ward’s minimum variance method. The graph shows the 4 clusters that are used to define the 4 AF phenotypes for ICU patients. b) Membership map with a uniform size for the micro-clusters to show the distribution of the macro-cluster regions. c) The size of the micro-clusters in the membership map dictated by the number of participants assigned to it. d) Main characterising features for each of the phenotypes.

Fig. 6

Membership map with the probability distributions for different data points superimposed. Maps a) and b) show the probability distribution for two randomly selected participants from the general population taken from the UK Biobank database. Maps c) and d) show the probability distribution for two randomly selected patients from the critical care population taken from the MIMIC-IV database.