Predicting atrial fibrillation in primary care using machine learning.

Nathan R Hill,Alexander T Cohen,Matthew Lumley,Jason Gordon,Daniel M Sugrue,Daniel Ayoubkhani,Steven Lister,Mark O’Neill,Usman Farooqui,David Clifton,Ameet Bakhai,Phil Mcewan,Andres Azuero

doi:10.1371/journal.pone.0224582

Abstract

BackgroundAtrial fibrillation (AF) is the most common sustained heart arrhythmia. However, as many cases are asymptomatic, a large proportion of patients remain undiagnosed until serious complications arise. Efficient, cost-effective detection of the undiagnosed may be supported by risk-prediction models relating patient factors to AF risk. However, there exists a need for an implementable risk model that is contemporaneous and informed by routinely collected patient data, reflecting the real-world pathology of AF.MethodsThis study sought to develop and evaluate novel and conventional statistical and machine learning models for risk-predication of AF. This was a retrospective, cohort study of adults (aged ≥30 years) without a history of AF, listed on the Clinical Practice Research Datalink, from January 2006 to December 2016. Models evaluated included published risk models (Framingham, ARIC, CHARGE-AF), machine learning models, which evaluated baseline and time-updated information (neural network, LASSO, random forests, support vector machines), and Cox regression.ResultsAnalysis of 2,994,837 individuals (3.2% AF) identified time-varying neural networks as the optimal model achieving an AUROC of 0.827 vs. 0.725, with number needed to screen of 9 vs. 13 patients at 75% sensitivity, when compared with the best existing model CHARGE-AF. The optimal model confirmed known baseline risk factors (age, previous cardiovascular disease, antihypertensive medication usage) and identified additional time-varying predictors (proximity of cardiovascular events, body mass index (both levels and changes), pulse pressure, and the frequency of blood pressure measurements).ConclusionThe optimal time-varying machine learning model exhibited greater predictive performance than existing AF risk models and reflected known and new patient risk factors for AF.

Highlights

Analysis of 2,994,837 individuals (3.2% Atrial fibrillation (AF)) identified time-varying neural networks as the optimal model achieving an area under the receiver operating characteristic curve (AUROC) of 0.827 vs. 0.725, with number needed to screen of 9 vs. 13 patients at 75% sensitivity, when compared with the best existing model CHARGEAF
Atrial fibrillation (AF), the most common sustained heart arrhythmia[1], is associated with an approximately five-fold increase in stroke[2] and an increase in stroke severity compared to non-AF patients, resulting in higher morbidity and mortality[3, 4]
In the first stage of model building, we evaluated different baseline models including logistic least absolute shrinkage and selector operator (LASSO)[17], random forests[18], support vector machines[19], neural networks (NN)[20], Cox regression and published AF risk models (Framingham[13], ARIC[12] and CHARGE-AF[11])

Summary

Introduction

Atrial fibrillation (AF), the most common sustained heart arrhythmia[1], is associated with an approximately five-fold increase in stroke[2] and an increase in stroke severity compared to non-AF patients, resulting in higher morbidity and mortality[3, 4]. In England alone, 425,000 people are estimated to be living with undiagnosed AF[6] Given that these patients are at increased risk of stroke-related death or disability,[3, 4] early detection and effective management of AF have the potential to improve patient outcomes and alleviate the economic burden of AF and its sequelae. Both European and US guidelines recommend diagnosing AF based on a 12-lead electrocardiogram (ECG) or rhythm strip[7, 8]; 12-lead ECG use for AF detection in primary care has been reported not to be cost-effective, regardless of whether used in a systematic (e.g. patients >65 years) or targeted (including high-risk patients only) approach[9]. There exists a need for an implementable risk model that is contemporaneous and informed by routinely collected patient data, reflecting the real-world pathology of AF.

Objectives

Methods

Results

Conclusion