Predicting the risk of atrial fibrillation: is the machine learning the answer?

This editorial refers to ‘Detecting undiagnosed atrial fibrillation in UK primary care: validation of a machine learning prediction algorithm in a retrospective cohort study’, by S. Sekelj et al., pp. 598–605. Atrial fibrillation (AF) is the most common arrhythmia, particularly in the elderly.1 However, the true prevalence of the condition is unknown taking into account that paroxysmal and asymptomatic AF can be difficult to detect in general clinical practice; consequently, AF is often diagnosed incidentally or following a cardioembolic stroke. Therefore, an early detection and effective management of AF is crucial to improve patient outcomes and alleviate the economic burden of AF. Electrocardiography is the only effective method to diagnose AF but a standard 12-lead electrocardiogram is not sufficient to detect paroxysmal AF and longer monitoring such as event recorder electrocardiography is impractical and not cost-effective. Taking into account these considerations there is an urgent need in identifying patients at high risk for AF that should undergo further examinations to diagnose the arrhythmia while maintaining cost-effectiveness. The use of machine learning (ML) for predicting disease instead of traditional risk score for AF prediction may enable physicians to speed-up decisions making and improve efficiency, reliability, and accuracy. Before discussing the applications of ML in AF screening it is appropriate to explain briefly what ML is and its potential applications in the health care system. ML is a form of artificial intelligence (AI) broadly defined as the ability of a machine to autonomously acquire knowledge by extracting patterns from large data sets. While ML with the explosion of big data is nearly ubiquitous in all sectors of technology, from speech recognition to autonomous driving its introduction into the medical field has been much more controversial. However, in the last years, the huge increase of electronic health records containing different types of digitalized medical data and new tools to analyse them in an efficient way, offer tremendous potential to enhance medical research and clinical care.2 ML algorithms have recently been successfully employed to classify skin cancer using images with comparable accuracy to a trained dermatologist or in detecting known familial hypercholesterolaemia genetic mutations providing a ‘virtual’ genotyping of patients with this condition.3 There are essentially two types of learning techniques employed in ML: supervised and unsupervised learning. In supervised learning, a computer learns directly from of a large quantities of features which are associated with a known outcome, this process is known as ‘training the algorithm’. Once the algorithm is successfully trained, it is applied to the testing dataset without their associated outcomes. The predictions made by the algorithm are then compared to the known outcomes of the testing dataset to establish model performance. A common example of supervised learning in clinical cardiology is the prediction of either cardiovascular or all-cause mortality in patients free from cardiovascular disease.4 In contrast, in the unsupervised learning, there isn’t any pre-defined human input and the algorithm attempts to autonomously derive conclusions from unlabelled data. In the following section, regarding ML in AF, we focus our attention primarily on supervised ML, which to date has found the widest applicability. As discussed in the introduction, AF is notoriously difficult to screen because it is often transient and asymptomatic. A variety of clinical, electrocardiographic, and biochemical markers have been identified as risk factors for the development of new AF and from these risk factors, a number of risk prediction models have been developed. At present existing risk prediction tools include the Framingham, Atherosclerosis Risk In Communities (ARIC) and Cohorts for Aging and Research in Genomic Epidemiology (CHARGE-AF) models; however, all of these were developed in the USA and may not be transferrable across populations. Furthermore, these models have been created using epidemiological methods assuming linearity in relationships between certain individual characteristics and the outcomes but this assumption may be inaccurate and it can be improved using supervised ML.5 In a large-scale retrospective analysis using data from almost 3 million individuals with no history of AF in the UK, an ML algorithm achieved an area under the curve of 0.83 for AF prediction as opposed to 0.73 with the CHARGE-AF score and 0.70 with logistic regression analysis.6 However, although these results are promising, the algorithm proposed by Hill et al.6 has not yet been applied to other dataset and it is unknown how well the model will perform with different population-based data. In order to overcome such limitation, in this current edition of the European Journal of Preventive Cardiology, Sekelj et al.7 externally validated the algorithm using coded primary care data from the DISCOVER dataset, which is one of Europe’s largest patient-level datasets, containing data from approximately 2.5 million patients across North West London (NWL) at any given time. In brief, the authors showed that compared with the development study the algorithm displayed a good discriminative performance in distinguishing between patients with AF and those without AF (AUROC 0.87) in the DISCOVER database. Among patients aged ≥65 years, the discriminatory power was slightly weaker but was nevertheless acceptable (AUROC 0.71). Although the present study confirmed the performance of the ML-based algorithm proposed by Hill in identifying patients at highest risk of AF there are some limitation that should be highlighted: results from this study are not generalizable to patient populations in other regions of the UK or other countries, the relationships between risk factors and their importance may vary in patients aged ≥30 years compared with patients aged ≥65 years so the algorithm would need to be recalibrated for a new population (patients aged ≥ 65 years). The question in how to apply the ML algorithms to different populations is a crucial point in the development of ML techniques keeping in mind that the algorithms need to be revalidated or adapted in this case. Siontis et al. in a recent paper suggested the utilization of blockchain technology to create a decentralized marketplace to share in a secure and traceable way a large amounts of patient data across institutions for the retraining and testing AI algorithm. Although the integration of blockchain technology with AI in cardiovascular medicine could facilitate secure integration of data from healthcare platforms one potential ethical concern regards the monetizing of the medical data behind data-sharing based on blockchain technology.8 The growing diffusion of smart devices capable to monitor physical activity, sleep, resting and heart rate variability, weight and oxygen saturation may provide an opportunity to refine AF prediction through ML algorithm. In this context, the development of smartphone application that collects and syncs data from electronic health record, pharmacies, and personal devices like activity tracker can be used to create and test ML algorithm for AF prediction. Moreover, the ability of ML to analyse a large amount of data would allow us to refine AF prediction model including additional risk factors such as psychological distress that is associated with an increased risk of incident AF.9 One of the major limitation of the risk models used for AF prediction is that they predict the mid-to-long-term risk of AF, for example, 5 years or 10 years, rather than the contemporaneous risk of unrecognized AF. To overcome this limitation a group of Mayo Clinic investigators developed an ML model to predict AF based on the electrocardiographic signature of AF invisible to human eye present during sinus rhythm.10 Data from implantable cardiac monitors have demonstrated that years of monitoring may be required to detect clinically significant AF; however, the great spread of wearable sensors capable to measure heart rate provide an opportunity for AF detection. In this context, a recent paper by Tison et al.11 showed that a commercially available Smartwatch can passively detect AF using a readily available mobile application based on an ML algorithm. Conflict of interest: none declared. The opinions expressed in this article are not necessarily those of the Editors of the European Journal of Preventive Cardiology or of the European Society of Cardiology.