Risk prediction of new AF: is there a role for artificial intelligence?

Atrial fibrillation (AF) is the most common cardiac arrhythmia encountered in clinical practice.1 AF conveys significant health risks, most notably from a fivefold increased risk of stroke.2 Strokes attributable to AF are of a greater severity, cause greater disability and mortality, and result in greater healthcare costs than non-AF related strokes.3 Approximately 40% of people with AF are asymptomatic and are described as having ‘silent’ AF.4 It is therefore unsurprising that over 12% of people are found to have AF at the first presentation of a stroke. Early identification and treatment of AF before the occurrence of stroke is, therefore, of paramount importance to improve population health. AF screening has been recommended but is yet to be implemented. Recommendations currently suggest screening in patients aged 65 years or older5 but the yield of interventions for detecting new AF varies considerably according to the age of populations screened and the interventions tested.6 Identifying those at high risk of developing new AF, including those with ‘silent AF’, would enable better targeting of screening interventions and consequently has the potential to improve the effectiveness and cost-effectiveness of future AF screening programmes. A variety of clinical, electrocardiographic (ECG) and biochemical markers have been identified as risk factors for the development of new AF. From these risk factors a number of risk prediction models have been developed (Table 1). While all studies have utilised large patient datasets to derive and validate risk prediction models, only some have been developed using unselected primary care populations and all have been developed in non-UK populations. Studies of risk prediction models for the development of new AF. AF: atrial fibrillation; ECG: electrocardiogram; BMI: body mass index; BP: blood pressure; SBP: systolic blood pressure; MI: myocardial infarction; LVH: left ventricular hypertrophy; IHD: ischaemic heart disease; PAC: premature atrial complexes; TIA: transient ischaemic attack; COPD: chronic obstructive pulmonary disease. Studies of risk prediction models for the development of new AF. AF: atrial fibrillation; ECG: electrocardiogram; BMI: body mass index; BP: blood pressure; SBP: systolic blood pressure; MI: myocardial infarction; LVH: left ventricular hypertrophy; IHD: ischaemic heart disease; PAC: premature atrial complexes; TIA: transient ischaemic attack; COPD: chronic obstructive pulmonary disease. Existing models only have moderate abilities to predict new AF and do not evaluate the incremental effects of different risk predictors for AF. Furthermore, risk models predominantly use clinical characteristics to predict the occurrence of new AF and, of those that use ECG parameters, only a few ECG markers have been utilised. No models have incorporated biochemical markers for AF risk prediction. There remains an opportunity to improve the predictive abilities and risk stratification against AF outcomes by combining clinical, ECG and biochemical parameters. Research suggests that existing risk prediction models may not be transferrable across populations. The US derived CHARGE-AF prediction tool was applied to the UK EPIC Norfolk cohort (n = 24,020). Although CHARGE-AF was found to have reasonable discrimination (C-statistic (95% confidence interval (CI) 0.81) 0.75–0.85) there was weak calibration of this model with a nearly twofold overestimation of AF incidence.7 Therefore, an important consideration for future research would be to develop and validate prediction models using data from the target population. Existing risk prediction models have been developed using epidemiological methods which can be improved on using newer techniques, such as ‘artificial intelligence’ (AI) or ‘machine learning’. Previous approaches to developing risk prediction models rely on statistical modelling of ‘survival rates’ associated with patient baseline characteristics. These methods assume linearity in relationships between individual characteristics on outcomes (e.g. the relationship between increasing age and developing new AF is linear) and linearity in the effect of multiple characteristics on outcomes (e.g. the increased risk of developing AF from the combined effects of hypertension and increased age is linear). These assumptions may be inaccurate and novel approaches to determining risk prediction, such as AI technology, may overcome such limitations. Machine learning is computer-based learning using AI technology to derive and enable pattern recognition within routinely collected clinical data.8,9 Machine learning could be used to analyse integrated healthcare datasets better for the derivation of better clinical evidence as it can overcome the assumptions of linearity when modelling multiple interacting risk factors. Machine learning was recently found to improve the accuracy of predicting all-cause mortality than conventional risk factor modelling in a prospective cohort of patients (n = 502,628) from UK-Biobank.10 Moreover, machine learning enables pattern recognition within data with a potential to identify new risk factors for disease that may previously have been unrecognised. With the rapid emergence of AI technology, researchers are increasingly questioning if there is an opportunity to improve the methodological interrogation of routinely collected patient data to improve healthcare outcomes. With regard to AF, should researchers be considering who may develop AF in the future to enable better targeting of AF detection interventions? There is now an opportunity to use technological advances to identify new risk factors for AF within large patient datasets and to develop tools for predicting those at greatest risk of developing AF. Indeed, primary care systems currently have risk prediction tools embedded within clinical systems (e.g. Q-Risk) so translation of such research into clinical practice is no longer a barrier. The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article. The author(s) received no financial support for the research, authorship, and/or publication of this article.