Digital Pathways to Innovate Atrial Fibrillation Care with Photoplethysmography and Deep Learning

Abstract

Digital health is transforming the landscape of medical diagnostics and patient care. It is an umbrella term used to describe a range of technologies including, electronic medical records, telemedicine, wearable devices, mobile health applications and artificial intelligence (AI).(1) These technologies serve various purposes, including disease detection, aiding patient treatment, and managing health information. These technologies are reshaping how health care is delivered and managed. Despite the anticipated benefits of these technologies for patient outcomes, the evidence linking digital health technologies to improved clinical outcomes remains sparse. However, the absence of evidence is not the evidence of absence. Many of these technologies were developed to address medical problems that are common and affect clinical outcomes. The onus lies on physicians and researchers to find out how to apply these technologies in a way that benefits the patient. Atrial fibrillation (AF) is the most common heart rhythm disorder and it is a major risk factor for stroke, heart failure, and other cardiovascular complications.(2) An average person’s lifetime risk to develop AF is approximately 25%.(3) Currently, the prevalence of AF in Europe is estimated between 2% and 4%, but is expected to double from 2010 to 2060, primarily owing to population aging and the increasing prevalence of AF risk factors, leading to a substantial economic and public health burden.(4) A diagnosis of AF is associated with a 5-fold increased stroke risk and 2-fold increased risk of all-cause mortality.(5) Early diagnosis and appropriate treatment, particularly with oral anticoagulation in individuals at high risk for stroke or systemic embolism, may mitigate this substantial morbidity and mortality. Because AF often remains asymptomatic and AF episodes may be of short duration, early detection with traditional methods is challenging.(6,7) Digital health technologies offer an innovative solution to these challenges. Increasingly more devices, marketed directly to the consumer, such as smartphones, smartwatches and fitness trackers, are equipped with heart rate monitoring technology, mostly based on electrocardiography (ECG) or photoplethysmography (PPG).(8) Whereas ECG requires specialized hardware to measure voltages, PPG is an optical technology that can be derived from a digital camera.(9–11) Therefore, a PPG-based application can turn virtually every smartphone in to a rhythm monitoring device, making the technology widely available.(12) However, challenges remain. According the 2020 European Society of Cardiology (ESC) guidelines for the diagnosis and management of AF, the diagnosis requires an ECG showing irregular R-R intervals (when atrioventricular conduction is not impaired) and the absence of distinct repeating P waves.(4) This definition does not include a diagnosis with PPG. Therefore, the validation of PPG for the diagnosis of AF was identified as one of the gaps in evidence and is the first challenge addressed in the thesis. Most of the existing evidence regarding AF applies to clinical AF which is defined as AF that is documented on a surface ECG. (13–15) Asymptomatic AF, detected through screening by novel digital technologies, is referred to as subclinical AF.(16) It is unsure whether the existing evidence on clinical AF can be transferred to subclinical AF. Hence, the 2020 ESC guidelines and the 2022 US Preventive Services Task Force highlighted the insufficient evidence to recommend universal screening for AF, underscoring the importance of physician judgment in screening decisions.(4,17) Whereas digital devices could democratize AF screening, the effectiveness of digital AF screening, impact on AF management, and subsequently on AF outcomes, remains unknown.(18) This is the second challenge addressed in this thesis. The benefit from screening largely depends on the balance between the benefit of stroke reduction and harm imposed by the bleeding risk from the treatment of screening-detected AF.(19–21) Among the scores that stratify the thromboembolic risk in patients with AF, the CHA2DS2-VASc score is the most widely used in clinical practice due to its simplicity.(22,23) This score has been validated for the prediction of stroke in AF, as well as for the prediction of AF.(24) The dual function of the CHA2DS2-VASc risk score results from the considerable overlap of risk factors for AF and for stroke. Hypothetically, targeted screening to the CHA2DS2-VASc risk score can identify individuals who are more likely both to display AF upon screening and simultaneously to benefit from treatment. However, the benefit on clinical outcomes in studies that followed this hypothesis was marginal.(19,21,25,26) Therefore better patient characterization is needed beyond clinical risk factors. Digital health applications leveraging deep learning may improve patient phenotyping and identify patients that could benefit from screening. This is the third challenge addressed in this thesis. Specific situations, such as the perioperative phase, are known to increase the likelihood of AF. These situations are associated with inflammation and an increased adrenergic tone, thereby acting as a ‘AF stress test’ on the heart, revealing subjects with an atrial substrate vulnerable to the induction and maintenance of AF.(27) For example, AF occurs in 30% to 50% of patients after cardiac surgery, known as postoperative atrial fibrillation (POAF).(28) POAF is known to be associated with a 2-fold increase in early adverse outcomes such as stroke and mortality, which persists during long-term follow-up.(28,29) POAF is well described during the hospitalization phase. However, there is paucity of data on POAF after the hospitalization phase. The problem, until recently, was that POAF monitoring after the hospitalization phase was cumbersome with either short duration ECG monitoring or invasive devices.(30–32) With digital health technology, POAF can be monitored after the hospitalization phase with noninvasive ECG- or PPG-based devices.(33) However, the effect of home-based POAF monitoring with digital devices has not been studied and the effect on AF management and subsequent clinical outcomes remains unknown. This is the fourth challenge addressed in this thesis. Besides stroke prevention with OAC, the management of AF includes rate and rhythm control to reduce symptoms and improve quality of life.(4,34,35) Early rhythm control is associated with better clinical outcomes.(36) AF ablation, through pulmonary vein isolation (PVI), is the most effective treatment for the maintenance of sinus rhythm,(4,37) and is superior to antiarrhythmic drug (AAD) therapy in the management of AF.(38) However, AF recurrence after ablation is common.(35,39) Monitoring recurrence is essential for assessing treatment effectiveness, atrial arrhythmia burden and attributing patient symptoms to an atrial arrhythmia. The conventional monitoring approach relies on periodic (longterm) ECGs and is constrained by limited monitoring time, measurement dispersion, and cost.(40) The risk of AF recurrence after ablation can be predicted prior to the procedure with clinical risk scores (e.g. APPLE score).(41) However, these algorithms are impractical and their use in clinical practice is limited.(42) The risk of AF recurrence after ablation is increased if low-voltage areas (LVA) (atrial fibrosis) is present.(43) In such cases, additional ablation of these areas has been suggested.(44–46) While LVA can only be measured invasively with electroanatomical voltage mapping, the presence of LVA can be predicted using clinical risk scores or cardiac magnetic resonance imaging. (41,47–49) However, the predictive performance is limited and both are impractical. Digital health applications may overcome the current challenges associated with AF monitoring after ablation using novel rhythm monitoring techniques. Additionally, deep-ECG learning may improve patient phenotyping to predict the presence of LVA, and ultimately improve the selection of ablation techniques. This is the fifth challenge addressed in this thesis. The recent advances in technology combined with the need to manage patients remotely during the coronavirus disease-19 (COVID-19) pandemic, have led to a rapid adaptation of the use of digital devices in clinical practice.(50,51) The increased connectivity of data and availability of monitoring devices may enable more patients to be treated at home instead of in the hospital. This concept proved particularly appealing during the covid pandemic that caused a tremendous overload on the healthcare system in nearly every country.(52) This concept was addressed as a sixth challenge in this thesis. This doctoral dissertation aims to explore the use of digital health technologies to improve patient care. This thesis will focus on PPG-based digital health applications, generating novel insights in the clinical application of this technology to address the outlined challenges and improve AF management. Additionally, this thesis will focus on ECG-based deep learning applications, gaining novel and mechanistic insights in digital patient phenotyping to address the outlined challenges and personalize AF management plans.