Perspectives on Sex- and Gender-Specific Prediction of New-Onset Atrial Fibrillation by Leveraging Big Data.

Atrial fibrillation (AF), the most common sustained cardiac arrhythmia, has a large impact on quality of life and is associated with increased risk of hospitalization, morbidity, and mortality. Over the past two decades advances regarding the clinical epidemiology and management of AF have been established. Moreover, sex differences in the prevalence, incidence, prediction, pathophysiology, and prognosis of AF have been identified. Nevertheless, AF remains to be a complex and heterogeneous disorder and a comprehensive sex- and gender-specific approach to predict new-onset AF is lacking. The exponential growth in various sources of big data such as electrocardiograms, electronic health records, and wearable devices, carries the potential to improve AF risk prediction. Leveraging these big data sources by artificial intelligence (AI)-enabled approaches, in particular in a sex- and gender-specific manner, could lead to substantial advancements in AF prediction and ultimately prevention. We highlight the current status, premise, and potential of big data to improve sex- and gender-specific prediction of new-onset AF.

Introduction

Atrial fibrillation (AF), the most common cardiac arrhythmia, markedly increases the risk of hospitalization, morbidity, and mortality (1–3). Over the last two decades advances regarding the clinical epidemiology and management of AF have been established (1–3). Recent evidence indicates that both sex and gender play a role in the development and progression of cardiovascular disease, drug reactions, and healthcare utilization (4–7). While sex refers to the biology; including chromosomes, gene expression, hormone levels, and their function, gender comprises the socio-cultural attributes; including socially constructed roles, behaviors, expressions, and identities (4–7). In the field of AF, however, sex and gender implications remain understudied (2, 8–10). Notably, the age-adjusted prevalence, incidence, and lifetime risk of AF are higher in men than in women (1–3). More specifically, it is suggested that several risk factors (hypertension, smoking, alcohol intake, obesity, history of diabetes mellitus, history of myocardial infarction or history of heart failure), carry a differential impact on AF risk in men and women (1, 8–11). It has also been suggested that AF-related adverse outcomes and response to various treatment modalities differ between men and women (1–3, 8–11). Nevertheless, AF remains to be a complex and heterogeneous disorder and a comprehensive sex- and gender-specific approach to predict new-onset AF is lacking.

The past decade has witnessed an exponential growth in recorded data in the healthcare sector. The massive amount of recorded information, i.e. big data, has turned to a topic of special interest, because of its great potential. Leverage of big data, using artificial intelligence (AI)-enabled approaches, provides an opportunity to further improve prediction of AF (12, 13). Use of various sources of big data such as electrocardiograms (ECGs), electronic health records (EHRs), and wearable devices, in particular in a sex- and gender-specific approach, could lead to substantial advancements in AF prediction and ultimately prevention. Here, we review the current status and highlight the premise and potential of these data sources to improve sex- and gender-specific prediction of new-onset AF.

Leveraging Big Data for Prediction of New-onset Atrial Fibrillation

Over the past decade, several AF risk prediction scores have been developed and validated using more traditional research methods (14–22). These prediction scores predominantly use traditional cardiovascular risk factors such as age, sex, race, height, weight, hypertension, diabetes mellitus, coronary heart disease, and heart failure which are readily obtainable clinical variables (14–22). In general, these scores have a moderate to good performance (C-statistic ranging between 0.65–0.78) (14–22). Over the past years, there has been an exponential interest in using various big data sources to further improve the AF risk prediction beyond the traditional AF risk factors (23, 24). Specifically, multiple studies have employed AI-enabled algorithms to evaluate new-onset AF prediction by leveraging various big data modalities including the clinical data, ECGs, EHRs, and wearable devices (23–42). Some of these studies showed that AI-enabled AF prediction models performed similar to or better than established traditional AF prediction models (25, 27–30). Furthermore, targeted AF screening using a machine learning (ML) risk prediction algorithm showed the potential to enhance AF screening and to improve the cost-effectiveness of AF screening through an efficient use of limited healthcare resources (31).

A variety of studies have highlighted the potential predictive capacity of AI to assess the risk of new-onset AF from a 12-lead ECG with acceptable to excellent performance (area under the receiver operating characteristic curve ranging between 0.70–0.90) (32, 33). The potential of the ECG to predict AF might be explained by the fact that the AF substrate is caused by electrical and structural remodeling of the heart (2). AI-enabled algorithms for ECG assessment could mark the very early stages of remodeling, not yet being detected by the cardiologist using routine measures, to predict new-onset AF. However, previous studies mainly predicted new-onset AF risk within a short time period (<1 year) (33). Thereby, their value for primary prevention is limited, as the time window may not be long enough to intervene in individuals at high AF risk. In a recent study (32), a convolutional neural network for 10-second ECG measures was trained to infer 5-year risk of new-onset AF. The investigators concluded that their method had similar predictive utility as the widely accepted clinical risk factor model: the Cohorts for Heart and Aging Research in Genomic Epidemiology AF score (CHARGE-AF) (14). Yet, the combination of both (clinical risk factors and ECG) provided the greatest predictive accuracy with good discrimination and calibration. These findings underscore the capacity of AI-enabled approaches to improve AF prediction in an inexpensive, non-invasive, widely available, and point-of-care testing manner.

Two studies utilized EHRs to develop AF prediction models (41, 42). EHRs contain real-time, patient-centered data that are instantly available to patients and authorized healthcare providers. The increasing digitalization of healthcare systems makes EHRs more and more widely available, making them particularly useful to predict new-onset AF. A recent study developed an AF prediction model for a 6-month time period using 200 most common EHR features of 2,252,219 individuals. However, this ML approach did not substantially perform better than a logistic regression model using traditional AF risk factors (41). Another study developed a model to predict new-onset AF over a 2-year time period with good discrimination (C-statistic of 0.81) using a 10-variable model compromised of covariates commonly available in the EHRs of 53,552 subjects. This study, however, did not compare the developed prediction model to previous validated AF risk prediction models (42). Another study evaluated the improvement in 5-year AF risk prediction when adding novel variables identified by ML to the CHARGE-AF enriched score (27). Although this study was not conducted using EHRs, it did include clinical, serological, echocardiographic, and cardiac imaging information that are increasingly becoming more available within EHRs. This method, however, did not significantly improve AF prediction in comparison to the CHARGE-AF enriched score (27).

The emergence of wearable devices constitutes another major source to improve AF prediction and management (23). Wearable devices often use photoplethysmography or ballistocardiography to monitor an individual's cardiac rhythm. Other forms of wearable devices include cardiac implantable electronic or patch-based ECG devices that have proven to be useful in selected patient populations to detect AF or assess the risk of stroke in AF patients beyond the CHA2DS2-VASc score (43, 44). Such wearable devices are simple to apply and enable real-time continuous monitoring of the heart. This makes the use of wearable devices promising as one could use this electrical information to predict new-onset AF. Nonetheless, the use of wearable devices and AI algorithms is currently mainly limited to AF detection rather than new-onset AF prediction.

Premise of Big Data in Prediction of New-Onset Atrial Fibrillation

Leveraging big data by AI-enabled approaches could offer great opportunities to improve AF prediction. First, AI methods could be used to overcome statistical issues that are potentially challenging in traditional approaches. In particular, for complex diseases such as AF, simultaneous use of hundreds of quantitative biomarkers may lead to problems such as multi-collinearity, non-linearity, complex interactions, and the possibility of over-fitting (45–47). AI methods, including random (survival) forest method, a non-parametric ML decision tree-based approach, have been proposed to overcome such challenges (45, 47). Second, AI approaches allow for data mining purposes to automatically extract more valuable information from unstructured and complex datasets to improve AF prediction. Ultimately, use of AI would allow combination of various extensive, annotated data libraries into multidimensional datasets which include genotyping, imaging, clinical, and other subphenotypic information. These multidimensional datasets could be used to identify different AF subphenotypes which then could be utilized to improve AF prediction, prevention, and management in a potentially sex and gender-specific manner. Particularly, the inclusion of multilayered high-throughput omics data in such datasets seems promising, given the vast contribution of genetic studies (genome-wide association studies, experimental and in silico candidate gene studies, and Mendelian randomization studies) to advance AF pathophysiology. Recently, a data-driven cluster analysis of 9,749 AF patients, using 60 clinical characteristics, led to identification of 4 cluster AF phenotypes (48). The 4 AF phenotypes were: AF with limited risk factors, younger AF patients with comorbid behavioral disorders, AF patients with tachycardia-bradycardia with device implantation due to sinus node dysfunction, and AF with atherosclerotic vascular disease (48). Another cluster analysis of 2,458 AF patients, using 46 variables, identified 3 cluster AF phenotypes including younger paroxysmal AF, persistent/permanent AF with left atrium enlargement, and atherosclerotic comorbid AF in elderly (49). Another study used hierarchical clustering analyses to identify distinct phenotypes of primary mitral regurgitation which is also considered a heterogeneous disease, as it is the case with AF (50). As such, AI applications could further aid in improving subphenotypic AF classifications to further unravel the complexity and heterogeneity of AF. Based on data-driven approaches, rather than hypothesis-driven approaches with a-priori assumptions, leverage of big data with AI methods can also identify and prioritize AF biomarkers within the realm of AF risk prediction.

Potential of Big Data for Sex- and Gender-Specific Prediction of New-onset Atrial Fibrillation

While big data sources such as ECGs, EHRs, and wearable devices could improve AF prediction and management, their potential for a sex- and gender-specific approach to AF needs further attention.

In particular, recent electrophysiological evidence highlights sex differences with regard to cardiac cellular electrophysiology and their translation to the ECG parameters. It is well documented that sex hormones affect the action potential morphology and cellular electrophysiology through their influence on ion channel function and current densities. Specifically, men have a shorter action potential duration, have a more prominent phase 1 repolarization, and shorter phase 3 repolarization than women (51). It is hypothesized that primarily inward depolarizing L-type Ca2+ current and outward repolarizing K+ currents modified by sex hormones are responsible for these sex differences in action potential morphology. Moreover, women have a higher heart rate, lower heart rate variability, shorter and taller P waves, shorter PR interval, shorter and smaller QRS complexes, longer QT and QTc interval, longer JT interval, wider and smaller T waves, smaller J point, and smaller ST segment in comparison to men (51). From a clinical perspective, the longer QTc interval in women makes women more prone to drug-induced QTc prolongation which may result in torsades de pointes (51). The higher J point and ST segment in men may explain the higher prevalence of J-wave syndromes (Brugada syndrome, and early repolarization syndrome) in men (51). Next to these sex hormones' modulations on a cellular level, it is thought that the smaller size of the women's heart, at least partially, explains some of the documented ECG sex-differences. Noteworthy, these sex differences in sinus rhythm generally persist when men and women develop AF (52).

EHRs hold real-time, patient-centered data from men and women that reflect sexual and biological differences such as obesity and hypertension, among others. However, lack of precise and inclusive documentation of gender, the socio-culturally constructed characteristics of men and women, in EHRs is notable (7, 53). Although, gender documentation may be incomplete in the narrow sense (no actual documentation of i.e. cis-, trans-, or non-binary gender identity), EHRs provide more information in a broad sense on socio-cultural habits, thereby representing gender. More specifically, EHRs could provide information on socio-culturally determined characteristics, roles or habits; including but not limited to socio-economic status, physical and social behaviors that influence e.g. physical activity, social interaction, medication use and adherence to medication, and healthcare use (5–7). Such information could shed more light on gender-related factors that may impact AF. With regards to wearable devices, previous evidence has shown a higher burden of atrial and ventricular arrhythmia in women using a wearable cardioverter-defibrillator, compared to men (54). This is in line with the evidence underpinning the existence of sex-specific ECG features, as mentioned earlier. In addition, similar to EHRs, wearable devices may give more insight into gender-related behavioral habits including physical activity, working hours, sleep, caloric and fluid intake, and other social activities. The latter variables have indeed been suggested to be of added value when modeling sex and gender differences in various health domains, although data to support this claim within the AF field is lacking (55).

The sex-related dimensions in ECGs, and gender-related features in EHRs and wearable devices, as discussed earlier, are recognizable and should yet be fully explored to improve AF prediction and management. AI-enabled algorithms may be able to detect more sex-specific characteristics and gender-related behavioral patterns within these big data sources that might not be apparent on a more macroscopic level or by using traditional statistical methods. This may eventually improve our understanding of sex- and gender-related differences in AF.

Discussion

This review highlighted the current status, premise, and potential of big data to improve prediction of new-onset AF in a sex- and gender-specific manner. While leveraging big data using AI-enabled algorithms offers major opportunities to further advance AF prediction, adoption of a sex- and gender-specific approach is still lacking.

Ample challenges remain before AI-enabled algorithms can be adopted for prediction, prevention, and management of AF. First, the interpretability (transparency and explainability) of AI and exact definition of how all the different methods work is yet difficult (56). This so called “black box” is dependent on the type of AI algorithms that is being used (in particularly deep learning). Application of such algorithms warrants careful and thorough examination of the methods before their implementation, because an algorithm that is intransparent and/or unexplainable may lead to erroneous conclusions that could potentially harm a patient. Second, validation and calibration of AI-enabled algorithms for AF prediction while using external data sources are essential before such algorithms could be widely adopted, implemented, and used within the AF field. Third, the classification codes of clinical variables, drugs, and diseases in different countries and hospitals are different. This leads to challenges with regard to data extraction and harmonization. Careful standardization to harmonize the data, derived from multiple sources, and to integrate all the data modalities within a multidimensional dataset is warranted. Fourth, various ethical issues such as privacy, transparency, informed consent, and trust should be taken care of, and the potential for criminal and malicious use and contested ownership of data should be carefully considered (57). Lastly, rigorosity of AI-enabled algorithms depends on the objectivity, quality, and size of the data used to train them. False, low quality, non-representative study sample, and/or missing data will result in invalid models, while also limiting the generalizability of such models (58). This latter limitation is in particular of concern, as it further magnifies the sex and gender inequalities that already exist within research.

Large, diverse, and multidimensional data sources carry the potential to improve AF prediction and management. Yet, various challenges remain before AI-enabled algorithms can be adopted and implemented. Enhancement of personalized and precision medicine in AF warrants taking into account the complexity of sex and gender dimensions in big data sources and methods, while also overcoming the challenges that currently accompany the use of AI-enabled algorithms.

Author Contributions

Conceptualization, project administration, and validation: SG and MK. Data curation, investigation, methodology, resources, writing—original draft, and writing—review and editing: SG, ZL, and MK. Funding acquisition and supervision: MK. Software and visualization: SG. All authors contributed to the article and approved the submitted version.

Funding

This study was further supported by the Senior Scientist Grant from Dutch Heart Foundation (03-004-2021-T050).

Conflict of Interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher's Note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.