Atrial fibrillation (AF) is a common arrhythmia associated with an elevated risk of stroke, heart failure, and cardiovascular mortality [1, 2, 3]. Unfortunately, AF often goes unnoticed and untreated due to its frequent asymptomatic or minimally symptomatic nature, especially in cases of paroxysmal AF (PAF) [4]. In China, over one-third of AF patients remain unaware of their condition, with a higher prevalence observed in specific demographics such as individuals aged 45–54 years and $\ge$75 years, men, rural residents, those with lower education, and patients without major vascular comorbidities [5]. Thus, strategies for screening and identifying undetected AF are of great importance in stroke prevention.

Screening for AF using just a standard 12-lead electrocardiography (ECG) is challenging. While continuous monitoring using cardiac implantable electronic devices (e.g., pacemaker, implantable defibrillator, implantable loop recorder) has improved AF detection rates [6, 7], these methods have provided few benefits and contributed to increased healthcare costs. Multiple wearable devices such as smartwatches, smartphones, and digital ECG patches have been proposed as tools for AF screening. However, to date, these screening modalities are neither practical nor affordable [8, 9, 10, 11].

While artificial intelligence-enabled electrocardiography (AI-ECG) has recently gained attention as a potent and cost-effective tool for detecting or predicting PAF during sinus rhythm (SR), no prior study has explored the potential of conventional printed ECG (pECG) as a screening tool for PAF. The use of AI-pECG has the potential to bridge a clinical gap, especially in developing countries like China. In this study, our objectives were two-fold: (1) to assess the suitability and accuracy of a deep-learning model designed to detect AF using printed ECG records, and (2) to develop a deep learning-based WeChat applet for AF screening that can provide immediate results in real-world clinical settings.

In this study, we presented a novel deep-learning algorithm designed for the identification of PAF based on printed 12-lead ECG records obtained during SR. Our innovative deep-learning model, coupled with a companion WeChat applet, provides a practical solution for clinicians, particularly those in developing countries, to conduct real-time AF screening with an impressive AUROC of 0.871 and high sensitivity (87.5%). The study also highlights features of ECG that are critical to identifying PAF. While these features may be challenging to visualize, they are identifiable through the AI algorithm. To the best of our knowledge, this study is the first attempt to integrate pECG records and AI into a model for AF detection. The mobile user interface has proven to be intuitive, user-friendly, and compatible with a range of widely-used devices.

Early diagnosis of AF, particularly PAF, is clinically beneficial but presents significant challenges. Approximately one-third of AF patients are asymptomatic, and the sporadic nature of symptoms hinders precise ECG monitoring during AF episodes [14]. Previous studies focused on P-wave indices, such as duration, dispersion, area, abnormal axis, and P-wave terminal force in lead V1 during SR, as predictors for the development of AF [15, 16, 17, 18]. Advanced interatrial block has also been identified as a significant predictor of new-onset AF, AF recurrence, and ischemic stroke [19]. The independent predictive value of QRS duration for new-onset AF in women was highlighted by Aeschbacher et al. [20]. The QRS duration reflects the conduction speed through the specialized cardiac conduction system and ventricular myocardium [21]. It was demonstrated that this variable is intricately linked to cardiac structural and functional abnormalities [21]. However, existing methods have shown moderate performance in providing clinical utility. Standard SR-ECG readings do not always represent normal atrial contraction uniformly [22]. A substantial proportion of patients with post-cardioversion AF continue to exhibit persistent disorganized appendage contraction [22]. This suggests that there may be multiple wavelets, undetectable to the human eye, that precede the typical ECG waveform changes in individuals predisposed to AF. These factors collectively underscore the complexity of early PAF detection and the need for more refined diagnostic methods.

Recent studies have emphasized the potential of AI algorithms to address these challenges in AF detection and prediction. Attia et al. [23, 24] developed an automatic AF detection method using a CNN model based on short-term normal ECG signals, indicating that a shorter period (30 days) between AF and SR enhances diagnostic accuracy. In a follow-up study, they analyzed surface ECGs from a substantial cohort comprising 59,212 patients, both before and after the first documented AF episode. This study revealed the AI-ECG probability of AF gradually increases prior to the first AF episode, experiences a temporary decrease 1–2 years post-AF, and then resumes a continuous increase thereafter. Baek et al. [25] reported that an AI model showed that the optimal interval for detecting subtle changes in PAF occurred within 0.24 seconds before the QRS complex in the 12‑lead ECG during SR. This insight contributes to the evolving landscape of AI-based predictive tools, which estimate the probability of developing AF from surface ECGs recorded during sinus rhythm, showing significant promise in clinical practice.

All previous AI-ECG studies utilized digital ECG data, which is not universally available in developing countries [26, 27]. The majority of hospitals in developing countries do not yet benefit from the digital recording and storage of ECG data, and do not have the ability to automatically reanalyze paper-based ECG records. The collection of extensive, high-quality ECG data remains a challenge in these settings. However, deciphering the patterns that pECG records may reveal is an essential step toward developing a more accurate method for detecting PAF in low-resource clinical scenarios.

In our study, the AI model enabled the identification of subtle yet crucial differences between the pECGs obtained during SR from patients with PAF and those obtained from participants without a history of AF, even with relatively limited data. Although our model achieved a slightly lower ROC and accuracy compared with the models used in other recently published studies, our findings remain promising for clinicians treating patients in low-resource environments.

The preliminary screening tool could be used in the following manner: if the result indicates the presence of PAF, healthcare professionals should recommend more intensive monitoring for the patient. Conversely, for patients identified as non-AF, the routine AF education and observation would be recommended. This strategy could expedite the time to PAF diagnosis, thereby contributing to the prevention of ischemic stroke. Additionally, it may mitigate unnecessary medical interventions and decreasing the risk of bleeding associated with false-positive results during initial assessments. While it is premature to make any changes in the current AF recommendations, our findings represent a significant step forward in the field of AF screening.

The perception of AI algorithms as “black boxes” due to their opaque decision-making processes is a notable concern. In medical settings understanding the rationale behind diagnostic or therapeutic recommendations is essential for building trust and ensuring patient compliance [28]. This lack of transparency can be a significant barrier to the integration of AI systems in healthcare. To address this issue, researchers have investigated attention mechanisms such as Grad-CAM [29, 30, 31]. Kim et al. [32] observed a consistent focus in their model’s attention during predictions across window periods of 1, 2, and 4 weeks. The model’s identified focal points, specifically around one QRS point and one point between the T and P waves, provide valuable insights for exploring the mechanisms of AF through AI-ECG analysis.

In this study, we observed that in specific images, the highlighted red regions (indicative of the model’s focus) may concentrate on specific waveforms or electrocardiographic areas. Interestingly, these focal points might shift or appear in different positions in other images. This variability could indicate the model’s sensitivity to individual differences in diverse samples or stem from subtle uncontrollable variations in light and color among numerous print-scanned data. We will undertake further in-depth analyses of printed data to enhance result accuracy and clarity.

A few important limitations of the present study warrant discussion. First, our dataset, obtained from a single tertiary care teaching hospital in China, is relatively smaller in comparison to other large-scale AI-ECG studies. In order to ensure a robust sample size for model training in both groups, we designed the cohort with a higher prevalence of AF. Ongoing efforts are directed towards a prospective multicenter study aimed at evaluating the model’s performance across a more diverse, ostensibly healthy population. Second is the potential for undetected AF in patients categorized as non-AF, as our inclusion criteria were restricted to those who underwent electrophysiological procedures. It is reasonable to hypothesize that some false-positive patients in the non-AF group may be undiagnosed AF cases. Third, the image quality check process is current undergoing refinement, which limits the effectiveness of our WeChat applet. Finally, all pECGs recorded in our study were in a 12-by-1 ECG format, raising uncertainty about the generalizability of our results to other distributions. Our future studies aim to further explore the intricacies of pECGs to advance their clinical utility in precision medicine.

Utilizing deep-learning for the analysis of 12-lead printed ECG records has demonstrated its capability to accurately detect PAF during SR. This technology holds promise as a reliable screening tool in real-world clinical settings, particularly in regions with limited access to digital ECG analysis. The ongoing exploration of printed ECG patterns aims to unlock additional insights for precision medicine and clinical practice.

AF, atrial fibrillation; ECG, electrocardiography; SR, sinus rhythm; pECG, printed electrocardiography; PAF, paroxysmal atrial fibrillation; AUROC, area under the receiver operating curve; PRAUC, area under the precision-recall curve; AI-ECG, artificial intelligence-enabled electrocardiography; Grad-CAM, gradient-weighted class activation map; CNN, convolutional neural network.

The data underlying this article will be available upon reasonable request to the corresponding author.

YZ, DYZ, XPL and SDH made substantial contributions to conception and design. YZ, YC, YT, LS, and YJW were contributed to the acquisition of data. YZ and DYZ prepared the first draft of the article. YZ, DYZ, SJG and GDW performed the data analysis. YC, SJG, GDW, YT, LS, YJW and SDH were involved in critical revision of the content. YZ and XPL made the final revision. All authors read and approved the final manuscript. Each author has actively contributed to the work and is committed to being accountable for all aspects of the research.

The studies involving human participants were reviewed and approved by the Ethics Committee of Beijing Chaoyang Hospital. Patient’s informed consent is waived. The corresponding documentation bears the reference number 2022-QI-14-1.

Not applicable.

This research was funded by the National Natural Science Foundation of China, grant number 82270325 and 62102008.

DYZ, SJG and GDW are employees of HeartVoice Medical Technology and the company’s apparatus was used in the experiment, but there is no conflict of interest. Other authors also declare no conflict of interest.

Supplementary material associated with this article can be found, in the online version, at https://doi.org/10.31083/j.rcm2507242.