Robust Artificial Intelligence Tool for Atrial Fibrillation Diagnosis: Novel Development Approach Incorporating Both Atrial Electrograms and Surface ECG and Evaluation by Head‐to‐Head Comparison With Hospital‐Based Physician ECG Readers

Abstract

Background

Atrial fibrillation (AF) increases risk of embolic stroke, and in postoperative patients, increases cost of care. Consequently, ECG screening for AF in high‐risk patients is important but labor‐intensive. Artificial intelligence (AI) may reduce AF detection workload, but AI development presents challenges.

Methods and Results

We used a novel approach to AI development for AF detection using both surface ECG recordings and atrial epicardial electrograms obtained in postoperative cardiac patients. Atrial electrograms were used only to facilitate establishing true AF for AI development; this permitted the establishment of an AI‐based tool for subsequent AF detection using ECG records alone. A total of 5 million 30‐second epochs from 329 patients were annotated as AF or non‐AF by expert ECG readers for AI training and validation, while 5 million 30‐second epochs from 330 different patients were used for AI testing. AI performance was assessed at the epoch level as well as AF burden at the patient level. AI achieved an area under the receiver operating characteristic curve of 0.932 on validation and 0.953 on testing. At the epoch level, testing results showed means of AF detection sensitivity, specificity, negative predictive value, positive predictive value, and F1 (harmonic mean of positive predictive value and sensitivity) as 0.970, 0.814, 0.976, 0.776, and 0.862, respectively, while the intraclass correlation coefficient for AF burden detection was 0.952. At the patient level, AF burden sensitivity and positive predictivity were 96.2% and 94.5%, respectively.

Conclusions

Use of both atrial electrograms and surface ECG permitted development of a robust AI‐based approach to postoperative AF recognition and AF burden assessment. This novel tool may enhance detection and management of AF, particularly in patients following operative cardiac surgery.

Nonstandard Abbreviations and Acronyms

AFL

atrial flutter

F1

harmonic mean of positive predictive value and sensitivity

Clinical Perspective.

What Is New?

• The use of epicardial atrial electrograms and surface ECGs in combination offer a novel means for determining a true atrial fibrillation (AF) diagnosis in postoperative cardiac patients, thereby permitting creation of a robust artificial intelligence tool for detection of AF during surface ECG monitoring following cardiac surgery.

• AF detection in postoperative patients is important to prevent AF‐related complications but is currently labor intensive; artificial intelligence offers the potential to detect new postoperative AF more efficiently and at lower cost.

What Are the Clinical Implications?

• The next step is to expand application of the artificial intelligence algorithm to multiple cardiac surgery centers to determine its potential for both recognizing and treating postoperative AF earlier, thereby reducing length of hospital stay and cost of care.

Atrial fibrillation (AF) affects ≈2% to 4% adults globally and is encountered in 30% to 60% postoperative patients. ¹ Further, risks associated with unrecognized and untreated AF are substantial. In particular, AF is associated with approximately one‐third of embolic strokes ² , ³ and a 2‐fold increase in mortality associated with stroke. ⁴ In addition, AF following surgery complicates postoperative care and often prolongs hospitalization.

Studies have shown that screening for AF can be cost‐saving by identifying those individuals with prolonged AF burden warranting anticoagulation and permitting early initiation of AF treatment (whether by rhythm or rate control), thereby reducing hospital length of stay for postoperative patients. ¹ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ However, due to the paroxysmal and often asymptomatic nature of AF, diagnostic screening using current techniques is often inadequate. ⁵ In particular, effective AF detection necessitates labor‐intensive ECG monitoring; the latter is a challenging task given the substantial magnitude of ECG data that must be reviewed. Inevitably, clinicians will need reliable computer‐aided ECG diagnostic analysis to not only identify AF but also quantify durations of continuous AF occurrences and overall AF burden for patients. Artificial intelligence (AI) systems may prove helpful in this regard, but their development demands a well‐curated and structured AF database for AI‐based tool training and testing. ¹⁰

This prospective study was undertaken to provide a robust AF‐annotated ECG database platform for subsequent AI development by using postoperative atrial electrogram recordings and ECG tracings to assist true AF annotation in patients with cardiac surgery. Thereafter, performance assessment of the AI platform was undertaken when applied to ECG alone in a separate test group of patients following operative cardiac surgery.

METHODS

Study Design and Oversight

This investigator‐initiated, prospective, single‐center study was designed to enroll consecutive patients. Each qualified patient received ≈7‐day continuous ECG monitoring and data collection of both single‐lead surface ECG and single‐channel atrial electrogram during their postoperative hospitalization. The atrial electrogram was used to assist expert ECG physicians in establishing a more precise diagnosis of true AF than would be obtained by using conventional surface ECG alone. Each patient's ECG record was to be divided into 30‐second nonoverlapping segments (epochs) with annotations of AF or non‐AF as input data structured for a deep learning AI development as well as subsequent test purposes. The data that support the findings in this study are available from the corresponding author upon reasonable request.

Patient Population

The database was constructed from consecutive adult patients consenting to undergo indicated but otherwise routine cardiovascular surgery between July 2, 2020 and February 25, 2021 (Figure 1). Inclusion criteria were: (1) adults aged 18 to 80 years; (2) nonurgent or nonemergent off‐pump coronary artery bypass graft surgery and valvular surgery; (3) left ventricular ejection fraction > 40%; (4) patients undergoing heart surgery for the first time; (5) no history of adult congenital heart disease; (6) good compliance and ability to complete follow‐up; and (7) atrial epicardial leads implanted during hospitalization. Exclusion criteria were: (1) history of prior cardiac surgery; (2) heart surgery accompanied by surgery of another organ system; (3) liver and kidney function abnormality (results exceeding the upper limit of normal value by 3 times) or with coronary artery bypass graft surgery or valvular surgery contraindications; (4) patients with other diseases who required radiotherapy, chemotherapy, or long‐term hormone treatment; (5) patients who declined to consent to the surgical procedure; (6) patients who were unable to have epicardial leads implanted during hospitalization; (7) patients in need of other critical care such that the data collection became inconvenient or not applicable; for instance, those requiring pressor agents or antiarrhythmic drugs; and (8) left ventricular ejection fraction ≤40.

Figure 1. Flow diagram of study design.

A total of 1032 consecutive patients scheduled for surgery were screened for study eligibility. The ECG/electrogram recordings in a total of 659 patients were divided chronologically into 3 data sets: learning/training (263), learning/validation (66), and testing (330), with a ratio of 40%:10%:50%. EF indicates ejection fraction.

Sample Size

AI development in general requires as large and accurately annotated or labeled structured data as possible. ¹ , ¹⁰ , ¹¹ , ¹² For the structuring purpose, we divided continuous surface ECG into 30‐second epochs based on European Society of Cardiology (ESC) guidance, which recommends single‐lead surface ECG strip with no less than 30 seconds’ duration for AF diagnosis. ¹ The data acquisition plan was to collect data continuously on each patient for 7 days; our goal was to record a number of patients comparable to 328 patients by Hannun et al. ¹³ Since we record 7‐day continuous data (604 800 seconds) on each patient, our database would also be comparable to the 1 million 10‐second ECG strips used by Attia et al. ¹² We distributed training, validation, and testing data sets at a ratio of 40%:10%:50% (Figure 1).

The study was conducted in accordance with the principles of the Declaration of Helsinki. All records were deidentified and individual patient consent was obtained for medical/surgical interventions. The protocol was approved by the ethics committee of the General Hospital of Northern Theater Command and registered in the Chinese Clinical Trial Registry (http://www.chictr.org.cn/searchproj.aspx registration number: ChiCTR2000029310). It was overseen by the ethics committee and the Science Discipline Department. Both committees had the authority to stop the trial for safety or efficacy concerns.

Data Acquisition and Preparation

The data acquisition system was capable of recording one surface ECG channel and another atrial signal channel derived from a pair of Teflon‐coated stainless cardiac pacing leads that were secured routinely during cardiac surgery to the atrial epicardium for therapeutic pacing purpose. The latter electrodes provided an opportunity to record unipolar atrial signals synchronized with simultaneous single‐lead surface ECG (Figures 2 and 3).

Figure 2. Examples of unipolar atrial signal synchronized with simultaneous single‐lead surface ECGs.

Examples of continuous recording of surface ECG (top line in each section) and simultaneous atrial epicardial electrograms (second line in each section). A, The electrogram reveals both atrial fibrillation and corresponding ventricular activity (synchronous with QRSs from surface ECG). Atrial signals are more difficult to appreciate on the ECG alone, and given periods of relatively regular QRSs, the atrial fibrillation diagnosis may be missed when ECG readers are confronted with a high volume of recordings. The electrogram offered rhythm confirmation, thereby enhancing artificial intelligence (AI) learning capability. B, Both surface ECG (top line in each section) and atrial electrograms (second line in each section) are shown from a patient in whom the irregularity of the ECG suggests atrial fibrillation, whereas the discrete atrial electrograms at cycle length>200 ms favor the diagnosis of atrial flutter. In this case the electrogram facilitated the correct rhythm diagnosis for AI learning.

Figure 3. Annotation schematic.

Surface ECG data acquisition was based on single‐lead ECG (→) and concurrent atrial electrogram signal acquisition derived from a pair of Teflon‐coated, stainless cardiac pacing lead (→), a, b were labeled as AF and c, d were labeled as non‐AF with the assistance of atrial signal; and then the atrial signal was eliminated and the annotated surface ECG data (a’, b’, c’, d’) platform was established and ready for AI training, validation, and testing. The designation “ADD & Norm” are known steps adopted by the transformer architecture that contribute to the stable training of the model as noted in Vaswani et al ²⁰ and in He et al. ²¹ AAO indicates ascending aorta; CPL, cardiac pacing lead; RA, right atrium; and RV, right ventricle.

Annotation Procedure and Data Structure

A committee of 10 cardiologists and physicians experienced in Holter ECG diagnosis (Data S1) from various hospitals in China provided independent annotations of AF or non‐AF for all collected data. Unlike conventional postoperative monitoring methods that rely on surface ECG alone, the committee members were permitted to use both ECG and the additional atrial signals. Each record was randomly assigned to 2 different committee members to distinguish AF from all other non‐AF cardiac rhythms, including sinus rhythm, premature ventricular contractions, premature atrial contractions, abnormal conductions, and atrial tachycardia or atrial flutter (AFL).

The 2020 ESC guideline recommends the following criteria for AF diagnosis: (1) the surface ECG shows absolutely irregular RR intervals; (2) there are no distinct P waves on the surface ECG; and (3) the atrial cycle length is <200 ms (>300 beats per minute). ¹ , ⁶ , ⁷ , ⁸ , ⁹

From a practical perspective, conventional methods can only adopt criteria (1) and (2) for AF diagnosis since only surface ECG is available. Terms of absolutely irregular or distinct P waves relied heavily on experienced physicians to interpret; therefore, introducing interobserver or even intraobserver variability in AF diagnosis at epochs level. Challenges for annotating AF based on surface ECG alone may arise in the presence of AFL or atrial tachycardia or in conjunction with abnormal atrial ventricular conduction (eg, intermittent atrioventricular block), especially when no atrial signal is available or single‐channel surface ECG recording quality is inadequate for reviewers to identify P waves.

The atrial electrogram signal in our study provides opportunity to diminish the above annotation challenges. For example, absolutely irregular in criterion (1) can be determined by directly measuring intervals of atrial activity, whereas no distinct P waves in criterion (2) can be confirmed by absence of atrial signal or presence of multiple atrial signals between QRS complexes, while also deternmining atrial cycle length or rate in criterion (3). The latter can be assessed directly by counting numbers of atrial spikes within an observational or calculation window (eg, 30‐second epoch or 10 QRS waves). Therefore, our method was able to yield trusted or true AF diagnostic labeling at epochs level for surface ECG.

The goal of our committee was to create AF or non‐AF annotations for each epoch of surface ECG throughout the record. A conservative approach based on consensus was set for diagnosing AF, especially when criteria (1) and (2) were hard to discern despite the atrial rate averaging around 200 ms. In such cases, the outcome would be decided in favor of AF for the ECG epoch instead of AFL or atrial tachycardia. If a decision could not be made, the epoch would be discussed within the committee to establish a consensus (see Data S1 for further amplification of the decision‐making process).

In cases where QRS waves from surface ECG were indistinguishable or pacing was activated for clinical reasons, the start and end of such periods were annotated to identify ECG epochs falling within that time segment. Such epochs did not participate in AI development and testing.

Classifier Design and Supervised AI Training

We designed an AF classifier based on commonly used transformer‐based deep learning models. The overall model architecture is shown in Figures 1 and 3. A 30‐second surface ECG segment has a total of 30s×256=7680 digits (ECG recorder sampling rate was 256 Hz) as the input for the transformer‐based classifier. A 1‐dimensional convolution layer with kernel size 64 and stride size 32 of the classifier takes on the input and outputs a matrix of 240 (a calculation window of 64 digits with 32 overlap throughout the 7680 digits) by 512 (local features dimension). Then the outputs were fed into the 6‐layer transformer to generate a final matrix of 240×512, upon which a probability of AF (for final classification result) is calculated for the input surface ECG. Parameters inside the classifier were all trainable. The residual connection and layer normalization were set between modules within the classifier to stabilize the training process.

A common and simple machine‐learning procedure was adopted to train the AI classifier: stochastic gradient descent–based optimization for the classifier's trainable parameters on a training data set, while machine‐learning progress was monitored on a validation data set at the end of each training iteration. One training iteration was considered complete when an ergodic condition was met: all surface ECG segments with AF annotation in the training data set had to be chosen at least once (ie, ergodic) during a random selection to form a batch of 128 segments (64 AF and 64 non‐AF). Then the AF classifier was applied on the validation data to obtain the F1 (the harmonic means of positive predictive value and sensitivity) score of this iteration. Training iterations were repeated until the F1 score stopped increasing; then the model from the iteration with the highest F1 score was chosen. The learning curves are shown in Figure S1, illustrating the F1 score and loss on the validation set for each iteration. Thus, the training process was completed for the transformer‐based classifier. Finally, the AF cutoff threshold for the classification probability was chosen based on receiver operating characteristic analysis on the last round of validation results (Figure 4).

Figure 4. Atrial fibrillation (AF) burden for each patient record among learning and testing data sets.

Patient data are displayed as sorted incrementally according to AF burden. A, Histograms of the learning and testing sets (x axis: the frequency of total AF burden; y axis: total AF burden for each patient). B and C, QQ plot of AF burden in the learning and testing data sets (x axis: observed value of total AF burden for each patient; y axis: expected normal value of total AF burden conforming to the normal distribution). The distribution of AF burden was the same across categories of patient sets, statistically tested by independent‐samples Kolmogorov‐Smirnov test (P=0.751).

Training computation was conducted on a machine with one NVIDIA Tesla V100 GPU using the PyTorch framework. During the training process, we employed cross‐entropy as the loss function and optimized the model with Adam (β=0.9, 0.98). The learning rate was increased linearly from 0 to 1e−4 in the first 10, 000 updates and then decreased linearly to 0 in the following 90,000 updates. Detailed information is included in Data S1.

Statistical Analysis

The trained AI was evaluated by assessing its diagnostic accuracy, the area under the curve of the receiver operating characteristic, sensitivity, specificity, negative predictive value, positive predictive value, and F1. Table S1 defines precision, sensitivity, specificity, F1, true‐positive, true‐negative, false‐positive, and false‐negative rates.

Intraclass correlation coefficient of AF burden at the patient level was calculated by reliability analysis with 2‐way random‐effects model and the type of absolute agreement, using SPSS Statistics for Windows version 21.0 (IBM). AF burden for each patient record among learning and testing data sets were compared by independent‐samples Kolmogorov–Smirnov test.

Epoch‐level AF detection by the trained AI enabled calculation of AF episodes and burden for each patient. Patient‐level performance was assessed against the committee AF diagnosis (AF episodes lasting for <30 seconds were excluded for this performance calculation). Sensitivity and positive predictivity of AF episodes as well as burden at the patient level were assessed by running epicmp and sumstats from www.physionet.org.

RESULTS

A total of 10 981 803 annotated surface ECG epochs were derived from recordings (learning: 6.0±1.0 days; testing: 5.6±1.3 days) obtained in 659 adult patients (age range, 27–80 years) to form the database for AI development purpose. The average patient age was 61 years. Age, sex, left ventricular ejection fraction, left atrial diameter, and CHA₂DS₂‐VASc or HAS‐BLED scores did not differ significantly between the learning and testing subgroups. Both subgroups had similar frequency of postoperative AF (Table 1) as well as overall AF burden for each patient (Figure 4).

Table 1.

Patient Demographics in the Learning and Testing Data Sets

	Total	Learning	Testing	P value
	(N=659)	(N1=329)	(N2=330)
Age, y
Mean±SD	61.1±9.3	61.5±9.0	60.7±9.5	0.31
Median [minimum, maximum]	63.0 [27.0, 80.0]	63.0 [27.0, 80.0]	62.0 [27.0, 80.0]
EF, %
Mean±SD	55.9±6.1	55.9±6.3	55.9±6.0	0.87
Median [minimum, maximum]	58.0 [30.0, 67.0]	58.0 [30.0, 67.0]	57.0 [30.0, 66.0]
FS, %
Mean±SD	29.2±3.8	29.2±3.9	29.2±3.8	0.89
Median [minimum, maximum]	30.0 [15.0, 38.0]	30.0 [15.0, 38.0]	30.0 [15.0, 36.0]
LA, mm
Mean±SD	39.4±7.5	39.8±7.9	39.0±7.2	0.13
Median [minimum, maximum]	38.0 [26.0, 87.0]	38.0 [29.0, 81.0]	38.0 [26.0, 87.0]
Sex, n (%)
Male	446 (67.7)	218 (66.3)	228 (69.1)	0.44
Female	213 (32.3)	111 (33.7)	102 (30.9)
Heart rhythm, n (%)				0.07
Sinus rhythm	339 (51.4)	158 (48.0)	181 (54.8)
AF	68 (10.3)	42 (12.8)	26 (7.9)
POAF	252 (38.2)	129 (39.2)	123 (37.3)
Surgery, n (%)				0.27
Coronary artery bypass	420 (63.7)	207 (62.9)	213 (64.5)
Valve replacement	199 (30.2)	97 (29.5)	102 (30.9)
Combined	22 (3.3)	12 (3.6)	10 (3.0)
Other	18 (2.7)	13 (4.0)	5 (1.5)
CHA2DS2‐VASc score, n (%)				0.76
<2	224 (34.0)	110 (33.4)	114 (34.5)
≥2	435 (66.0)	219 (66.6)	216 (65.5)
HAS‐BLED score, n (%)				0.50
<3	563 (85.4)	278 (84.5)	285 (86.4)
≥3	96 (14.6)	51 (15.5)	45 (13.6)

Data are reported as mean±SD and median (minimum, maximum) for quantitative data, and as number (%) for qualitative data. Differences in the characteristics between data sets were evaluated by using t test or χ² test. AF indicates atrial fibrillation; EF, ejection fraction; FS, fractional shortening; LA, left atrium; and POAF, postoperative atrial fibrillation.

A total of 5 675 617 annotated ECG epochs (511 652 AF and 4 013 713 non‐AF in training, and 222 729 AF and 927 523 non‐AF in validation) from 329 patients were used for learning, and a total of 5 306 186 annotated ECG epochs (635 200 AF and 4 670 986 non‐AF) from 330 patients were used for testing, allowing the compilation of a balanced data set. See Tables S2 and S3 for detailed information and performance on epochs level.

In the validation data set, the area under the receiver operating characteristic curve was 0.932 (Figure 5), and the mean of the sensitivity, specificity, negative predictive value, positive predictive value, and F1 for AF patient recognition were 1.000, 0.853, 1.000, 0.868, and 0.930, respectively (Table 2). Detected AF burden for each patient correlated with true AF burden with an intraclass correlation coefficient of 0.983.

Figure 5. Diagnostic performance of atrial fibrillation (AF) and AF burden in the validation set and testing set.

A, The area under the receiver operating characteristic (ROC) curve in the validation set for AF recognition was 0.932. B, The area under the ROC curve in the testing set for AF recognition was 0.953. C, The intraclass correlation coefficient (ICC) for AF burden in the validation set was 0.983. D, The ICC of AF burden in the testing set was 0.952 based on artificial intelligence (AI) and the gold standard. AUC indicates area under the curve.

Table 2.

Performance Metrics

	Validation set (95% CI)	Test set (95% CI)
AUROC curve	0.932 (0.878–0.982)	0.953 (0.933–0.969)
PPV	0.868 (0.625–1.000)	0.776 (0.658–0.881)
Sensitivity	1.000 (0.857–1.000)	0.970 (0.903–1.000)
NPV	1.000 (0.800–1.000)	0.976 (0.936–1.000)
Specificity	0.853 (0.556–1.000)	0.814 (0.723–0.902)
F1 score	0.930 (0.750–1.000)	0.862 (0.758–0.925)
Accuracy	0.925 (0.750–1.000)	0.876 (0.805–0.927)
ICC	0.983 (0.967–0.990)	0.952 (0.939–0.961)

The 95% CI was evaluated by self‐boosting, with a sampling frequency of 2000 iterations. AUROC indicates area under the receiver operating characteristic; F1, harmonic mean of positive predictive value and sensitivity (an index to provide test accuracy); ICC, intraclass correlation coefficient; NPV, negative predictive value; and PPV, positive predictive value.

In the testing data set, the area under the receiver operating characteristic curve was 0.953 and the mean of the sensitivity, specificity, negative predictive value, positive predictive value, and F1 for AF patient recognition were 0.970, 0.814, 0.976, 0.776, and 0.862, respectively (Table 2), and the intraclass correlation coefficient of AF burden was 0.952 (Figure 5).

In the testing data set, sensitivity and positive predictivity (reported as median and interquartile range [IQR]) of AF episode detection at the patient level were 100% (IQR, 99%–100%) and 100% (IQR, 100%–100%), respectively. The IQRs for episode duration concordance were 99% to 100% and 96% to 100% for sensitivity and predictivity, respectively.

We also recorded the time consumed by AI diagnosis. AI could diagnose 621.7±12.3 segments (30‐second epochs) per second. For a 24‐hour ECG diagnosis, AI required 4.63±0.09 seconds.

DISCUSSION

This study offers several principal findings. First, a novel methodology utilized epicardial atrial electrograms to assist experienced ECG readers in diagnosing true AF on a single‐lead surface ECG. Second, the latter approach provided a method for developing a robust data collection and structuring system for development of an AI‐based AF detection tool applicable to surface ECG alone. This method was particularly designed for application in postoperative cardiac surgery patients, but may prove useful to quantify AF burden in other settings. Finally, given the multiple risks associated with unrecognized or untreated AF, AI development using techniques described herein is important since diagnosis of AF and the ability to quantify AF burden are crucial factors in ensuring timely and cost‐efficient assessment and treatment of at‐risk populations. ¹⁰ , ¹³ , ¹⁴

Importance of Establishing True AF for AI Development

The surface ECG provides essential information about the presence, and nature of cardiac arrhythmias and AI may prove helpful in screening for such rhythm disturbances. ¹³ , ¹⁴ However, without accurate arrhythmia annotation, AI may not be able to utilize embedded information or characteristics (some may be beyond human ECG readers recognition) in the surface ECG to establish a correct diagnosis. Thus, surface ECG with robust and correct annotations of arrhythmic types is crucial for AI development. We hypothesized that adding atrial electrogram signals as an aid in establishing robust ECG diagnostic annotations may be beneficial for AF detection by using surface ECG in AI applications.

Previous AI arrhythmia studies have relied on surface ECG alone to create a database for a supervised AI learning. Hannun et al, ¹³ Attia et al, ¹² and Zhu et al ¹⁵ demonstrated promising AI performance in surface ECG diagnosis of cardiac arrhythmias. In these studies, for purposes of AI development, a committee of physicians overread surface ECGs to provide arrhythmia annotations to construct the database suitable for AI. However, AF and AFL were not separated and AI ability to assess AF burden quantification was not determined.

The average agreement rate among physicians on AF diagnosis in prior studies was 72%. ¹⁶ This implies that there are challenges of cardiac arrhythmia diagnosis without atrial activity information. In our study, by adding the atrial signal, the interobserver disagreement rate (cases that had to be resolved by the committee) was relatively low (6%); the cases that caused disagreement arose primarily when the atrial cycle length was ≈200 ms (the threshold for AF or AFL classification). Therefore, as a result of adding the atrial signal during AI development, the subsequent ECG challenge of AF diagnosis is reduced to measurement errors instead of human experts' opinions.

AF Burden

AF burden (defined as the duration of time the patient is in AF rhythm divided by the total recording time), as well as its ventricular rate, longest episode, and frequency of episode occurrence, may provide important insights into stroke risk assessment, as opposed to a binary diagnosis of patients having or not having AF. In one report, ¹⁷ patients with AF without any AF episodes lasting ≥6 minutes and with a total AF burden <6 hours/day for 30 consecutive days had anticoagulants discontinued, otherwise anticoagulation was started or continued. The authors found that such an approach is feasible and could decrease anticoagulation by 75% with few adverse events. ¹⁷ Therefore, accurately and effectively assessing AF burden can be important, and we believe our study is the first to establish a robust methodology for research and validation on stroke risk assessment by AF burden monitoring. In the case of postoperative care, AI systems may permit earlier recognition of AF and thereby allow for earlier interventions as well as facilitating follow‐up.

Finally, our novel AI development methodology permitted distinguishing AF from AFL, a feature not present in prior studies. The value of this additional capability is 2‐fold. First, while clinically both AF and AFL currently warrant anticoagulation, ¹⁸ , ¹⁹ the evidence for anticoagulation is stronger for AF. Second, ablation is considered more effective for AFL.

Limitations

Our study has several limitations. First, the window size for atrial rate calculation may yield different atrial rates especially when atrial activities are varying around the threshold of 200 ms (300 beats per minute). However, since we can now link AF diagnosis directly to AF burden, such a limitation becomes a measurement error instead of diagnostic opinion. Second, the study did not examine the proposed AI method on an external data set. However, the test cohort, as was the case with the training and validation set, was also derived from patients after surgery. Third, patients undergoing open heart cardiac surgery were inevitably relatively sick in terms of cardiac function or cardiac arrhythmia types. Therefore, the database in our study lacks individuals without evident structural heart disease or nonsurgical cardiac patients. On the other hand, it seems reasonable to begin with a postsurgery patient population for both ethical and practical reasons. In regard to the former, atrial leads placement is standard practice in cardiac surgery patients and consequently is not an additional procedure. In terms of the latter, AF frequency is high in postcardiac surgery (30%–60% AF occurrence) and the impact of postsurgery AF on hospital costs and length of stay is substantial.

CONCLUSIONS

Atrial electrograms recorded with surface ECG from postcardiac surgery patients enhance the development of robust deep learning algorithms for automated ECG detection of postoperative AF and its burden. Such a method has the potential to be of diagnostic value and reduce the cost of monitoring in postoperative cardiac patients. As more data are collected, AI‐based tools such as the one studied herein may diminish the duration of hospital stays with consequent reduced cost of care.

AF affects ≈2% to 4% adults and increases the risk of embolic stroke. Further, AF commonly occurs following surgery (particularly cardiac surgery) and both complicates postoperative care and often prolongs hospitalization. Screening for AF in high‐risk populations may both improve health care and reduce costs of care. However, detecting AF is currently labor‐intensive, requiring review of prolonged ECG monitoring. AI may facilitate AF detection but currently presents unresolved challenges in terms of diagnostic accuracy. In this study, a novel method employing atrial electrograms in conjunction with ECG recordings was used to develop a robust AI tool for AF recognition and AF burden determination. Such a method, while designed for postoperative cardiac patients, has the potential to be of diagnostic utility in large populations of community‐dwelling at‐risk individuals such as postsurgery patients, individuals with structural heart disease, and the elderly.

Sources of Funding

This work was supported by the Liao Ning Revitalization Talents Program (XLYC2001001) and the Provincial Key R & D Program (2020JH2/10300156).

Disclosures

None.

Supporting information

Data S1

Tables S1–S3

Figure S1