Risk-Guided Atrial Fibrillation Screening With Artificial Intelligence–Enabled Electrocardiogram Models A VITAL-AF Trial Analysis

Abstract

BACKGROUND

Screening for atrial fibrillation (AF) may lead to earlier detection and initiation of preventive measures. Current AF screening approaches using a guideline age-based threshold of ≥65 years have shown limited yield.

OBJECTIVES

In an AF screening trial, we assessed whether the screening effect was larger among individuals at elevated AF risk using validated clinical and electrocardiogram (ECG)-based artificial intelligence (AI) risk models.

METHODS

VITAL-AF was a cluster-randomized trial of patients aged ≥65 years treated at 1 of 16 primary care practices affiliated with Massachusetts General Hospital. Patients randomized to a screening practice were screened using a single-lead ECG. Among VITAL-AF participants without prevalent AF with at least one 12-lead ECG within 3 years before enrollment, we estimated AF risk using 3 validated models derived outside of VITAL-AF: the Cohorts of Heart and Aging Research in Genomic Epidemiology-AF (CHARGE-AF) clinical score, an AI-based model using a 12-lead ECG alone (ECG-AI), and a model combining ECG-AI and CHARGE-AF (CH-AI). Two-year incident AF discrimination was assessed by the time-dependent area under the receiver-operating characteristic curve (AUROC) and average precision. AF screening effect was defined as the difference in 2-year incident AF diagnosis rate (per 100 person-years) in screening vs control across AF risk deciles.

RESULTS

Of 30,630 VITAL-AF participants without prevalent AF, 16,937 had pretrial ECG and clinical data. Each score discriminated 2-year AF risk according to AUROC (CHARGE-AF: 0.711 [95% CI: 0.671-0.749]; ECG-AI: 0.784 [95% CI: 0.743-0.819]; CH-AI: 0.788 [95% CI: 0.754-0.824]) and average precision (0.0952 [95% CI: 0.0836-0.112]; 0.132 [95% CI: 0.113-0.157]; 0.133 [95% CI: 0.117-0.159]). An AF screening effect was observed in the top decile of CH-AI (AF diagnosis rate in screening 10.07/100 person-years [95% 8.28-11.87] vs 7.76 [95% 6.30-9.21] in control, P < 0.05), corresponding to a difference in AF diagnosis rate of 2.32/100 person-years (95% CI: 0.01-4.63) and number-needed-to-screen of 43 per year.

CONCLUSIONS

Use of ECG-based AI and clinical factors identified individuals at particularly high risk for AF who may benefit from screening. Findings suggest a trade-off between increasing AF screening efficiency and decreasing population coverage (ie, restriction of the screening pool). Future studies are needed to determine whether a risk-based approach is optimal or whether consideration of additional clinical- and systems-level factors (eg, access, health care system engagement) can further refine AF screening strategies. (Screening for Atrial Fibrillation Among Older Patients in Primary Care Clinics [VITAL-AF]; NCT03515057).

CENTRAL ILLUSTRATION Estimated AF Risk and AF Screening Efficiency

Summary of the study and its implications. In the VITAL-AF trial of AF screening embedded in routine primary care, we assessed whether the AF screening effect varied according to AF risk estimated using several AF risk models, including CH-AI, a validated combination model incorporating clinical factors, and ECG-AI. We found a significant effect of screening among individuals in the top decile of estimated AF risk using CH-AI (bar graph, middle left), representing a multifactorial trade-off between greater screening efficiency and decreasing population coverage (ie, smaller screening pool) (scatterplot, middle right). In the scatterplot, colored points indicate statistically significant screening effects, and corresponding 2-year absolute risk estimates using CH-AI are shown for such points. Findings highlight current challenges and future opportunities for risk-guided AF screening. AF = atrial fibrillation; CH-AI = Cohorts for Aging Research in Genomic Epidemiology-Atrial Fibrillation plus artificial intelligence–based electrocardiogram; CHARGE-AF = Cohorts for Aging Research in Genomic Epidemiology-Atrial Fibrillation; ECG = electrocardiogram; ECG-AI = artificial intelligence–based electrocardiogram; ESC = European Society of Cardiology.

Atrial fibrillation (AF) is a common arrhythmia associated with substantial morbidity related to higher rates of heart failure, stroke, and premature cognitive decline.¹ Because AF may be asymptomatic and is commonly episodic, up to 25% of AF cases are undiagnosed.^2,3 As a result, there is substantial interest in mass screening to detect unrecognized AF, which may in turn facilitate preventive interventions (eg, anticoagulation in patients with stroke risk factors)⁴ designed to prevent AF-related morbidity.

However, contemporary assessments of population-based AF screening have demonstrated limitations (Box 1). Most screening efforts have used simple guideline-based age criteria (eg, screening individuals aged ≥65 years as endorsed in guidelines from the European Society of Cardiology⁵), leading to screening of many individuals at very low risk of AF. ⁶ As a result, discrete or single time point screening trials (eg, VITAL-AF [Screening for Atrial Fibrillation in Older Adults at Primary Care Visits], D2AF [Detecting and Diagnosing Atrial Fibrillation]) reported no increase in AF yield compared with usual care. More intensive screening interventions (eg, single-lead electrocardiogram [ECG] screening twice weekly for 1 year in REHEARSE-AF⁷ (Assessment of Remote Heart Rhythm Sampling Using the AliveCor Heart Monitor to Screen for Atrial Fibrillation) or long-term continuous screening with implantable loop recorders in LOOP [Implantable Loop Recorder Detection of Atrial Fibrillation to Prevent Stroke]⁸ ) have resulted in more AF detection but represent substantial resource utilization, tend to detect very-low-burden AF whose link to adverse outcomes is less clear,⁹ and have failed to demonstrate meaningful reductions in AF-related adverse events (eg, stroke, heart failure).^8,10

BOX 1.

Summary of Relevant AF Screening Trials and ECG-Based Artificial Intelligence Models

Study	Population	Screening Intervention	Effect of Screening on Newly Diagnosed AF (Time Point)
Randomized trials of general population AF screening
SAFE (Fitzmaurice et al, 2007)41	Primary care, aged ≥65 y	12-lead ECG or irregular pulse-guided 12-lead ECG	RD: 0.59%; 95% CI: 0.20-0.98 (1 y)
REHEARSE-AF (Halcox et al, 2017)7	Primary care, aged >65 y with CHA2DS2-VASc ≥ 2	Single-lead ECG twice weekly for 1 y	HR: 3.90; 95% CI: 1.4-10.4 (1 y)
mSToPS (Steinhubl et al, 2018)42	Insurance plan members aged ≥ 75 y, aged ≥55 y with male sex and ≥1 AF risk factor, or aged ≥65 y with female sex and ≥1 AF risk factor	Immediate vs delayed (4 mo) 14-d patch monitor	RD: 3.0%; 95% CI: 1.8-4.1 (4 mo)
D2AF (Uittenbogaart et al, 2020)33	Primary care, aged ≥65 y	Pulse palpation, blood pressure measurement with AF algorithm, and single-lead ECG	OR: 1.06; 95% CI: 0.84-1.35 (1 y)
LOOP (Svendsen et al, 2021)8	Aged 70-90 y with ≥1 additional stroke risk factor	Implantable loop recorder	HR: 3.17; 95% CI: 2.81-3.59 (64.5 mo)
VITAL-AF (Lubitz et al, 2022)15	Primary care, aged ≥65 y	Single-lead ECG	RD: 0.13%; 95% CI: −0.16-0.42 (1 y)
SAFER (Mant et al, 2024)43	Primary care, aged ≥70 y	Single-lead ECG 4 times per day for 3 wk	Ongoing (4 y)
GUARD:AF (Lopes et al, 2024)10	Primary care, aged ≥70 y	14-d patch monitor	RD: 1.7% per y (15.3 mo)
AMALFI (Wijesurendra et al, 2025)44	Primary care, aged ≥65 y with CHA2DS2-VASc ≥3 (men) or CHA2DS2-VASc ≥4 (women)	14-d patch monitor	Ratio of proportions: 1.26; 95% CI: 1.02-1.57
Study	Model Objective	Model Performance
Studies using ECG-based artificial intelligence to estimate AF risk
Attia et al12	Detection of prevalent paroxysmal AF from non-AF ECG	Area under the curve: 0.87 (95% CI: 0.86-0.88)
Raghunath et al13	Prediction of 1-y incident AF from non-AF ECG	Area under the curve: 0.85 (95% CI: 0.84-0.85)
Khurshid et al14	Prediction of 5-y incident AF from non-AF ECG	5-y time-dependent area under the curve: 0.75 (95% CI: 0.74-0.76)

AF = atrial fibrillation; AMALFI = Active Monitoring for AtriaL Fibrillation; D2AF = Detecting and Diagnosing Atrial Fibrillation; ECG = electrocardiogram; GUARD-AF = Reducing Stroke by Screening for Undiagnosed Atrial Fibrillation in Elderly Individuals; LOOP = Implantable Loop Recorder Detection of Atrial Fibrillation to Prevent Stroke; mSToPS = mHealth Screening to Prevent Strokes; RD = risk difference; REHEARSE-AF = Assessment of Remote Heart Rhythm Sampling Using the AliveCor Heart Monitor to Screen for Atrial Fibrillation; SAFE = Screening for Atrial Fibrillation in the Elderly; SAFER = Screening for Atrial Fibrillation with ECG to Reduce stroke; VITAL-AF = Screening for Atrial Fibrillation in Older Adults at Primary Care Visits.

Current findings compel the need for better approaches to selecting AF screening candidates. Approaches informed by estimated AF risk have potential to optimize precision and yield. Clinical risk scores such as the Cohorts for Aging Research in Genomic Epidemiology-AF (CHARGE-AF) have been extensively validated,¹¹ but clinical use has been limited by cumbersome calculation and potential misclassification of score inputs. Emerging ECG-based artificial intelligence (AI) models (ECG-AI) can estimate AF risk using the 12-lead ECG, a routine diagnostic test available in most clinic settings.^12–14 However, ECG-AI models have been tested primarily in retro-spective and hospital-based evaluations (Box 1), and thus their performance in the context of a real-world AF screening trial and potential utility for improving the yield of an AF screening intervention remains unknown. Here, we sought to assess whether the effect of single-lead ECG screening in the VITAL-AF trial varied as a function of estimated AF risk according to 3 AF risk models: the CHARGE-AF risk score, a previously validated ECG-AI model, and their combination (CH-AI).

METHODS

STUDY DESIGN AND PARTICIPANTS.

The design, conduct, and primary outcome results for the VITAL-AF (Screening for Atrial Fibrillation Among Older Patients in Primary Care Clinics; NCT03515057) trial have been published previously.¹⁵ Briefly, VITAL-AF was a pragmatic cluster-randomized trial of AF screening using a handheld, single-lead ECG device embedded in primary care practices. Sixteen Mass General Brigham practices (8 community-based and 8 on-campus) were randomized 1:1 to either AF screening or usual care. Eligible patients were aged ≥65 years and had an outpatient clinic appointment at a participating primary care practice with a primary care physician, nurse practitioner, or physician’s assistant. Enrollment took place between July 31, 2018 and October 8, 2019. Participants were followed for clinical outcomes for up to 2 years. Given the pragmatic nature of VITAL-AF, individuals with prevalent AF were enrolled; however, in the current analysis, we focused only on individuals without prevalent AF. Given our specific interest in the application ofAI models using the 12-lead ECG, we included only VITAL-AF participants with at least one 12-lead ECG performed within 3 years before enrollment. The research protocol was approved by the Mass General Brigham Institutional Review Board. Participants provided informed consent to participate.

ESTIMATION OF AF RISK.

We estimated AF risk using 3 previously published models: CHARGE-AF, an 11-component clinical score; ECG-AI, a deep learning–based model to estimate future AF risk using a single 12-lead ECG; and CH-AI, a linear combination of CHARGE-AF and ECG-AI. ECG-AI and CH-AI were developed in a primary care cohort at Massachusetts General Hospital. All 3 models have been validated in multiple populations including prospective cohort studies and health care–related datasets.^{11,14,16–18} CHARGE-AF was developed outside the Mass General Brigham system. For ECG-AI and CH-AI, to avoid evaluation of model performance among individuals included in model training (which may result in inflated model performance), we used versions of both models fit with VITAL-AF participants excluded. As a result, the version of CH-AI used in the current analysis had slightly different coefficients than previously published (ie, 0.36611 × CHARGE-AF + 0.43383 × logit-transformed ECG-AI).¹⁴ Therefore, our process ensured no models were informed by participants evaluated in the current analysis.

Baseline demographics and clinical factors were ascertained at the time of enrollment using data collected from electronic health records. Oral anti-coagulant exposure was ascertained using prescription data and included apixaban, dabigatran, edoxaban, rivaroxaban, and warfarin.

OUTCOMES.

The primary outcome event was incident AF, which was ascertained following a tiered protocol. New AF events in the screening period were captured by a validated electronic health record–based algorithm comprising International Classification of Diseases, 10th Revision, codes for AF or atrial flutter or a 12-lead ECG with AF or atrial flutter in the diagnostic statement. Subsequently, all potential incident AF events were manually adjudicated by 2 independent research nurses who reviewed all relevant components of the electronic health record. Given studies suggesting only moderate real-world accuracy,^19,20 single-lead ECG readings were not used to adjudicate new AF events. Any discrepancies in AF ascertainment were resolved through consensus or, in cases of continued uncertainty, by a study cardiologist. A similar process was used for the new AF events from the year before the study period.

STATISTICAL ANALYSIS.

Person-time began at enrollment and ended at the earliest of incident AF diagnosis, death, or end of the study period. Because the incidence of death during the study period was low (3.9%), death was treated as a censoring event. We first assessed discrimination of 2-year incident AF using the CHARGE-AF, ECG-AI, and CH-AI models. Discrimination was quantified using the time-dependent area under the receiver-operating characteristic curve (AUROC) as well as time-dependent average precision, which provides a complementary assessment emphasizing the ability to detect high-risk individuals. As time-dependent metrics, both AUROC and average precision values incorporated survival time and accounted for censoring using inverse probability of censoring weights.²¹ Test characteristics at specific operating points were similarly calculated. Discrimination was additionally assessed across subgroups of men vs women and among individuals enrolled in community-based vs on-campus practices. The cumulative risk of incident AF stratified by quartiles of each score was plotted using the Kaplan-Meier method.

Model calibration was assessed by plotting deciles of estimated AF risk vs observed AF incidence with corresponding smoothed curves obtained using adaptive hazard regression and by calculating the integrated calibration index, a measure of average prediction error weighted by the empirical risk distribution.²² As all 3 models were originally designed to estimate 5-year AF risk, we recalibrated the average predicted risk for each model to the observed 2-year AF hazard in VITAL-AF (ie, calibration-in-the-large) to obtain 2-year AF probabilities. Overlap in categorization of high-risk status across each model was summarized using Venn diagrams for the top quartile and top decile of risk. Reclassification across models was quantified by decile-based net reclassification indices (NRIs) and visualized using Sankey plots across quartiles and deciles of estimated AF risk.²³ We tabulated characteristics of individuals at high risk according to each model as well as AF cases among individuals not labeled as high risk according to any model.

Because all 3 models discriminated incident AF, we proceeded to assess whether the effect of AF screening (ie, difference in the person-time incidence rate of new AF diagnosis in screening vs control) varied across strata of estimated AF risk using each score. The number-needed-to-screen was defined as the inverse of the AF screening effect (ie, the number of people needing to be screened per year for each additional AF diagnosis).²³ Given our interest in the observed effect of AF screening, group assignment was performed as-treated (ie, individuals who received screening at least once, regardless of clinic assignment, were assigned to the screening group). In the absence of clinically accepted 2-year risk thresholds to guide AF screening, we first stratified estimated AF risk across sample quartiles. Noting potential heterogeneity in the highest quartile, we proceeded to assess the screening effect across more granular categories (ie, deciles). We tested for heterogeneity in the screening effect across deciles by fitting a Cox proportional hazards model with new AF diagnosis as the outcome and including an interaction term between group (screening vs control) and AF risk decile (both as a continuous variable and as a binary category for top decile vs bottom nine). Because the assessment of risk deciles was performed post-hoc and correction for multiple testing was not performed because of limitations in sample size, corresponding associations are exploratory and should be considered hypothesis-generating.

We performed several secondary analyses. First, we repeated quantification of AF screening effect according to each AF risk score within the intention-to-treat population (ie, group assignment was on the basis of whether the patient’s first visit was to an intervention vs control practice). Second, we calculated reclassification indices (ie, net gains in sensitivity and specificity)²⁴ using the top decile of CH-AI vs the guideline-based age threshold of ≥65 years (ie, all individuals in the analysis) and the top decile of CHARGE-AF. We also calculated percentile-based NRI (divided into decile categories) across the 3 AF risk models. Third, we calculated time-dependent diagnostic test characteristics (ie, sensitivity, specificity, precision/positive predictive value, negative predictive value, and F1 score) at 2 years for each model at top quartile- and top decile-based thresholds. Fourth, we assessed associations between CH-AI risk and rates of screen-detected AF (defined as a new AF diagnosis within 90 days after a positive 1-lead ECG screening result). Fifth, we assessed the AF screening effect in the top 5% of AF risk (instead of deciles) using each model. Sixth, we assessed the screening effect in the top decile of CH-AI across on-campus vs community practices. Seventh, we fit a version of our Cox proportional hazards models testing for interaction between group (screening vs control) and top AF risk decile with additional adjustment for clinic.

Analyses were performed using R version 4.4.0 (R Project for Statistical Computing). A significant effect was defined as either a 2-sided P < 0.05 or a 95% CI excluding the null.

RESULTS

STUDY SAMPLE.

Of 30,630 VITAL-AF study participants without prevalent AF, we excluded 447 for absent clinical data and 13,246 for absent ECG available within 3 years, resulting in 16,937 individuals in the primary analysis sample (Figure 1). Mean patient age was 75 ± 7.0 years, 57% were women, and 83% were White. Rates of cardiovascular comorbidity were appreciable (Table 1). Of the sample, 7,989 received screening and 8,948 received usual care.

FIGURE 1. Overview of Study Sample and Predicted Risk Distributions.

(A) Patient flow through the study, including relevant exclusions. (B) Smoothed probability density plots for the CHARGE-AF clinical score (top), ECG-AI model (middle), and CH-AI combined model (bottom). Models were converted to predicted 2-year AF probabilities (see text) and are shown separately for individuals with incident AF (red) and individuals without incident AF (blue). AF = atrial fibrillation; CH-AI = Cohorts for Aging Research in Genomic Epidemiology-Atrial Fibrillation plus electrocardiogram-based artificial intelligence; CHARGE-AF = Cohorts for Aging Research in Genomic Epidemiology-Atrial Fibrillation; ECG = electrocardiogram; ECG-AI = electrocardiogram-based artificial intelligence.

TABLE 1.

Sample Characteristics

	Overall (N = 16,937)	Screened (n = 7,989)	Control (n = 8,948)
Age, y	74.7 ± 7.0	74.6 ± 6.9	74.8 ± 7.1
Race
White	13,985 (83)	6,577 (82)	7,408 (83)
Black	863 (5.1)	436 (5.5)	427 (4.8)
Hispanic	369 (2.2)	196 (2.5)	173 (1.9)
Other	1,404 (8.3)	639 (8.0)	765 (8.5)
Unknown	316 (1.9)	141 (1.8)	175 (2.0)
Sex
Female	9,654 (57)	4,624 (58)	5,030 (56)
Male	7,283 (43)	3,365 (42)	3,918 (44)
Height, cm	164.8 ± 10.7	164.9 ± 10.8	164.8 ± 10.7
Weight, kg	76.9 ± 17.8	76.7 ± 17.7	77.1 ± 17.8
Systolic blood pressure, mm Hg	131.1 ± 16.4	131.3 ± 16.2	130.9 ± 16.6
Diastolic blood pressure, mm Hg	73.9 ± 8.7	74.4 ± 8.5	73.4 ± 8.8
Current smoker	856 (5.1)	367 (4.6)	489 (5.5)
Primary care visits in preceding year	5.2 ± 5.2	5.0 ± 5.1	5.5 ± 5.4
Use of antihypertensive medication	13,427 (79)	6,308 (79)	7,119 (80)
Diabetes	4,707 (28)	2,220 (28)	2,487 (28)
History of myocardial infarction	1,866 (11)	893 (11)	973 (11)
History of heart failure	2,648 (16)	1,257 (16)	1,391 (16)
CHARGE-AF scorea	13.7 ± 0.99	13.6 ± 0.98	13.7 ± 1.00
ECG-AI scorea	−0.57 ± 0.75	−0.59 ± 0.75	−0.55 ± 0.76
CH-AI scorea	4.6 ± 0.79	4.6 ± 0.78	4.6 ± 0.79

Values are mean ± SD or n (%).

^aScores can be converted to 2-y risk of AF in VITAL-AF as follows: CHARGE-AF: 1 – 0.9627745^{exp(ΣβX−13.65969)}; ECG-AI: 1 – 0.9640965^{exp(ΣβX+0.5710836)}; CH-AI: 1 – 0.9662516 ^{exp(ΣβX−4.59939)}, where ΣβX refers to the listed score value.

AF = atrial fibrillation; CH-AI = Cohorts for Aging Research in Genomic Epidemiology-Atrial Fibrillation plus electrocardiogram-based artificial intelligence; CHARGE-AF = Cohorts for Aging Research in Genomic Epidemiology-Atrial Fibrillation; ECG-AI = electrocardiogram-based artificial intelligence.

Baseline characteristics were similar between both groups, including distributions of estimated AF risk (Table 1). The median months from ECG to study entry was also similar for screening (10.5 months [IQR: 4.5-19.8 months]) vs usual care (11.0 months [IQR: 4.9-19.9 months]). Compared with participants enrolled in on-campus practices, participants enrolled in community practices had moderately higher rates of comorbidity (eg, diabetes: 37% vs 24%; P < 0.01) and more primary care office visits in the preceding year (6.9 vs 4.6; P < 0.01) (Supplemental Table 1). Individuals excluded for missing ECG or absent clinical data had demographic characteristics largely similar to those of included individuals but had moderately lower rates of coISP]-morbidity (eg, diabetes: 19% vs 28%; P < 0.01) and fewer primary care office visits in the preceding year (3.1 vs 5.2, P < 0.01) (Supplemental Table 2).

DISCRIMINATION OF 2-YEAR AF RISK.

At 2 years, 664 new AF cases were diagnosed over a median of 1.8 years (IQR: 1.6-1.9 years), corresponding to a 2-year incidence of 4.5% (95% CI: 4.1-4.8). CHARGE-AF, ECG-AI, and CH-AI each discriminated 2-year incident AF according to both AUROC (CHARGE-AF: 0.700 [95% CI: 0.654-0.739]; ECG-AI: 0.780 [95% CI: 0.747-0.810]; CH-AI: 0.781 [95% CI: 0.740-0.814]) and average precision (CHARGE-AF: 0.0935 [95% CI: 0.0810-0.110]; ECG-AI: 0.129 [95% CI: 0.110-0.151];

CH-AI: 0.131 [95% CI: 0.113-0.153]) (Figure 2). Compared with CHARGE-AF, CH-AI had greater discrimination according to AUROC (improvement: 0.081 [95% CI: 0.053-0.11]) and average precision (improvement: 0.0380 [95% CI: 0.0246-0.0535]).

FIGURE 2. Model Performance for 2-Year Incident AF Prediction.

(A) Discrimination of 2-year AF risk using CHARGE-AF (orange), ECG-AI) (green), and CH-AI (blue). The left plot depicts receiver-operating characteristic curves and the right plot depicts precision-recall curves. The legends depict the area under the curve (left) and the average precision (right).

(B) Calibration plots of predicted 2-year AF risk (x-axis) vs observed 2-year AF incidence (y-axis). Each plot contains 10 points representing score deciles and a smoothed curve obtained using adaptive hazard regression. On each plot is shown the integrated calibration index, a time-dependent error metric where lower values demonstrate smaller average error (see text). Models are represented by the same colors as in (A). PPV = positive predictive value; other abbreviations as in Figure 1.

Test characteristics for each model at quartile- and decile-based thresholds are shown in Supplemental Table 3. Distributions of 2-year estimated AF risk according to each model are shown in Figure 1. Cumulative risk of incident AF stratified by quartiles of each score are shown in Figure 3. Discrimination of AF risk using CH-AI was similar among individuals enrolled in community practices (AUROC: 0.781 [95% CI: 0.716-0.840]) vs on-campus practices (AUROC: 0.790 [95% CI: 0.748-0.834]) and was slightly higher among women (AUROC: 0.798 [95% CI: 0.753-0.842]) vs men (AUROC: 0.746 [95% CI: 0.681-0.803]).

FIGURE 3. Cumulative Risk of Incident AF Stratified by Quartiles of AF Risk.

Plotted is the Kaplan-Meier cumulative risk of incident AF stratified by quartiles of the 3 scores analyzed in the current study: CHARGE-AF (top left), ECG-AI (top right), and CH-AI (bottom). Abbreviations as in Figure 1.

CALIBRATION AND RECLASSIFICATION.

Calibration was generally good across each model, with the lowest prediction error observed for ECG-AI (0.0029 [95% CI: 0-0.069]), followed by CH-AI (0.0034 [95% CI: 0-0.0086) and CHARGE-AF (0.017 [95% CI: 0.013-0.021]), with CHARGE-AF demonstrating a general tendency to overestimate AF risk (Figure 2). Calibration of 2-year AF risk in the top quartile (CHARGE-AF: predicted 15.5% vs observed 9.0%; ECG-AI: 10.9% vs 10.3%; CH-AI: 10.6% vs 10.6%) and top decile (CHARGE-AF: 24.5% vs 12.4%; ECG-AI: 16.7% vs 15.0%; CH-AI: 16.0% vs 15.7%) was favorable for ECG-AI and CH-AI compared with CHARGE-AF.

Absolute risk estimates corresponding to quartiles and deciles of each score are shown in Supplemental Table 4. Among 6,251 individuals in the top quartile of any score, 2,229 (36%) were in the top quartile of all scores, whereas among 2,797 individuals in the top decile of any score, 635 (28%) were in the top decile of all scores (Figure 4). Reclassification across risk quartiles and deciles was greatest using CHARGE-AF vs ECG-AI, compared with other combinations (Figure 4 and Supplemental Figure 1).

FIGURE 4. Concordance and Reclassification Among AF Risk Models.

(A) Scatterplots demonstrating correlation between the standardized linear predictors of CHARGE-AF and ECG-AI (left), CHARGE-AF and CH-AI (middle), and ECG-AI and CH-AI (right). Pearson correlations are shown on the bottom right of each plot. (B) Venn diagrams summarizing the distribution of model overlap among individuals in the top quartile of any score (left) or top decile of any score (right). (C) Sankey plots summarizing flow from the top quartile of CHARGE-AF to CH-AI (left) and the top decile of CHARGE-AF to CH-AI (right). Abbreviations as in Figure 1.

Decile-based NRI was favorable using CH-AI (NRI: 0.18 [95% CI: 0.11-0.25]; NRI+: 14.4% [95% CI: 7.320.9]; NRI−: 3.6% [95% CI: 2.3-4.9]) and ECG-AI (NRI: 0.09 [95% CI: 0.02-0.16]; NRI+: 7.3% [95% CI: 5.914.8]; NRI−: 1.5% [95% CI: 1.0-2.8]) vs CHARGE-AF as well as using CH-AI (NRI: 0.11 [95% CI: 0.04-0.17]; NRI+: 8.2% [95% CI: 1.8-14.2]; NRI−: 2.3% [95% CI: 1.1-3.5]) vs ECG-AI. Favorable reclassification was also observed using the top decile of CH-AI to identify high-risk patients compared with either all individuals aged ≥65 years (ie, the full study sample) (NRI: 0.25 [95% CI: 0.21-0.29]; NRI+: −66.1% [95% CI: −61.8 to −69.9]; NRI−: 91.1% [95% CI: 90.7-91.6]) or the top decile of CHARGE-AF (NRI: 0.07 [95% CI: 0.04-0.11]; NRI+: 7.1% [95% CI: 3.7-10.4]; NRI−: 0.33% [95% CI: −0.12 to 0.79]).

High-risk individuals according to CHARGE-AF tended be older and had higher comorbidity and more primary care office visits in the preceding year compared with individuals labeled as high-risk according to ECG-AI, whereas high-risk individuals using CH-AI were intermediate. AF cases who were not high risk according to any model tended to have fewer primary care office visits and were more frequently women (Supplemental Tables 5 and 6).

AF SCREENING EFFECT ACCORDING TO AF RISK.

There was no significant screening effect (ie, difference in incidence rate of new AF diagnosis in screening vs control) in the overall sample (0.09 per 100 person-years [95% CI: −0.27 to 0.45]). When assessed across quartiles of each AF risk model, the AF screening effect was consistently highest in the top quartile of risk, although effects were not statistically significant (difference in AF diagnosis rate in screening vs control using CHARGE-AF: 0.49 per 100 person-years [95% CI: −0.62 to 1.60]; ECG-AI: 0.84 per 100 person-years [95% CI: −0.21 to 1.90 per 100 person-years]; CH-AI: 0.61 per 100 person-years [95% CI: −0.53 to 1.75 per 100 person-years]) (Table 2).

TABLE 2.

Screening Effect Across Quartiles of AF Risk

Score	Quartile	AF Incidence in Screened Group (95% CI) (per 100 person-y)	AF Incidence in Control Group (95% CI) (per 100 person-y)	Difference in AF Incidence (95% CI) (per 100 person-y)
CHARGE-AF	1	1.02 (0.69-1.35)	0.62 (0.36-0.88)	0.40 (−0.03 to 0.82)
	2	1.10 (0.75-1.46)	1.11 (0.78-1.45)	−0.01 (−0.50 to 0.48)
	3	2.22 (1.71-2.73)	2.44 (1.94-2.94)	−0.22 (−0.94 to 0.49)
	4	5.62 (4.79-6.46)	5.13 (4.40-5.86)	0.49 (−0.62 to 1.60)
ECG-AI	1	1.08 (0.73-1.44)	0.79 (0.50-1.08)	0.29 (−0.16 to 0.75)
	2	1.00 (0.66-1.34)	1.24 (0.88-1.60)	−0.24 (−0.74 to 0.25)
	3	2.37 (1.84-2.89)	2.86 (2.32-3.41)	−0.50 (−1.25 to 0.26)
	4	5.33 (4.53-6.13)	4.49 (3.80-5.17)	0.84 (−0.21 to 1.90)
CH-AI	1	0.76 (0.48-1.05)	0.51 (0.28-0.75)	0.25 (−0.12 to 0.62)
	2	1.33 (0.94-1.73)	0.90 (0.59-1.20)	0.44 (−0.06 to 0.93)
	3	1.97 (1.49-2.44)	2.58 (2.07-3.10)	−0.62 (−1.32 to 0.09)
	4	5.93 (5.07-6.79)	5.32 (4.57-6.06)	0.61 (−0.53 to 1.75)

Abbreviations as in Table 1.

Because of numerically greater effects at the highest levels of risk, we proceeded to assess the screening effect across more granular groups (ie, deciles). Using CH-AI, the screening effect was significant in the top decile (ie, estimated 2-year AF risk >9.4%; AF diagnosis rate in screening: 10.07 per 100 person-years [95% CI: 8.28-11.87 per 100 person-years] vs 7.76 per 100 person-years [95% CI: 6.30-9.21 per 100 person-years] in control subjects; difference: 2.31 per 100 person-years [95% CI: 0.01-4.63 per 100 person-years]). The observed screening effect in the top decile of CH-AI translated to an additional AF case per 43 individuals screened each year. Findings were qualitatively similar using the top decile of CHARGE-AF and ECG-AI, but the screening effect was not statistically significant (Figure 5). In a Cox proportional hazards model with incident AF as the outcome, there was no interaction between screening group and CH-AI risk decile as a continuous variable (P = 0.70), but there was an interaction between screening group and CH-AI risk in the top decile vs bottom nine (interaction HR: 1.39 [95% CI: 1.01-1.92]; P = 0.04). Clinical characteristics between the screening and control groups remained generally well-balanced within high-risk groups (Supplemental Tables 7 and 8).

FIGURE 5. AF Screening Effect Across Deciles of Estimated Risk.

New AF diagnosis rate (per 100 person-years) across deciles of estimated AF risk using CHARGE-AF (A), ECG-AI (B), and CH-AI (C). Individuals in the screening group are depicted in blue, whereas individuals in the control group are depicted in green. Error bars represent 95% CIs. Abbreviations as in Figure 1.

SECONDARY ANALYSES.

A larger AF screening effect was seen among individuals in the top 5% of estimated AF risk using CH-AI (AF diagnosis rate in screening: 13.71 per 100 person-years [95% CI: 10.61-16.82 per 100 person-years] vs 9.42 per 100 person-years [95% CI: 7.16-11.68 per 100 person-years] in control subjects; difference: 4.29 per 100 person-years [95% CI: 0.45-8.12 per 100 person-years]), corresponding to an additional AF case per 23 individuals screened each year (Supplemental Figure 2). The observed screening effect in the top decile of CH-AI was higher in community practices (difference in AF diagnosis rate in screening vs control: 5.56 per 100 person-years [95% CI: 0.66-10.47]) vs on-campus practices (1.31 per 100 person-years [95% CI: −1.41 to 4.02 per 100 person-years]). Findings were generally similar when group assignment was performed according to the intention-to-treat principle (group assignment was on the basis of the patient’s first visit to a study practice during the study period) (Supplemental Table 9, Supplemental Figures 3–5). The rate of screen-detected AF increased with greater CH-AI risk (Supplemental Table 10). The estimated interaction effect between screening group and CH-AI risk in the top decile vs bottom nine was similar after adjustment for study practice but was no longer statistically significant (interaction HR: 1.37 [95% CI: 0.99-1.89]; P = 0.06).

DISCUSSION

Among >16,000 participants in the VITAL-AF trial of AF screening embedded in routine primary care, we systematically assessed variation in the AF screening effect according to AF risk estimated using 3 validated models: the CHARGE-AF clinical score; ECG-AI, an ECG-based deep learning model; and CH-AI, a combination of CHARGE-AF and ECG-AI. We had 2 primary findings. First, each AF risk model discriminated 2-year AF in a real-world sample of primary care patients, with important differences among individuals identified as high risk across models. Second, an AF screening effect was observed among individuals in the top 10% of AF risk using CH-AI, in whom screening was associated with 1 new AF case for every 43 persons screened each year. Our results suggest the potential of precision-guided AF screening using contemporary risk models while providing insight into current limitations, trade-offs, and the potential importance of considering systems-level factors beyond AF risk (Central Illustration).

Our results support and extend prior analyses investigating targeted approaches to AF screening. Based on randomized trial evidence from the early 2000s demonstrating higher AF detection with pulse palpation or 12-lead ECG, guidelines from the European Society of Cardiology and the Cardiac Society of Australia and New Zealand recommend mass screening of individuals aged ≥65 years.^27,28 More recent trials, including the VITAL-AF study on which the current analysis is based, have shown minimal yield of AF screening when targeting all individuals aged ≥65 years in the context of contemporary care. Prior analyses have suggested that leveraging AF risk factors beyond age (eg, clinical factors, genetic risk^29,30 and N-terminal pro–B-type natriuretic peptide³¹) may facilitate more efficient AF case detection. One study applying an ECG-based AI AF risk model demonstrated about a 3-fold increase in AF diagnoses compared with historical control subjects.³² The current analysis extends prior observations by quantifying the specific potential of AI-enabled risk models to increase the yield of AF detection within the context of a real-world AF screening trial comprising individuals meeting current guideline-based criteria for AF screening.

Our findings support the following implications. First, AI-enabled analysis of the 12-lead ECG, with or without clinical factors, is a feasible mechanism to identify older primary care patients at elevated risk for near-term AF. Although several ECG-based AI models have been validated for AF prediction, most have been tested in retrospective and hospital-based evaluations.^13,14 Here, we provide important evidence that both clinical and AI-based models stratify AF risk within a large sample of primary care patients meeting current guideline-based criteria for AF screening (ie, the very population in which AF risk estimation is most useful). Of note, the performance of CHARGE-AF (AUROC: 0.710) was on the lower end of the range reported in prior validation studies (AUROC: ~0.7-0.8),^16,17 consistent with reduced informativeness of clinical factors in older populations with high comorbidity burden.¹⁸ Importantly, our analyses demonstrated that individuals identified as high risk were not the same across each model. The AI-based algorithms tended to identify younger individuals with less traditional cardiac comorbidity, suggesting particular value for prioritizing individuals who may otherwise be missed by standard risk stratification. Yet, CH-AI did provide slightly higher discrimination and favorable reclassification vs ECG-AI, suggesting that consideration of clinical factors alongside AI-based risk signals provides some complementary value. Of note, AF cases missed by all 3 models tended to be younger, have less health care engagement, and were more commonly women (despite higher discrimination among women vs men using CH-AI), highlighting important opportunities for future model improvement. On balance, our results support the validity of contemporary AF risk stratification methods in older primary care patients who may be considered for AF screening and specifically the incorporation of AI-enabled ECG analysis.

Second, application of contemporary AF risk models to guide AF screening may represent a multifactorial trade-off between increasing AF screening efficiency and decreasing population coverage (ie, restriction of the screening pool). Consistent with the primary outcome results of VI-TAL-AF,^33,34 there was no difference in new AF diagnoses with screening in our overall sample.¹⁵ However, we did observe a substantial screening effect (2.6 per 100 person-years, or number-needed-to-screen of 43 people per year) in the top decile of CH-AI. The observed screening effect is favorable to the incremental AF yield reported in other discrete time point screening studies (generally 1 for every 100 individuals screened, if not more).²³ Notably, the AF screening effect appeared to inflect at the top decile of CH-AI (corresponding to an estimated 2-year AF risk of >9.4%) rather than increase gradually. We posit the inflection may relate to the specific screening strategy applied (discrete time point single-lead ECG), which may require individuals to be at particularly high risk of incident AF (or pre-existing unrecognized AF) to have AF detected at appreciable rates over no screening.^35,36 Estimated AF risk may also enrich for greater AF burden, which is more likely to be detected using brief discrete time point screening. Of note, because AF yield varies based on screening modality (eg, it is lower with a single time point vs implanted loop recorders), our findings may differ with application of other screening methods. Nevertheless, for the screening intervention considered here, our observations clearly reveal a trade-off between AF screening efficiency and population selectivity (ie, excluding many individuals from screening eligibility). Yet, risk-based approaches are well accepted in other clinical contexts (eg, aortic aneurysm screening for middle-aged men with a smoking history).³⁷ We acknowledge that the modality we used for AF risk estimation (ECG) is itself an AF screening method; nevertheless, we submit that the ECG possesses multiple advantages, including widespread clinical availability, potential to facilitate opportunistic risk stratification using previously collected tracings,³² and ready applicability to consumer-facing devices, where AF risk estimates from non-AF ECGs could guide the frequency and intensity of subsequent screening. Furthermore, we submit that several key insights provided by our findings (eg, heterogeneity in risk profiles, efficiency and coverage trade-offs, importance of systems-level factors) should possess relevance for precision-guided AF screening beyond the specific screening strategies or modalities used. Of note, because discrimination of our best performing models remained only moderate (AUROC: 0.78), future work is warranted to assess whether consideration of additional features (eg, biomarkers, genetic risk scores) may allow sufficient enrichment of AF risk to optimize AF screening efficiency without severely restricting the screening pool.

Third, consideration of factors beyond AF risk estimation may be a key component of optimized AF screening. Although our observation of a screening effect in the top decile of CH-AI is promising, a converse interpretation is that despite application of contemporary AF risk models, we could only identify an AF screening effect in 10% of the population.³⁸ To this end, it is notable that we observed larger AF screening effects among individuals identified as high risk using CH-AI compared with ECG-AI, despite similar degrees of risk discrimination and preserved balance in clinical characteristics between screening and control groups. Potential drivers of this difference are therefore difficult to identify with certainty and may have included modest improvements in calibration at the high end of the risk distribution using CH-AI vs ECG-AI. Findings further suggest that individuals identified using CH-AI (who tended to be older, have slightly greater health care system engagement, and have more comorbidity when compared with those identified by ECG-AI) had characteristics predictive not only of AF, but AF that would otherwise go undetected without screening.³⁸ The potential importance of systems-level factors is also supported by our observation of a larger targeted screening effect using CH-AI in community-based vs on-campus practices. Recent work suggests that direct estimation of AF screening effects, which may include systems-level factors such as access to care, health care engagement, and other sociodemographic variables,³⁸ is an alternative approach to improving AF screening efficiency. A key limitation of screening effect models (and relative advantage of risk-based approaches), however, is a tendency for effect models to be highly specific to the intervention and setting used for model development. Therefore, future strategies may leverage both underlying AF risk and predicted AF screening effectiveness in a composite approach specific to the AF screening intervention applied. Regardless of enrichment mechanism, key implementation steps (eg, integration of risk estimates into the electronic health record and potential deployment of AF risk estimation using digital health technologies including consumer wearable or handheld ECGs)³⁹ will be critical to facilitate scalable identification of AF screening candidates. In future work, we seek to deploy models such as CH-AI to facilitate opportunistic interventions whereby individuals identified as high risk on the basis of pre-existing clinical data are offered proactive screening, potentially in a randomized design. Ultimately, the most important proof of risk-guided screening will require demonstration that targeted AF detection translates into meaningful improvements in AF-related outcomes including ischemic stroke, death, and systemic thromboembolism.

STUDY LIMITATIONS.

Our analysis should be considered in the context of design. First, although we submit analysis of a primary care population meeting current age-based guidelines for AF screening (ie, aged ≥65 years) is a strength, we were unable to evaluate whether use of AF risk may identify younger individuals in whom AF screening may be effective. Second, given our specific interest in the potential utility of ECG-based AI models, we excluded participants without a baseline ECG, which introduces bias toward individuals with greater comorbidity and interaction with the health care system. We submit that such selection (affecting both screening and control groups) cannot explain our observation of a screening effect in the top decile of CH-AI, yet we acknowledge that prospective evaluations are needed to evaluate risk-based AF screening using protocolized ECG acquisition. Studies are also needed to assess whether digital health technology (eg, mobile single-lead ECG) may extend the reach of ECG-based risk stratification to individuals with more limited access to care. Third, we were unable to assess the effect of screening on AF-related adverse events (eg, stroke, bleeding) as function of estimated AF risk, because VITAL-AF was not powered for these outcomes. Fourth, because our analysis is post-hoc, the screening effect is no longer randomized and is subject to potential confounding. However, we note that baseline factors, including estimated AF risk, were similar between groups (both overall and within high-risk categories). Fifth, we did not adjust for multiplicity due to sample size, and therefore our observations regarding screening effects in the top quantiles of CH-AI are exploratory and should be considered hypothesis-generating. We cannot exclude chance variation at the high end of the predicted risk distribution. Yet, we note that directionally consistent effects across 3 different AF risk models and consistent findings in the top 5% risk category support the validity of our findings. Sixth, although our study included 16 primary care practices including 8 community health centers and 8 on-campus clinics with varying patient characteristics, our analysis sample remains subject to demographic (eg, predominantly White) and geographic specificity, and generalizability of our findings to other populations remains unclear. Future studies are needed to assess AF risk estimation to guide AF screening in more diverse samples. Seventh, given our intent to evaluate AI-enabled analysis of the ECG (ie, a relatively inexpensive test available in most clinic settings), we did not assess the potential utility of other AF risk signals (eg, biomarkers,³¹ imaging characteristics,³¹ genetic scores)^29,30,40 to guide AF screening. Eighth, to maximize the applicability of our findings, we evaluated previously published external models without modification. Accordingly, we did not rederive or fine-tune AF risk models in VITAL-AF or modify underlying architectures (eg, use a more complex approach than the simple linear combination of ECG-AI and CHARGE-AF used in the published version of CH-AI¹⁴), any of which may have increased predictive performance but also potentially reduced generalizability. Ninth, given VITAL-AF’s design as a pragmatic AF screening trial embedded in routine primary care, AF was adjudicated using clinical diagnosis. Although our incidence aligns with other similarly designed studies³³ and may enrich for higher burden AF, which is more clinically action-able, true AF incidence may be underestimated.

CONCLUSIONS

In VITAL-AF, a large clinical trial of discrete time point AF screening embedded in routine primary care, the use of ECG-based AI and clinical factors was able to identify individuals at particularly high risk for AF who may benefit from screening. Overall findings suggest a multifactorial trade-off between increasing AF screening efficiency and decreasing population coverage (ie, restriction of the screening pool). Future studies are needed to determine whether a risk-based screening approach is optimal or whether consideration of additional clinical (eg, biomarkers) or systems-level (eg, access to care, health care system engagement, screening setting) factors can further refine AF screening strategies.

ABBREVIATIONS AND ACRONYMS

AF

atrial fibrillation

AUROC

area under the receiver-operating characteristic curve

CH-AI

Cohorts for Aging Research in Genomic Epidemiology-Atrial Fibrillation plus electrocardiogram-based artificial intelligence

CHARGE-AF

Cohorts for Aging Research in Genomic Epidemiology-Atrial Fibrillation

ECG

electrocardiogram

ECG-AI

electrocardiogram-based artificial intelligence

NRI

net reclassification index

APPENDIX

For supplemental tables and figures, please see the online version of this paper.

CODE AVAILABILITY.

The ECG-AI AF risk model used in the current analysis is publicly available under an academic use license.²⁵ The CHARGE-AF risk model weights have been published previously,¹¹ and the CH-AI weights are included in Methods. Scripts underlying the statistical analysis are also publicly available.²⁶

DATA AVAILABILITY

VITAL-AF trial data contain protected health information and cannot be shared publicly.