This study was conducted in accordance with the 1975 Declaration of Helsinki. It was approved by the Medical Ethics Committee of Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine (Approval Number: XHEC-C-2024-190-1) and registered at the Chinese Clinical Trial Registry (Registration ID: ChiCTR2400094326; URL: http://www.chictr.org.cn). All patients provided written informed consent for study participation.

According to the preliminary experiment, we used G*power 3.1 software (Heinrich Heine University Düsseldorf, Germany) to estimate the required sample size. The analysis was conducted with the following parameters: the effect size f² = 0.15, 1 − β = 80%, α = 0.05, and the number of predicted variables was 5–7. Finally, 92–103 participants were needed.

This prospective observational cohort study initially enrolled 145 AF patients who underwent RFCA at Xinhua Hospital affiliated to Shanghai Jiao Tong University School of Medicine, between May 2024 and February 2025.

All AF patients were diagnosed via standard 12-lead ECG or 24-hour Holter monitoring, and classified according to the 2023 ACC/AHA/ACCP/HRS guidelines (paroxysmal AF: self-terminating within 48 h; persistent AF: sustained > 7 days or requiring cardioversion; long-standing persistent AF: Continuous > 12 months) (1). Antiarrhythmic drugs (e.g., amiodarone) were discontinued for ≥ 5 half-lives pre-procedure. Within 48 h before the procedure, transesophageal echocardiography was performed to exclude intracardiac thrombus, supplemented by cardiac computed tomography angiography when clinically feasible. All procedures were performed under local anesthesia and guided by the CARTO3 navigation system (Biosense Webster, Inc., Irvine, USA). The THERMOCOOL SMARTTOUCH SF catheter was used as its 56-hole tip irrigation facilitating cooling at low flow rate, thus easing the fluid management process. Pulmonary vein isolation (PVI) was performed in all patients. Additional ablation including left atrial roof line, anterior septal, posterior and inferior lines, mitral isthmus (MI) and cavo-tricuspid isthmus (CTI) lines, complex fractionated electrograms (CFAE) modification, and ablation of ganglionated plexi and extra-PV triggers, were performed when deemed necessary. For patients not achieving sinus rhythm post-ablation, low-energy (≤15 J) intracardiac cardioversion was delivered via catheters positioned in the right atrium and coronary sinus/left atrium. All operations were performed by experienced physicians (>50 annual cases).

Inclusion criteria were: AF patients aged between 40 and 80 years old who underwent RFCA in the last 3–12 months (3), documented sinus rhythm with a heart rate ranging from 60 to 100 beats per minute, and with written informed consent given. Exclusion criteria included: patients with contraindications to CPX as defined by the American Heart Association, significant musculoskeletal system diseases (e.g., fractures, serious soft tissue injuries) or severe chronic diseases (e.g., cerebrovascular, pulmonary, hepatic, or renal impairment), patients with cognitive dysfunction, and those who had participated in other intervention trials within the past 90 days. After excluding atrial flutter (n = 1), non-consent (n = 1), and severe respiratory comorbidities (n = 2), data from 141 participants were analyzed (Supplementary Figure S1).

The retrospective data of all patients included: (1) Demographic and anthropometric data: Age, gender, height, weight, and waist circumference, body mass index (BMI), blood pressure, smoking history, alcohol consumption, medical history, and current medications were recorded. The BMI was calculated as weight (kg) divided by height squared (m²). (2) Laboratory data: Fasting venous blood samples were collected in the morning to measure N-terminal pro-B-type natriuretic peptide (NT-proBNP), hemoglobin (HGB), glucose (GLU), serum creatinine (Cr), low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), total cholesterol (TC), triglycerides (TG) and estimated glomerular filtration rate (eGFR) levels. (3) Cardiac function data: Two-dimensional transthoracic echocardiography was used to measure left ventricular end-diastolic dimension (LVEDD), left atrial anteroposterior diameter (LAD), left ventricular ejection fraction (LVEF) and systolic pulmonary artery pressure (PAP).

Cardiopulmonary exercise test was conducted on an electronically braked cycle ergometer according to American Heart Association guidelines (13). The baseline phase included 3 min of seated rest. The warm-up phase consisted of 3 min of unloaded cycling (55–65 rpm). In the incremental phase, the workload was continuously increased at a constant rate of 10–15 watts per minute (ramp protocol) until the patient experienced voluntary fatigue or met the termination criteria. The recovery phase was 3 min of unloaded cycling (30 rpm). Throughout the process, output from a continuous 12-lead electrocardiogram and pulse oximetry were monitored, and blood pressure was measured every 3 min. VO_2peak was directly measured using the Quark PFT4 Ergo metabolic cart (COSMED, Italy) and defined as the highest 30 s average value during maximal effort, normalized to body weight (ml·kg⁻¹·min⁻¹). METs_max was calculated as VO_2peak divided by 3.5 ml·kg⁻¹·min⁻¹ (13).

(1) Time up-and-go test (TUG): Participants sat on a 46 cm high chair with back against the chair, arms resting on the chair's arms. Upon receiving the “Go” command, they stood up and walked at a comfortable and safe pace to a line on the floor 3 m away, turned, returned to the chair, and sat down again. The total completion time was recorded in seconds (19). (2) Five-times sit-to-stand test (FTSTS): Participants performed five consecutive sit-to-stand cycles from a 46 cm high chair placed against a wall. The initial position required the ankles to be in a neutral alignment, with feet flat and the arms folded across the chest. Upon receiving the “Go” command, participants were instructed to fully extend their knees and hips during the standing phase and ensure complete contact with the chair during the sitting phase (20). The total time to complete five cycles was recorded. (3) 6-minute walk distance (6MWD): The 6MWD was performed according to the American Thoracic Society guidelines on a 30 m indoor walkway with colored cones marking turn-around points (21). A certified cardiac rehabilitation therapist assessed the baseline heart rate and blood pressure, and gave standardized instructions. Participant then walked for six min at their self-selected maximal pace. Post-test measurements, including 6MWD, heart rate and blood pressure during the recovery phase were recorded.

All participants abstained from caffeine for ≥12 h and fasted for ≥3 h before the tests, and wore comfortable clothing and shoes. To minimize the interference of fatigue, all motor function tests were completed within one week. TUG and FTSTS were performed on the same day with 5 min seated recovery interval between them, whereas the 6MWD commenced precisely 30 min after the completion of TUG/FTSTS. CPX was conducted on a separate day.

All statistical analyses were conducted using SPSS 25.0 (IBM Ink., Armonk, NY) under the supervision of a medical statistician expert. Normally distributed continuous variables were analyzed using Student's t-test, reported as mean ± SD. Non-normally distributed continuous variables were compared via the Mann–Whitney U test with median (Q1, Q3) presentation. Categorical variables were analyzed by the Chi—square (χ²) test, and the results were presented as percentages (%). Participants were randomly allocated into a derivation cohort (n = 105) and a validation cohort (n = 36) at a ratio of 3:1. Variables demonstrating statistical trends (P < 0.1) in the correlation analysis with VO_2peak and METs_max were initially selected as candidate independent variables. Stepwise regression (forward entry α ≤ 0.05, backward retention α ≥ 0.1) was employed to develop the predictive equations. The robustness of the models was evaluated through multiple means. Normality was assessed using histograms of standardized residuals. Multicollinearity was examined by calculating the variance inflation factor (VIF), with a threshold of VIF < 5. Residual independence was evaluated using the Durbin—Watson statistic, with an acceptable range of 1.5–2.5. The goodness-of-fit of the models was determined by F-tests from ANOVA, and reported as adjusted R². For internal validation, predicted and measured values in the validation cohort were compared using Pearson correlation coefficients and paired Student's t-test, with Bland-Altman agreement analysis to evaluate accuracy and systematic bias. For all analyses, P < 0.05 indicated statistically significant.

This study enrolled 141 AF patients who underwent RFCA, with detailed characteristics presented in Table 1. The average age was 68.7 ± 7.58 years, including 62.4% males and 56.7% persistent AF patients. Comorbidities included hypertension (96 patients), diabetes (27 patients), CAD (83 patients), and prior stroke (31 patients). Among these patients, 54.6% received ACEIs/ARBs/ARNI therapy, 81.6% NOACs, 86.5% amiodarone, 53.2% β-blockers, 23.4% antiplatelet agents, and 68.1% statins. Echocardiographic parameters demonstrated a mean LAD of 38.98 ± 4.96 mm, LVEED of 47.6 ± 4.61 mm, and LVEF of 65.72 ± 5.21%. Motor function tests revealed FTSTS (13.85 ± 4.08 s), TUG (9.35 ± 2.81 s), and 6MWD (397.0 8 ± 60.4 m). CPX results indicated VO_2peak and METs_max were 16.29 ± 3.20 ml·kg⁻¹·min⁻¹ and 4.63 ± 0.98 ml·kg⁻¹·min⁻¹, respectively. No significant differences existed in demographic, blood biochemistry, motor function, and cardiopulmonary function parameters between validation and derivation cohorts (P > 0.05).

As summarized in Table 2, VO_2peak in AF patients was significantly correlated with multiple indicators including age (r = −0.296, P = 0.002), height (r = 0.301, P = 0.002), BMI (r = −0.269, P = 0.005), gender (r = −0.451, P < 0.001), LAD (r = −0.306, P = 0.002), NT-proBNP (r = −0.379, P < 0.001), hemoglobin (r = 0.316, P = 0.001), eGFR (r = 0.309, P = 0.001), FTSTS (r = −0.303, P = 0.002), TUG (r = −0.253, P = 0.009), 6MWD (r = 0.388, P < 0.001), and systolic blood pressure (r = −0.326, P = 0.001). METs_max also showed significant correlations with multiple indicators including age (r = −0.305, P = 0.002), height (r = 0.229, P = 0.002), BMI (r = −0.272, P = 0.005), gender (r = −0.448, P < 0.001), LAD (r = −0.301, P = 0.002), NT-proBNP (r = −0.369, P < 0.001), hemoglobin (r = 0.314, P = 0.001), eGFR (r = 0.302, P = 0.002), FTSTS (r = −0.308, P = 0.001), TUG (r = −0.263, P = 0.007), 6MWD (r = 0.4, P < 0.001), and systolic blood pressure (r = −0.332, P = 0.001). Other indicators such as weight, LVEF, and lipid levels showed no significant correlation with VO_2peak and METs_max.

The final regression equation was: VO_2peak (ml·kg⁻¹·min⁻¹) = 35.080 − (0.286 * BMI [kg/m²]) − (1.927 * Sex [male = 0; female = 1]) − (1.090 * ln NT-proBNP [pg/ml]) + (0.011 * 6MWD [m]) − (0.039 * SBP [mmHg]) − (0.512 * GLU [mmol/L]).

Standard error of estimate (SEE) = 2.236 ml·kg⁻¹·min⁻¹, R = 0.731, adjusted R² = 0.506.

As shown in Tables 3, 4, all variables in the multiple linear regression model were significantly associated with VO_2peak (P < 0.05), with variable influence ranked as: NT-proBNP > BMI > Sex > 6MWD > SBP > GLU.

The tolerance of the VO_2peak prediction model was between 0.807 and 0.973, and the VIF ranged from 1.028 to 1.239, confirming that the variables were independent of each other. The model demonstrated satisfactory goodness-of-fit, with adjusted R² of 0.506, SEE of 2.236 ml·kg⁻¹·min⁻¹, and significant F-statistic (F = 18.738, P < 0.001). The histogram of residuals (Figure 1A) combined with the Kolmogorov–Smirnov test (P = 0.20) supported that the residuals were normally distributed. Scatterplot analysis (Figure 1B) revealed that residual dispersion remained relatively constant across the prediction range, which confirmed homogeneity of variance. The Durbin-Watson statistic (1.862) indicated that the residuals of this regression equation were independent of each other.

Internal validation analyses (Table 5) demonstrated strong agreement between predicted and measured VO_2peak. Pearson correlation analysis was significant (r = 0.616, P < 0.01), with no systematic bias detected via paired Student's t-test (P > 0.05). The Bland-Altman analysis (Figure 1C) showed a mean bias of −0.13 (95% LoA: −5.20 to 4.93) between the two values, indicating that the predicted value was slightly lower than the measured value, but the difference was within acceptable limits. These results initially validate the robustness of the VO_2peak prediction model integrating NT-proBNP, BMI, Sex, 6MWD, SBP, and GLU.

The final regression equation was: METs_max (ml·kg⁻¹·min⁻¹) = 9.646 − (0.447 × Sex [male = 0; female = 1]) − (0.260 × ln NT-proBNP [pg/ml]) − (0.140 × GLU [mmol/L]) − (0.078 × BMI [kg/m²]) − (0.016 × SBP [mmHg]) + (0.004 × 6MWD [m]).

SEE = 0.719 ml·kg⁻¹·min⁻¹, R = 0.719, adjusted R² = 0.469.

Table 1 shows that all the variables included in the regression model were significantly associated with METs_max (P < 0.05) with variable influence ranked as: 6MWD > SBP > BMI > NT-proBNP > Sex > GLU.

Tables 3, 4 show the results of the multiple linear regression and the reliability test of the METs_max prediction model, respectively. The METs_max prediction model demonstrated acceptable multicollinearity metrics (tolerance > 0.1, VIF < 5), confirming variable independence. The regression exhibited satisfactory goodness-of-fit, with adjusted R² of 0.506, SEE of 2.236 ml·kg⁻¹·min⁻¹, and significant F-statistic (F = 16.3, P < 0.01). Residual diagnostics (Figures 2A,B) supported adherence to regression assumptions: normality was confirmed by Kolmogorov–Smirnov test (P = 0.2), homoscedasticity by the residual scatter-plot, and residual independence by the Durbin-Watson statistic (2.082).

Internal validation analyses (Table 5) revealed strong concordance between predicted and measured METs_max. Pearson correlation analysis was significant (r = 0.581, P < 0.01), with no systematic bias detected via paired Student's t-test (P > 0.05). The Bland-Altman analysis (Figure 2C) demonstrated excellent agreement between predicted and measured values, with a negligible mean bias of −0.00 (95% LoA: −1.59 to 1.59), confirming high concordance between the two values. These results initially validate the robustness of the METs_max prediction model incorporating 6MWD, SBP, BMI, NT-proBNP, Sex, and GLU.