Early differentiation between paroxysmal and persistent atrial fibrillation based on interpretable machine learning: a multicenter retrospective study

doi:10.21203/rs.3.rs-7302454/v1

Download PDF

Research Article

Early differentiation between paroxysmal and persistent atrial fibrillation based on interpretable machine learning: a multicenter retrospective study

https://doi.org/10.21203/rs.3.rs-7302454/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Aims

Atrial fibrillation (AF) is a common arrhythmia associated with increased risks of stroke and heart failure. Early differentiation between paroxysmal and persistent AF at first diagnosis is critical for guiding treatment decisions. This study aimed to develop an interpretable machine learning model based on structured electronic health records (EHR) to distinguish AF subtypes and identify key contributing factors.

Methods and results

In this multicenter, retrospective cohort study, data were collected from three tertiary hospitals in China between January 2013 and January 2023. A total of 11,986 patients with suspected AF were screened, of whom 4155 patients with first-diagnosed AF were included (paroxysmal: 2565 [61.3%]; persistent: 1620 [38.7%]). Structured EHR variables including clinical demographics, serological indicators, and echocardiographic parameters were extracted. Variable selection was performed using Spearman correlation and least absolute shrinkage and selection operator regression. Three machine learning algorithms were trained and externally validated. The CatBoost model achieved the best performance, with an area under the receiver operating characteristic curve of 0.876 (95% CI: 0.871–0.880) and accuracy of 0.808 (95% CI: 0.803–0.816). Sensitivity and specificity ranged from 0.802 to 0.811. Shapley additive explanations (SHAP) were used to interpret model outputs and identify variables most associated with AF subtype classification.

Conclusion

This multicenter study demonstrates that interpretable machine learning models based on structured EHR data can accurately distinguish paroxysmal from persistent AF at first diagnosis. The proposed model may facilitate early subtype-specific risk stratification and personalized treatment, potentially improve outcomes and reduce disparities in AF care across different medical conditions.

Paroxysmal atrial fibrillation

Persistent atrial fibrillation

Subtype classification

Machine learning

Multicenter retrospective study

Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia and is associated with an increased risk of stroke, death, heart failure, hospitalization, and cognitive decline^1–6. Early detection of AF offers a critical period for initiating appropriate therapies, modifying risk factors, and ensuring closer follow-up of high-risk individuals to prevent AF-related morbidity and mortality^7–9. Most AF screening is opportunistic, according to the latest European Society of Cardiology (ESC) and American Heart Association (AHA) guidelines^10–12. And screening also has shortcomings, such as easily overlooking patients with paroxysmal or short-term persistent AF^13,14. Current ESC/AHA guidelines emphasize that paroxysmal and persistent AF exhibit distinct treatment strategies and divergent long-term clinical prognoses^10,15,16. Patients with persistent AF are more likely to have adverse cardiovascular events, such as stroke and heart failure with preserved ejection fraction^11,17–21. Notably, paroxysmal and persistent AF account for approximately 38.9% and 39.2%, respectively, of all outpatient AF cases in China²². Therefore, early differentiation between paroxysmal and persistent AF has strong clinical application value. Currently, most researchers have paid attention to distinguishing AF subtypes by analyzing the AF attack phenomenon in ECG signals without evaluating from multiple angles, including previous medical history, serological tests, and cardiac ultrasound results^23,24. Although this has been noted by some researchers, it has been limited to explorations of small samples of people in single centers²⁵. Consequently, there is a lack of attention to the differentiation of paroxysmal and persistent AF.

Recent technological advances in artificial intelligence (AI) and machine learning (ML) provide opportunities to develop new tools to enable screening in risk factors and achieve more personalized treatment^26–28. In particular, ML models leveraging electronic health record (EHR) data have seen increasing application in cardiovascular medicine²⁹. As shown in Fig. 1A, in this study, we collected patient data from three tertiary hospitals in China and developed an interpretable ML model capable of accurately distinguishing paroxysmal AF from persistent AF of first-diagnosed AF patients. Additionally, we identified novel AF-related risk factors and implemented an online calculator to facilitate the broad application of our findings across Chinese and broader Asian populations. By enabling the timely and precise identification of AF subtypes, our approach aims to inform clinical decision-making, enhance preventive strategies, and pave the way for personalized care in AF management.

Datasets

This study was a multicenter retrospective study. The data were obtained from the EHR data of three tertiary medical institutions in different regions of China, including Sun Yat-sen Memorial Hospital of Sun Yat-sen University (SYSMH), Tungwah Hospital of Sun Yat-sen University (DH), and Dongguan Songshan Lake Tungwah Hospital (SSH).

Patient enrolment and data collection

Patients were enrolled based on the following criteria: (1) Hospitalization occurring between January 2013 and January 2023. (2) Documented evidence of AF rhythm either prior to or during hospitalization, confirmed by diagnostic modalities such as surface electrocardiography, 24-hour Holter monitoring, or pacemaker memory recordings. All patients had a first-time diagnosis of AF, though the exact AF subtype at the time of initial diagnosis was not specified. (3) AF was confirmed as the discharge diagnosis, classified as either paroxysmal or persistent AF according to the International Classification of Diseases, 10th Revision (ICD-10). Specifically, paroxysmal AF was coded as I48.x02 and persistent AF as I48.x00x007. Paroxysmal AF was defined as self-terminating episodes lasting less than 7 days, while persistent AF was defined as sustained episodes lasting longer than 7 days and requiring medical intervention^10,11. (4) Complete and detailed medical history records were available.

Patients were excluded if they met any of the following conditions: (1) History of rheumatic heart disease, congenital heart disease, primary valvular heart disease, cardiomyopathy, pericardial disease, cor pulmonale, malignancy, or recent major surgery. (2) Presence of systemic diseases potentially affecting cardiac structure or function, including acute myocardial infarction, thyroid dysfunction (hyperthyroidism or hypothyroidism), amyloidosis, pheochromocytoma, systemic lupus erythematosus, or severe infections. (3) Severe hepatic dysfunction, defined as alanine aminotransferase (ALT) levels exceeding three times the upper limit of normal and an ALT/AST ratio greater than 1. (4) Severe renal insufficiency, defined as an estimated glomerular filtration rate (eGFR) of less than 30 ml/min/1.73 m². (5) Incomplete clinical data, specifically defined as the absence of more than half of the required clinical variables or missing transthoracic echocardiography results.

All procedures in this study were performed in accordance with the ethical standards of the responsible committee for human experimentation in China and the Helsinki Declaration of 1975. This study was reviewed and approved by the Ethics Committee of SYSMH, DH, and SSH, respectively (SYSKY-2024-004-01, DHKY-2025-003-01, and SDHKY-2025-002-01).

Selection of variables

We collected demographic data, medication use, serological indicators, and baseline cardiac ultrasound data of all included subjects, totaling 50 variables. We performed Spearman correlation analysis on all the collected variables to initially screen out the variables that were correlated with the diagnosis of AF subtypes. Among them, P < 0.05 was considered statistically significant. Based on our previous studies, we added several variables that were considered to be correlated with the diagnosis of AF subtypes before further screening²⁵. Subsequently, we further screened the variables by Gradient Boosting Tree Recursive Feature Elimination (RFE) and Random Forest RFE. Through the above steps, we could comprehensively and accurately assess the importance of the variables and provide strong support for the subsequent model construction.

Machine Learning Algorithms

In this study, we conducted experiments using five widely adopted gradient boosting machine (GBM) algorithms: LightGBM (Light Gradient Boosting Machine), AdaBoost (Adaptive Boosting), GradientBoost, XGBoost (Extreme Gradient Boosting), and CatBoost (Categorical Boosting). Each algorithm offers unique advantages, enhancing the overall diversity and robustness of the analysis. LightGBM is a highly efficient gradient boosting framework. It is particularly known for its fast training speed and low memory usage, owing to its histogram-based method for decision tree learning. AdaBoost is one of the earliest boosting algorithms. The core mechanism of AdaBoost is to re-weight misclassified instances in each iteration, enabling the model to focus more on challenging cases, thus improving accuracy over time. GradientBoost is a classical gradient boosting algorithm that minimizes a loss function through the iterative training of decision trees. XGBoost is a highly efficient and scalable variant of gradient boosting, known for its exceptional performance in both speed and accuracy. CatBoost is an advanced gradient boosting algorithm that excels in handling classified features without the need for explicit preprocessing. Employing a range of machine learning models enabled us to comprehensively assess their performance on our datasets, providing insights into which models excel in addressing specific problems. We conducted experiments using five-fold cross-validation to comprehensively evaluate their performance on the dataset.

SHAP Interpretable Analysis for Machine Learning

SHAP (SHapley Additive exPlanations)³⁰ is a method rooted in game theory that allows for the quantification of each feature's contribution to a model's final prediction. While the tool itself provides transparency into model decisions, its real value lies in the insights it offers, helping us to better understand the inner workings of our machine learning models and the relationships between features and predictions. By calculating Shapley values, SHAP breaks down the model’s output into additive contributions from each feature, which not only aids in interpreting the model but also facilitates identifying which features most significantly impact the outcomes. The application of SHAP is particularly meaningful in our research, as it enables us to go beyond the "black box" nature of machine learning models and gain a deeper understanding of the driving forces behind their predictions.

Statistical Analysis

Normally distributed continuous variables were expressed as mean ± standard deviation (SD), and nonnormally distributed continuous variables were expressed as median (interquartile range, IQR). The distribution of continuous variables was assessed using the Shapiro–Wilk normality test, while the comparison of continuous variables was performed using Mann-Whitney U test. Classified variables were expressed as count and percentage. The x² test was used for comparisons of classified variables. Two‐tailed values of P < 0.05 were considered statistically significant. We used area under the receiver operating characteristic curve (AUC), sensitivity (SEN), specificity (SPE), accuracy (ACC), precision (PRE), recall, and F1 scores for evaluating the ability of the ML model to discriminate between paroxysmal and persistent AF. Displays restricted cubic spline (RCS) curves with 4 knots to test nonlinear relationships between independent variables and outcomes^31,32. SPSS Statistics Version 26.0 and Python 3.7.6 software were used for statistical analysis and graphics, and p < 0.05 was considered statistically significant.

Variable selection results

A detailed explanation of all the variables collected in this study can be found in Supplementary Table 1. After screening, we finally included 10 variables that can be divided into three categories: demographic data, cardiac ultrasound, and serological indicators. Demographic data included systolic blood pressure (SBP). Echocardiographic parameters included left atrial diameter (LA) and left ventricular ejection fraction (LVEF). These indicators were analyzed by routine transthoracic echocardiography (TTE) performed by a certified cardiologist at baseline and collected from the EHR. Serological parameters included white blood cell (WBC), neutrophils (NC), hemoglobin (Hb), N-terminal pro-brain natriuretic peptide (NT-proBNP), uric acid (UA), the ratio of low-density lipoprotein cholesterol to high-density lipoprotein cholesterol (LDL-C/HDL-C), and the atherosclerosis index of plasma (AIP). AIP is a logarithmically transformed ratio of TG (Total triglycerides) to HDL-C (High density lipoprotein cholesterol) in molar concentration (mmol/L), and it is mathematically derived from log (TG/HDL-C)³³. All serological indicators were obtained from the peripheral blood sample collected for the first time at baseline. The distribution of these 10 variables and their correlation with AF diagnostic subtypes in three independent centers are clearly shown in Fig. 2A1-A10 and B1-B3.

Baseline characteristics of participants

Initially, we collected 11,986 patients from DH, SYSMH, and SSH, including 6436, 4581, and 969 patients (Fig. 1B). According to the inclusion and exclusion criteria, we finally enrolled 4155 patients that including 2248 in DH, 1599 in SYSMH, and 338 in SSH. The number of patients with paroxysmal AF and persistent AF was 2565 (61.29%) and 1620 (38.71%), respectively. The proportions of patients with paroxysmal AF and persistent AF in the three centers were also close to 60% and 40% (1361 and 887 in DH, 1019 and 580 in SYSMH, 187 and 151 in SSH). The baseline characteristics of all variables finally included in the model are shown in Table 1.

Table 1
Baseline characteristics of participant
Variables	Total (n = 4185)	DH (n = 2248)	SYSMH (n = 1599)	SSH (n = 338)	p-Value
Age (years)	68 (59, 77)	69 (59, 78)	67 (60, 75)	69 (58, 78)	< 0.001*
Gender (n, %)					0.797
Male (%)	2468 (58.97)	1320 (58.71)	943 (58.97)	205 (60.65)
Female (%)	1717 (41.03)	928 (41.29)	656 (41.03)	133 (39.35)
SBP (mmHg)	131 (117, 147)	135 (118, 150)	127 (116, 142)	132 (118, 147)	< 0.001*
WBC (10⁹/L)	7.08 (5.73, 9.09)	7.43 (5.84, 9.74)	6.72 (5.67, 8.24)	7.17 (5.51, 9.19)	< 0.001*
NC (10⁹/L)	4.53 (3.36, 6.45)	4.86 (3.35, 7.23)	4.21 (3.35, 5.50)	5.07 (3.17, 7.34)	< 0.001*
Hb (g/L)	132 (118, 146)	131 (115, 145)	135 (123, 147)	130 (117, 145)	< 0.001*
AIP	0.032 (-0.156, 0.226)	0.014 (-0.180, 0.217)	0.055 (-0.128, 0.241)	0.009 (-0.173, 0.213)	< 0.001*
LDL-C/HDL-C	2.353 (1.733, 3.099)	2.256 (1.637, 3.045)	2.521 (1.944, 3.203)	2.055 (1.464, 2.792)	< 0.001*
NT-proBNP (pg/ml)	792 (209, 2337)	1144 (312, 3053)	470 (128, 1190)	1223 (324, 3469)	< 0.001*
UA (µmol/L)	384 (309, 468)	381 (301, 473)	391 (320, 468)	364 (289, 446)	< 0.001*
LA (mm)	37 (32, 43)	36 (31, 42)	38 (35, 43)	40 (35, 45)	< 0.001*
LVEF (%)	64 (57, 68)	62 (55, 67)	66 (61, 70)	59 (50, 65)	< 0.001*
Diagnose (n, %)					0.008*
0	2565 (61.29)	1361 (60.54)	1019 (63.73)	187 (55.33)
1	1620 (38.71)	887 (39.46)	580 (36.27)	151 (44.67)
Values are presented as n (%) as appropriate or the median [interquartile range (IQR)]. DH, Donghua Hospital of Sun Yat-sen University; SYSMH, Sun Yat-sen Memorial Hospital of Sun Yat-sen University; SSH, Dongguan Songshan Lake Donghua Hospital; SBP, systolic blood pressure; WBC, white blood cell; NC, Neutrophil count; Hb, hemoglobin; AIP, Atherogenic index of plasma; LDL-C, low density lipoprotein cholesterol; HDL-C, high density lipoprotein cholesterol; NT-proBNP, N-terminal brain natriuretic peptide precursor; UA, uric acid; LA, left atrial diameter; LVEF, left ventricular ejection fraction; Diagnose, 0 = paroxysmal AF, 1 = persistent AF.

Results of AF subtype prediction model

As described in the Methods section, we used five machine learning methods to build the model. We used the DH dataset with the largest sample size as the primary dataset and performed a five-fold cross-validation. As shown in Fig. 1C, in each cross-validation, the DH dataset was split into a training set and internal validation set in the ratio of 8:2, accompanied by independent external validation of the SYSMH and SSH datasets. This strategy helps eliminate the influence of data partitioning, ensuring more convincing and reliable results. The output variable of the model was the predicted AF diagnostic subtype, which was compared with the diagnosis recorded in the EHR at discharge. Comparing the evaluation indicators, we found that the model established in our study had good predictive performance and stable generalizability (as shown in Table 2).

Table 2
Results of model output indicators in multicenter
Model	ACC	Precision	Recall	AUC	F1 Score	SEN	SPE
DH
CatBoost	0.789 (0.766–0.825)	0.755 (0.700-0.812)	0.694 (0.656–0.727)	0.861 (0.835–0.898)	0.722 (0.701–0.764)	0.694 (0.656–0.727)	0.852 (0.798–0.892)
GradientBoost	0.791 (0.770–0.823)	0.753 (0.727–0.803)	0.702 (0.648–0.736)	0.859 (0.838–0.894)	0.726 (0.698–0.765)	0.702 (0.648–0.736)	0.849 (0.821–0.885)
LightGBM	0.780 (0.766–0.815)	0.738 (0.702–0.793)	0.688 (0.645–0.721)	0.855 (0.835–0.889)	0.711 (0.689–0.754)	0.688 (0.645–0.721)	0.840 (0.809–0.879)
XGBoost	0.774 (0.756–0.896)	0.727 (0.679–0.754)	0.688 (0.644–0.720)	0.846 (0.827–0.875)	0.706 (0.683–0.733)	0.688 (0.644–0.720)	0.831 (0.797–0.852)
AdaBoost	0.783 (0.768–0.803)	0.740 (0.716–0.769)	0.697 (0.661–0.726)	0.845 (0.826–0.875)	0.717 (0.695–0.739)	0.697 (0.661–0.726)	0.840 (0.811–0.863)
SSH
CatBoost	0.748 (0.735–0.762)	0.681 (0.669–0.690)	0.834 (0.817–0.867)	0.833 (0.828–0.840)	0.750 (0.736–0.767)	0.834 (0.817–0.867)	0.677 (0.665–0.696)
GradientBoost	0.734 (0.725–0.744)	0.670 (0.663–0.677)	0.813 (0.792–0.837)	0.825 (0.820–0.831)	0.734 (0.725–0.748)	0.813 (0.792–0.837)	0.668 (0.650–0.685)
LightGBM	0.737 (0.731–0.755)	0.674 (0.662–0.690)	0.812 (0.792–0.836)	0.821 (0.814–0.830)	0.737 (0.728–0.755)	0.812 (0.792–0.836)	0.676 (0.651–0.691)
XGBoost	0.728 (0.708–0.746)	0.666 (0.649–0.681)	0.804 (0.773–0.829)	0.821 (0.806–0.830)	0.728 (0.706–0.747)	0.804 (0.773–0.829)	0.666 (0.644–0.681)
AdaBoost	0.741 (0.725–0.764)	0.677 (0.657–0.699)	0.817 (0.786–0.842)	0.819 (0.808–0.830)	0.740 (0.722–0.764)	0.817 (0.786–0.842)	0.678 (0.642–0.701)
SYSMH
CatBoost	0.808 (0.803–0.816)	0.707 (0.698–0.719)	0.802 (0.788–0.812)	0.876 (0.871–0.880)	0.752 (0.747–0.761)	0.802 (0.788–0.812)	0.811 (0.802–0.820)
GradientBoost	0.802 (0.794–0.810)	0.702 (0.688–0.708)	0.790 (0.769–0.809)	0.872 (0.869–0.873)	0.743 (0.733–0.755)	0.790 (0.769–0.809)	0.809 (0.795–0.814)
LightGBM	0.807 (0.801–0.811)	0.706 (0.695–0.721)	0.801 (0.779–0.820)	0.875 (0.873–0.877)	0.750 (0.744–0.755)	0.801 (0.779–0.820)	0.810 (0.797–0.828)
XGBoost	0.800 (0.795–0.807)	0.697 (0.685–0.704)	0.793 (0.772–0.810)	0.873 (0.869–0.879)	0.742 (0.734–0.751)	0.793 (0.772–0.810)	0.804 (0.789–0.811)
AdaBoost	0.787 (0.775–0.795)	0.678 (0.666–0.686)	0.787 (0.762–0.803)	0.860 (0.853–0.866)	0.728 (0.711–0.740)	0.787 (0.762–0.803)	0.787 (0.782–0.791)
DH, Donghua Hospital of Sun Yat-sen University; SYSMH, Sun Yat-sen Memorial Hospital of Sun Yat-sen University; SSH, Dongguan Songshan Lake Donghua Hospital; ACC, accuracy; AUC, area under curve; CI, confidence interval; SEN, sensitivity; SPE, specificity.

Among the overall models built based on the DH dataset, the GradientBoost model had the highest accuracy of 0.791 (95%CI: 0.770–0.823), followed by the CatBoost model, which was 0.789 (95%CI: 0.766–0.825). In terms of AUC value comparison, the value of the GradientBoost model was 0.859 (95% CI: 0.838–0.894), which was slightly lower than the maximum value of 0.861 (95% CI: 0.835–0.898) of the CatBoost model. In terms of sensitivity and specificity, the GradientBoost model (0.702, 95%CI: 0.648–0.736) and the CatBoost model (0.852, 95%CI: 0.798–0.892) performed best, respectively. As for the two independent validation sets SSH and SYSMH, except that the highest specificity of SSH was achieved by the AdaBoost model (0.678, 95%CI: 0.642–0.701), the CatBoost model had the highest accuracy of 0.808 (95%CI: 0.803–0.816), AUC value of 0.876 (95%CI: 0.871–0.880), sensitivity of 0.802 (95%CI: 0.788–0.812) and specificity of 0.811 (95%CI: 0.802–0.820). We summarized the AUC for every center and the results of each fold of the five-fold cross validation for the five algorithms, as shown in Fig. 2C1-C3. We also plotted the AUC results of each fold in the five-fold cross-validation process for all centers in Supplementary Fig. 1.

Interpretation of AF subtype prediction model

As shown in Fig. 3A, the impact of different variables on the model output is illustrated by ranking their absolute SHAP values in descending order. The five variables that most significantly affect the diagnosis of different AF types are LA, NT-proBNP, Hb, LVEF, and UA. Figure 3B provides a more intuitive view of the relationship between these variables and AF diagnostic types. It can be observed that LA, LVEF, and UA affect the diagnosis of AF in a certain pattern. For example, LA displays a gradient transitioning from blue to red. There is a distinct color boundary near a SHAP value of 0, indicating a regular pattern between LA values and AF diagnostic types. Specifically, when LA values are lower, the model tends to predict paroxysmal AF, whereas higher LA values are associated with a diagnosis of persistent AF. Figure 3C shows the impact of all variables on sample classification across all samples, with red indicating a positive effect (persistent AF) on the model’s prediction and blue indicating a negative effect (persistent AF). Figure 3D further illustrates the influence of each variable on the model’s prediction for a specific sample using the SHAP method. Figure 3E shows how the top five variables most influential for distinguishing AF subtypes affect the model’s output as each variable changes. Compared to Fig. 2B, this view makes the impact of each variable’s trend even clearer. For example, as LA increases, its contribution shifts the model’s prediction toward persistent AF.

In the model for early differentiation of AF subtypes as paroxysmal or persistent, the top five variables were LA, NT-proBNP, LVEF, UA, and SBP. For these variables, we further obtained the mutual influence relationship between them and plotted them into a scatter plot (see Supplementary Fig. 2 for details). As shown in Supplementary Fig. 3, we also explored the relationship between these five variables and AF diagnostic subtypes in three independent centers by restricted cubic spline analysis.

Performance and explanation of AF subtype prediction models in subgroups

We divide the participants of every center into six groups: male under 60 years old, female under 60 years old, male 60–65 years old, female 60–65 years old, male 65 years old and above, female 65 years old and above. The model has achieved good prediction performance among different subgroups. We sorted the results of AUC between different centers according to male or female in different age subgroups to draw Supplementary Fig. 4. The order of A-F in Supplementary Fig. 4 is arranged according to age, and centers 1–3 are DH, SSH, and SYSMH, respectively. We show these results and the specific values of other evaluation indicators in Supplementary Tables 2–4. Similarly, we used the SHAP method to visualize the interpretability of the model. The SHAP graphs of all subgroups are summarized in Supplementary Fig. 5 of the Supplementary Materials. The most important and second influencing factors of any age subgroups are LA and NT-proBNP, which are similar to the overall model.

This study integrated EHR data (clinical history, serum indicators, and cardiac ultrasound) from three independent large tertiary hospitals in China and established an interpretable model based on machine learning to accurately distinguish paroxysmal and persistent AF in advance. We screened out new AF subtype-related factors, which not only provided a basis for further mechanism research, but also helped us build an online web calculator (http://123.56.120.106:8000/) that can accurately distinguish whether patients with first-time AF will have paroxysmal or persistent AF in the long term and give the probability. In addition, the variables affecting diagnosis in different age and gender subgroups were evaluated to confirm the importance and specificity of these variables in distinguishing early AF subtypes between paroxysmal and persistent AF.

AF is a common cause of stroke, heart failure, cardiovascular death, and dementia^8,34–36. The global burden of AF is rising³⁷, and Asia-Pacific patients account for the majority of AF patients worldwide, while Chinese patients are the fifth in estimated prevalence and the first in absolute prevalence in the Asia-Pacific region³⁸. China and the Asia-Pacific region face challenges including great disparities in access to healthcare and availability of diagnostic technology^38,39. The model and online calculator developed based on this study can help narrow the disparity in healthcare access between different regions and improve the long-term prognosis of AF patients caused by insufficient availability of diagnostic technology.

This study found 10 variables that can accurately distinguish paroxysmal from persistent AF in advance, which is partially similar to our previous research results²⁵. Among the key variables affecting the model obtained based on the SHAP method, LA is the most important. Previous studies have found that LA is one of the most important indicators for predicting new-onset AF and early distinguishing AF subtypes^21,40,41, and its increase is associated with the progression of AF from paroxysmal to persistent⁴². Our research has verified this view in the whole population and different subgroups. The latest studies suggest that the possible pathogenic mechanisms of this phenomenon are atrial cardiomyopathy and atrial fibrosis^43,44, and left atrial diameter is the clinical manifestation of these potential mechanisms.

Another crucial variable is UA, and our results suggest that higher UA levels are more likely to be diagnosed with persistent AF. This conclusion is consistent with the view of a recent published review⁴⁵. There is a significant difference in UA levels between paroxysmal AF and persistent AF⁴⁶, and this dose-response relationship can be observed in people with different disease backgrounds⁴⁷. The potential mechanism may be that high UA may not only increase the risk of AF through cardiovascular disease, but also directly affect the development of AF through mechanisms such as oxidative stress, inflammation, insulin resistance, and activation of the renin-angiotensin-aldosterone system, ultimately leading to electrical remodeling, changes in the autonomic nervous system, abnormal Ca²⁺ handling, and atrial remodeling^48–50. The incidence of patients with high UA is increasing year by year worldwide⁵¹, and it affects about 14.0% of adults in China or even more⁵². Therefore, the relationship between UA and AF subtype should be the focus of further research.

Our study also further confirmed the importance of SBP in differentiating AF subtypes, which is consistent with the conclusion of a recent study based on the China Cardiovascular Care Quality Improvement-Atrial Fibrillation⁵³. Higher blood pressure levels on admission to hospital for AF patients were associated with an increased risk of stroke/transient ischemic attack and heart failure (HF), while lower blood pressure levels were associated with an increased risk of HF and all-cause mortality. Patients with lower SBP are more likely to have persistent AF, which is accompanied by a greater risk of HF. Further research is needed to clarify the relationship between blood pressure and AF type and prognosis.

Studies have shown that there are clear gender differences in the epidemiology and risk factors of AF^54–56, and the relationship between increasing age and the progression of AF has been confirmed⁴². We conducted the subgroup analysis combining gender and age. The cut-off points were derived from the newly released Chinese AF Diagnosis and Management Guidelines based on Asian research evidence^57–60. We confirmed the stability of the model by obtaining similar model performance in different subgroups. Inter-group comparisons also yielded some meaningful conclusions, such as the highest AUC in the subgroup < 60 in male, and the highest AUC in the subgroup 60–64 in female. Our study achieved accurate differentiation of AF subtypes in younger age groups. A recent study concluded that younger age at first diagnosis of AF is associated with a higher risk of stroke⁶¹, so this study is conducive to early anticoagulation treatment for younger age subgroups based on this conclusion.

Nevertheless, our study also has some limitations. We only explored the distinction between two AF subtypes, while the latest ACC guidelines proposed four stages of AF evolution⁶². A larger sample size of AF patients with more subtypes and long-term follow-up are needed to achieve accurate multi-classification and progress prediction of AF subtypes²⁹. Although the patients came from different provinces, the three centers included in this study were all in China, which means that more centers in other districts and countries need to be added in the future to make the sample characteristics more extensive. While increasing research centers and participants, the balance of sample accounts between different centers is also something that needs to be balanced in our further researches.

In this study, we collected data from 11,986 patients across three tertiary hospitals in China and developed an interpretable machine learning model capable of accurately distinguishing early paroxysmal AF from persistent AF subtypes of first-diagnosed AF. Additionally, we identified novel AF-related risk factors and implemented an online calculator to facilitate the broad application of our findings throughout Chinese and wider Asian populations. By enabling timely and precise identification of AF subtypes, our approach aims to inform clinical decision-making, strengthen preventive strategies, and pave the way for personalized care in AF management.

ACC

accuracy

atrial fibrillation

AIP

atherogenic index of plasma

AUC

area under the curve

aortic valve flow velocity

BNP

brain natriuretic peptide

confidence interval

DNP

diastolic blood pressure

EHR

electronic health records

ESC

European Society of Cardiology

hemoglobin

HDL-C

high-density lipoprotein cholesterol

heart failure

ICD

International Classification of Diseases

IQR

interquartile range

left atrial diameter

LASSO

least absolute shrinkage and selection operator

LDL-C

low-density lipoprotein cholesterol

LVDd

left ventricular end-diastolic diameter

LVEF

left ventricular ejection fraction

machine learning

NT-proBNP

N-terminal pro-brain natriuretic peptide

RCS

restricted cubic spline

RFE

Recursive Feature Elimination

ROC

Receiver operating characteristic

SBP

systolic blood pressure

standard deviation

SEN

sensitivity

SHAP

SHapley Additive exPlanations

SPE

specificity

TTE

transthoracic echocardiography

uric acid.

Conflict of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Consent for publication

Not applicable.

Ethics approval and consent to participate

This study was reviewed and approved by the Ethics Committees of the Sun Yat-sen Memorial Hospital of Sun Yat-sen University (SYSKY-2024-004-01), the Ethics Committees of Tungwah Hospital of Sun Yat-sen University (DHKY-2025-003-01), and Dongguan Songshan Lake Tungwah Hospital (SDHKY-2025-002-01). The study adhered to the tenets of the Declaration of Helsinki.

Funding

This study is partially supported by National Natural Science Foundation of China (62176016, 72274127), National Key R\&D Program of China (No. 2021YFB2104800), Guizhou Province Science and Technology Project: Research on Q&A Interactive Virtual Digital People for Intelligent Medical Treatment in Information Innovation Environment (supported by Qiankehe[2024] General 058), Capital Health Development Research Project(2022-2-2013), Haidian innovation and translation program from Peking University Third Hospital (HDCXZHKC2023203), and Project: Research on the Decision Support System for Urban and Park Carbon Emissions Empowered by Digital Technology - A Special Study on the Monitoring and Identification of Heavy Truck Beidou Carbon Emission Reductions to Chao Tong. The Grant from Key Laboratory of Coronary Intraluminal Imaging and Functional Analysis of Dongguan City to Heng Li. And Shenzhen Medical Research Fund (B2302020), National Natural Science Foundation of China (82330021, 82270771), Shenzhen Science and Technology Program (KCXFZ20211020163801002, ZDSYS20220606100801004, SGDX20230116092459009), and Shenzhen Key Medical Discipline Construction Fund (SZXK002), Futian District Public Health Scientific Research Project of Shenzhen (FTWS2022001), Chinese Association of Integrative Medicine-Shanghai Hutchison Pharmaceuticals Fund (HMPE202202), China Heart House-Chinese Cardiovascular Association HX fund (2022-CCA-HX-090) to Hui Huang.

Author Contribution

Chao Tong, Heng Li and Hui Huang are the guarantors of the study. All authors (Sijin Li, Yuqi Zhang, Weijie Wu, Guang Li, Tucheng Huang, Chao Tong, Heng Li and Hui Huang) were involved in the conceptualization and design of the study. Sijin Li and Yuqi Zhang were responsible for the experiment. Data cleaning was done by Weijie Wu Guang Li and Tuchang Huang. Analysis and interpretation were done by Sijin Li and Yuqi Zhang under the supervision and withthe support of Chao Tong, Heng Li and Hui Huang. Drafting of the article was done by Sijin Li, Yuqi Zhang. All authors (Sijin Li, Yuqi Zhang, Weijie Wu, Guang Li, Tucheng Huang, Chao Tong, Heng Li and Hui Huang) revised and contributed to the intellectual content of the article. All authors (Sijin Li, Yuqi Zhang, Weijie Wu, Guang Li, Tucheng Huang, Chao Tong, Heng Li and Hui Huang) approved the final version of the article, including the authorship list.

Data availability

Some or all data sets generated during and/or analyzed during the present study are not publicly available but are available from the corresponding author on reasonable request.

Tsao CW, Aday AW, Almarzooq ZI, et al. Heart Disease and Stroke Statistics-2023 Update: A Report From the American Heart Association. Circulation. 2023;147(8):e93–621.
Petzl AM, Jabbour G, Cadrin-Tourigny J et al. Innovative approaches to atrial fibrillation prediction: should polygenic scores and machine learning be implemented in clinical practice? Europace 2024; 26(8).
Papanastasiou CA, Theochari CA, Zareifopoulos N, et al. Atrial Fibrillation Is Associated with Cognitive Impairment, All-Cause Dementia, Vascular Dementia, and Alzheimer's Disease: a Systematic Review and Meta-Analysis. J Gen Intern Med. 2021;36(10):3122–35.
Koh YH, Lew LZW, Franke KB, et al. Predictive role of atrial fibrillation in cognitive decline: a systematic review and meta-analysis of 2.8 million individuals. Europace. 2022;24(8):1229–39.
Qin D, Mansour MC, Ruskin JN, Heist EK. Atrial Fibrillation-Mediated Cardiomyopathy. Circ Arrhythm Electrophysiol. 2019;12(12):e007809.
Rienstra M, Tzeis S, Bunting KV et al. Spotlight on the 2024 ESC/EACTS management of atrial fibrillation guidelines: 10 novel key aspects. Europace 2024; 26(12).
Mairesse GH, Moran P, Van Gelder IC, et al. Screening for atrial fibrillation: a European Heart Rhythm Association (EHRA) consensus document endorsed by the Heart Rhythm Society (HRS), Asia Pacific Heart Rhythm Society (APHRS), and Sociedad Latinoamericana de Estimulacion Cardiaca y Electrofisiologia (SOLAECE). Europace. 2017;19(10):1589–623.
Rivard L, Friberg L, Conen D, et al. Atrial Fibrillation and Dementia: A Report From the AF-SCREEN International Collaboration. Circulation. 2022;145(5):392–409.
Kirchhof P, Camm AJ, Goette A, et al. Early Rhythm-Control Therapy in Patients with Atrial Fibrillation. N Engl J Med. 2020;383(14):1305–16.
Van Gelder IC, Rienstra M, Bunting KV, et al. 2024 ESC Guidelines for the management of atrial fibrillation developed in collaboration with the European Association for Cardio-Thoracic Surgery (EACTS). Eur Heart J. 2024;45(36):3314–414.
Joglar JA, Chung MK, Armbruster AL, et al. 2023 ACC/AHA/ACCP/HRS Guideline for the Diagnosis and Management of Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation. 2024;149(1):e1–156.
Writing Committee M, Joglar JA, Chung MK, et al. 2023 ACC/AHA/ACCP/HRS Guideline for the Diagnosis and Management of Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. J Am Coll Cardiol. 2024;83(1):109–279.
Lubitz SA, Atlas SJ, Ashburner JM, et al. Screening for Atrial Fibrillation in Older Adults at Primary Care Visits: VITAL-AF Randomized Controlled Trial. Circulation. 2022;145(13):946–54.
Uittenbogaart SB, Verbiest-van Gurp N, Lucassen WAM, et al. Opportunistic screening versus usual care for detection of atrial fibrillation in primary care: cluster randomised controlled trial. BMJ. 2020;370:m3208.
Tzeis S, Gerstenfeld EP, Kalman J et al. 2024 European Heart Rhythm Association/Heart Rhythm Society/Asia Pacific Heart Rhythm Society/Latin American Heart Rhythm Society expert consensus statement on catheter and surgical ablation of atrial fibrillation. Europace 2024; 26(4).
Natale A, Mohanty S, Sanders P, et al. Catheter ablation for atrial fibrillation: indications and future perspective. Eur Heart J. 2024;45(41):4383–98.
Nattel S, Harada M. Atrial remodeling and atrial fibrillation: recent advances and translational perspectives. J Am Coll Cardiol. 2014;63(22):2335–45.
Park JS, Cho I, Kim D, et al. Differentiating Left Atrial Pressure Responses in Paroxysmal and Persistent Atrial Fibrillation: Implications for Diagnosing Heart Failure With Preserved Ejection Fraction and Managing Atrial Fibrillation. J Am Heart Assoc. 2024;13(17):e035246.
Vanassche T, Lauw MN, Eikelboom JW, et al. Risk of ischaemic stroke according to pattern of atrial fibrillation: analysis of 6563 aspirin-treated patients in ACTIVE-A and AVERROES. Eur Heart J. 2015;36(5):281–a7.
Chiang CE, Naditch-Brule L, Murin J, et al. Distribution and risk profile of paroxysmal, persistent, and permanent atrial fibrillation in routine clinical practice: insight from the real-life global survey evaluating patients with atrial fibrillation international registry. Circ Arrhythm Electrophysiol. 2012;5(4):632–9.
Ogawa H, An Y, Ikeda S, et al. Progression From Paroxysmal to Sustained Atrial Fibrillation Is Associated With Increased Adverse Events. Stroke. 2018;49(10):2301–8.
Shi S, Tang Y, Zhao Q, et al. Prevalence and risk of atrial fibrillation in China: A national cross-sectional epidemiological study. Lancet Reg Health West Pac. 2022;23:100439.
Gavidia M, Zhu H, Montanari AN, et al. Early warning of atrial fibrillation using deep learning. Patterns (N Y). 2024;5(6):100970.
Petmezas G, Haris K, Stefanopoulos L, et al. Automated atrial fibrillation detection using a hybrid CNN-LSTM network on imbalanced ECG datasets. Biomed Signal Process Control. 2021;63:1021941746–8094.
Zhang Y, Li S, Mai P, et al. A machine learning-based model for predicting paroxysmal and persistent atrial fibrillation based on EHR. BMC Med Inf Decis Mak. 2025;25(1):51.
Daneshvar N, Pandita D, Erickson S, et al. Artificial Intelligence in the Provision of Health Care: An American College of Physicians Policy Position Paper. Ann Intern Med. 2024;177(7):964–7.
Rose C, Chen JH. Learning from the EHR to implement AI in healthcare. NPJ Digit Med. 2024;7(1):330.
Zhang Y, Li S, Wu W, et al. Machine-learning-based models to predict cardiovascular risk using oculomics and clinic variables in KNHANES. BioData Min. 2024;17(1):12.
Quer G, Topol EJ. The potential for large language models to transform cardiovascular medicine. Lancet Digit Health. 2024;6(10):e767–71.
Lundberg SM, Erion G, Chen H, et al. From local explanations to global understanding with explainable AI for trees. Nat Mach Intell. 2020;2(1):56–672522.
Gupta S, Glezerman IG, Hirsch JS, et al. Derivation and external validation of a simple risk score for predicting severe acute kidney injury after intravenous cisplatin: cohort study. BMJ. 2024;384:e077169.
Yang J, Wang T, Li K, Wang Y. Associations between per- and polyfluoroalkyl chemicals and abdominal aortic calcification in middle-aged and older adults. J Adv Res 2024.
Assempoor R, Daneshvar MS, Taghvaei A, et al. Atherogenic index of plasma and coronary artery disease: a systematic review and meta-analysis of observational studies. Cardiovasc Diabetol. 2025;24(1):35.
Hindricks G, Potpara T, Dagres N, et al. 2020 ESC Guidelines for the diagnosis and management of atrial fibrillation developed in collaboration with the European Association for Cardio-Thoracic Surgery (EACTS): The Task Force for the diagnosis and management of atrial fibrillation of the European Society of Cardiology (ESC) Developed with the special contribution of the European Heart Rhythm Association (EHRA) of the ESC. Eur Heart J. 2021;42(5):373–498.
Fabritz L, Crijns H, Guasch E et al. Dynamic risk assessment to improve quality of care in patients with atrial fibrillation: the 7th AFNET/EHRA Consensus Conference. Europace 2021; 23(3): 329 – 44.
Kim D, Yang PS, Sung JH, et al. Less dementia after catheter ablation for atrial fibrillation: a nationwide cohort study. Eur Heart J. 2020;41(47):4483–93.
Tan S, Zhou J, Veang T, Lin Q, Liu Q. Global, regional, and national burden of atrial fibrillation and atrial flutter from 1990 to 2021: sex differences and global burden projections to 2046-a systematic analysis of the Global Burden of Disease Study 2021. Europace 2025; 27(2).
Wong CX, Tse HF, Choi EK, et al. The burden of atrial fibrillation in the Asia-Pacific region. Nat Rev Cardiol. 2024;21(12):841–3.
Liu H, Chen M. Atrial Fibrillation Screening in Asia: Balancing Costs and Benefits for Optimal Outcomes. JACC Asia. 2025;5(1):172–3.
Lu R, Lumish HS, Hasegawa K, et al. Prediction of new-onset atrial fibrillation in patients with hypertrophic cardiomyopathy using machine learning. Eur J Heart Fail. 2025;27(2):275–84.
Jabbour G, Nolin-Lapalme A, Tastet O, et al. Prediction of incident atrial fibrillation using deep learning, clinical models, and polygenic scores. Eur Heart J. 2024;45(46):4920–34.
Padfield GJ, Steinberg C, Swampillai J, et al. Progression of paroxysmal to persistent atrial fibrillation: 10-year follow-up in the Canadian Registry of Atrial Fibrillation. Heart Rhythm. 2017;14(6):801–7.
Choi SH, Jurgens SJ, Xiao L, et al. Sequencing in over 50,000 cases identifies coding and structural variation underlying atrial fibrillation risk. Nat Genet. 2025;57(3):548–62.
Schotten U, Goette A, Verheule S. Translation of pathophysiological mechanisms of atrial fibrosis into new diagnostic and therapeutic approaches. Nat Rev Cardiol. 2025;22(4):225–40.
Lu Y, Sun Y, Cai L, et al. Non-traditional risk factors for atrial fibrillation: epidemiology, mechanisms, and strategies. Eur Heart J. 2025;46(9):784–804.
Wang X, Hou Y, Wang X, et al. Relationship between serum uric acid levels and different types of atrial fibrillation: An updated meta-analysis. Nutr Metab Cardiovasc Dis. 2021;31(10):2756–65.
Ding M, Viet NN, Gigante B, Lind V, Hammar N, Modig K. Elevated Uric Acid Is Associated With New-Onset Atrial Fibrillation: Results From the Swedish AMORIS Cohort. J Am Heart Assoc. 2023;12(3):e027089.
Masi S, Pugliese NR, Taddei S. The difficult relationship between uric acid and cardiovascular disease. Eur Heart J. 2019;40(36):3055–7.
Liu CH, Huang SC, Yin CH et al. Atrial Fibrillation Risk and Urate-Lowering Therapy in Patients with Gout: A Cohort Study Using a Clinical Database. Biomedicines 2022; 11(1).
Yu W, Cheng JD. Uric Acid and Cardiovascular Disease: An Update From Molecular Mechanism to Clinical Perspective. Front Pharmacol. 2020;11:582680.
Dehlin M, Jacobsson L, Roddy E. Global epidemiology of gout: prevalence, incidence, treatment patterns and risk factors. Nat Rev Rheumatol. 2020;16(7):380–90.
Du L, Zong Y, Li H, et al. Hyperuricemia and its related diseases: mechanisms and advances in therapy. Signal Transduct Target Ther. 2024;9(1):212.
Sun Z, Hao Y, Liu J et al. Blood pressure and in-hospital outcomes in patients hospitalized with atrial fibrillation: findings from the CCC-AF project. Hypertens Res 2025.
Kjerpeseth LJ, Igland J, Selmer R, et al. Prevalence and incidence rates of atrial fibrillation in Norway 2004–2014. Heart. 2021;107(3):201–7.
Al-Khayatt BM, Salciccioli JD, Marshall DC, Krahn AD, Shalhoub J, Sikkel MB. Paradoxical impact of socioeconomic factors on outcome of atrial fibrillation in Europe: trends in incidence and mortality from atrial fibrillation. Eur Heart J. 2021;42(8):847–57.
Sharashova E, Gerdts E, Ball J et al. Long-term pulse pressure trajectories and risk of incident atrial fibrillation: the Tromso Study. Eur Heart J 2025.
Ma C, Wu S, Liu S, Han Y. Chinese guidelines for the diagnosis and management of atrial fibrillation. Pacing Clin Electrophysiol. 2024;47(6):714–70.
Kim TH, Yang PS, Yu HT, et al. Age Threshold for Ischemic Stroke Risk in Atrial Fibrillation. Stroke. 2018;49(8):1872–9.
Li YG, Lee SR, Choi EK, Lip GY. Stroke Prevention in Atrial Fibrillation: Focus on Asian Patients. Korean Circ J. 2018;48(8):665–84.
Choi SY, Kim MH, Lee KM, et al. Age-Dependent Anticoagulant Therapy for Atrial Fibrillation Patients with Intermediate Risk of Ischemic Stroke: A Nationwide Population-Based Study. Thromb Haemost. 2021;121(9):1151–60.
Cheng YJ, Deng H, Wei HQ, et al. Association Between Age at Diagnosis of Atrial Fibrillation and Subsequent Risk of Ischemic Stroke. J Am Heart Assoc. 2025;14(4):e038367.
Ko D, Chung MK, Evans PT, Benjamin EJ, Helm RH. Atrial Fibrillation: A Review. JAMA. 2025;333(4):329–42.

No competing interests reported.

Supplementary.docx

Download PDF

Reviewers agreed at journal
09 Oct, 2025
Reviewers invited by journal
02 Sep, 2025
Editor assigned by journal
08 Aug, 2025
Submission checks completed at journal
08 Aug, 2025
First submitted to journal
05 Aug, 2025

You are reading this latest preprint version

Early differentiation between paroxysmal and persistent atrial fibrillation based on interpretable machine learning: a multicenter retrospective study

Status:

Version 1

Abstract

Figures

Introduction

Methods

Statistical Analysis

Results

Discussion

Conclusion

Abbreviations

Declarations

Conflict of Interest

Ethics approval and consent to participate

Funding

Author Contribution

Data availability

References

Additional Declarations

Supplementary Files

Status:

Version 1