The study included 4896 patients with AF in the MIMIC-IV database, with 2871(58.64%) of the cohort being male. Of the total number of patients, 1275 (26.04%) died during their hospital stay, while the remaining 3621 survived. During follow-up, 1432 patients (29.25%) died within 28 days, and 2436 patients (49.75%) died within 1 year. A statistically significant difference was observed in age between hospital survivors and non-survivors (p < 0.001). The systolic blood pressure (p < 0.001), diastolic blood pressure (p < 0.001)and SpO2 (p < 0.024) were both significantly lower in non-survivors, while the heart rate (p < 0.001) and respiratory rate (p < 0.001) were significantly higher in survivors. With regard to the results of laboratory tests, with the exception of sodium (p = 0.482), which demonstrated no significant differences, the majority of indicators exhibited marked disparities between the two groups. It is noteworthy that lactate dehydrogenase (p < 0.001) and serum creatinine (p < 0.001) levels were higher in non-survivors, whereas platelet levels (p < 0.001) exhibited the opposite trend. This resulted in a significantly higher EASIX in non-survivors (p < 0.001). In addition, non-survivors exhibited more severe conditions, as evidenced by higher scores on the SOFA, OASIS, APS III, SAPS II, and Charlson index, along with a lower GCS (p < 0.001). However, there was no significant difference in CHA2DS2–VASc scores between the two groups (p = 0.529). In addition, non-survivors had a significantly higher prevalence of sepsis (p < 0.001) and acute kidney injury (AKI) (p < 0.001) compared to survivors, and experienced a shorter hospitalization duration (p < 0.001) but a prolonged ICU stay (p < 0.001) (Table 1).

As indicated by the Boruta algorithm, 49 of the 71 variables most strongly associated with in-hospital mortality were confirmed (Fig. 2, Supplementary Table S2). In addition, using Lasso regression, we identified 57 highly relevant variables while optimizing the lambda to minimize multicollinearity (Fig. 3, Supplementary Table S2). The intersection of the Boruta-selected variables and those selected by Lasso resulted in 38 variables being retained for further analysis, which were considered to have a significant impact on in-hospital mortality in patients with atrial fibrillation. Notably, EASIX was retained in both selection methods, highlighting its relevance across both approaches. Considering both clinical significance and the need to mitigate multicollinearity, we retained not only the intersection variables but also included additional factors such as gender, race, CKD, COPD and the CHA2DS2–VASc score as correction factors. Ultimately, 45 variables were incorporated into the fully adjusted model.

Boruta algorithm conducted the feature selection for the relationship between EASIX and in-hospital mortality. The horizontal axis shows the name of each variable, while the vertical axis represents the Z value of each variable. The box plot depicts the Z value of each variable in the model calculation, with green boxes representing important variables, yellow representing tentative attributes, and red representing unimportant variables

Lasso regression conducted the feature selection for the relationship between EASIX and in-hospital mortality. A Variation characteristics of the coefficients of variables as the regularization parameter λ changes. The plot shows how the coefficients shrink towards zero with increasing λ, highlighting the importance of each variable; B selection process of the optimal value of the regularization parameter λ in the Lasso regression model, determined through cross-validation. The plot illustrates the relationship between the mean cross-validation error and log (λ). The dashed vertical lines indicate two key values of λ: the value that minimizes the mean cross-validation error (λ_min) and the largest value of λ within one standard error of the minimum (λ_1se), used for model selection

To comprehensively assess the role of EASIX, we analyzed it both as a continuous variable and categorized into quartiles. Participants were grouped by the EASIX quartiles at admission (Q1: < 4.56, Q2: 4.56–5.64, Q3: 5.64–6.84, and Q4: > 6.84) and their baseline characteristics are summarized in Supplementary Table S3.

The results from the multivariable logistic regression analysis (Table 2) indicated that a higher EASIX was significantly associated with an increased risk of in-hospital mortality (OR 1.28, 95% confidence interval [CI] 1.19–1.37), after adjusting for all factors identified through Boruta analysis, Lasso regression and clinical judgment. When comparing to the lowest quartile of EASIX (Q1) as a reference (Table 2, P for trend < 0.001), the odds of in-hospital death increased in Q2 (OR 1.76, 95% CI 1.34–2.32), Q3 (OR 1.94, 95% CI 1.45–2.60), and Q4 (OR 3.08, 95% CI 2.19–4.33). Furthermore, no evidence of a nonlinear relationship between EASIX and in-hospital mortality was found in the RCS model (Nonlinear P = 0.718) (Fig. 4). We evaluated the predictive performance of EASIX, SOFA, and CHA₂DS₂–VASc scores for in-hospital mortality using ROC analysis. The AUCs were 0.683 (95% CI 0.666–0.701) for EASIX, 0.705 (95% CI 0.688–0.721) for SOFA, and 0.506 (95% CI 0.488–0.524) for CHA₂DS₂–VASc. Delong's test indicated that EASIX had a slightly lower AUC than SOFA (P = 0.016) but was significantly higher than CHA₂DS₂–VASc (P < 0.001) (Fig. 5).

Restricted cubic spline regression analysis of EASIX with all-cause mortality. Restricted cubic spline regression analysis of EASIX with in hospital A 28-day, B 365-day C all-cause mortality

ROC curves for predicting in-hospital all-cause mortality

The association between EASIX and both 28-day and 365-day mortality was evaluated using Cox proportional hazards models. Multivariable Cox regression analysis revealed a significant correlation between elevated EASIX and increased risk of 28-day mortality (HR 1.21, 95% CI 1.16–1.26, P < 0.001). Compared to the lowest EASIX quartile (Q1, reference group, P for trend < 0.001), the hazard ratios (HRs) for 28-day mortality were notably higher in Q2 (HR 1.55, 95% CI 1.28–1.88), Q3 (HR 1.63, 95% CI 1.33–1.99), and Q4 (HR 2.36, 95% CI 1.88–2.96).A similar pattern was observed for 365-day mortality, with higher EASIX levels significantly correlating with increased risk of death at 1 year. Multivariable Cox regression confirmed that EASIX independently predicted 365-day mortality, where each 1-point rise in EASIX was associated with a 1.16-fold increased risk of death (HR 1.16, 95% CI 1.12–1.21) after adjusting for confounding variables. Furthermore, patients in the Q4, Q3, and Q2 quartiles had 1.39, 1.54, and 1.98 times the risk of 365-day mortality, respectively, compared to those in the lowest quartile (Q1) (Table 3). No evidence of a nonlinear relationship between EASIX and either 28-day (P = 0.302) or 365-day mortality (P = 0.547) was detected using restricted cubic spline (RCS) modeling (Fig. 4).

ROC analysis showed that for 28-day mortality prediction, EASIX (AUC 0.664, 95% CI 0.647–0.681) was not significantly different from SOFA (AUC 0.678, 95% CI 0.661–0.694; P = 0.117) but was significantly higher than CHA₂DS₂–VASc (AUC 0.540, 95% CI 0.523–0.557; P < 0.001). Similar findings were observed for 365-day mortality, where EASIX (AUC 0.649, 95% CI 0.633–0.664) showed no significant difference from SOFA (AUC 0.641, 95% CI 0.626–0.656; P = 0.334) but remained significantly superior to CHA₂DS₂–VASc (AUC 0.558, 95% CI 0.542–0.574; P < 0.001) (Supplementary Fig. S1).

Kaplan–Meier survival analysis stratified by EASIX quartiles revealed significant differences in both 28-day and 365-day mortality across quartiles (log-rank P < 0.0001), with the highest EASIX group showing the poorest survival outcomes (Fig. 6). Corrected pairwise comparisons indicated significant differences between the four groups, both for 28-day mortality and 365-day mortality (Bonferroni P < 0.05).

Kaplan–Meier survival analysis curves for all-cause mortality. Kaplan–Meier curves and cumulative incidence of 28-day (A) and 365-day (B) all-cause mortality stratified by EASIX

To further investigate whether the associations between EASIX and in-hospital, 28-day, and 365-day all-cause mortality held across various conditions, subgroup analyses were performed based on age, gender, race, hypertension, HF, MI, malignant tumors, CKD, COPD hyperlipidemia, stroke, and diabetes while adjusting for medications and interventions that are shown in the Table 1.

Stratified analysis of subgroups revealed that the association between EASIX and in-hospital mortality was more pronounced in older patients (P < 0.001, P for interaction = 0.018) and those of White descent (P < 0.001, P for interaction = 0.035), as evidenced by higher OR in these groups. Specifically, the OR for older patients was 1.23 (95% CI 1.17–1.29), while for White patients, it was 1.25 (95% CI 1.18–1.32), indicating that both age and race significantly modify the effect of EASIX on in-hospital mortality (Fig. 7).

Forest plots of stratified analyses of EASIX and in-hospital all-cause mortality

In the analysis of 28-day mortality, patients with a history of HF demonstrated a lower risk (P < 0.001, HR 1.13, 95% CI 1.09–1.18) compared to the control group (P < 0.001, HR 1.19, 95% CI 1.14–1.23) (P for interaction = 0.016).This finding suggests that the presence of HF may attenuate the effect of EASIX on 28-day mortality risk. Furthermore, no significant subgroup effects were observed in other strata, with all suggesting a significant association between EASIX and the outcome, further supporting the robustness of our conclusions (Supplementary Fig. S2).

When analyzing 365-day mortality, EASIX was significantly associated with mortality risk across all subgroups. However, significant interactions were observed for age (P for interaction = 0.010), race (P for interaction = 0.036), hypertension (P for interaction = 0.005), HF (P for interaction = 0.010), CKD (P for interaction = 0.003), and hyperlipidemia (P for interaction = 0.004), suggesting that these variables may modify the relationship between EASIX and long-term mortality risk (Supplementary Fig. S2).