The impact of the Elixhauser comorbidity index on length of stay in patients with coronary artery disease

Abstract

Objectives

To investigate the association between the modified Elixhauser Comorbidity Index (ECI) and length of stay (LOS) among patients with coronary artery disease (CAD).

Methods

A retrospective study was conducted using hospitalization data from the Hubei Provincial Traditional Chinese Medicine Comprehensive Statistics Platform between December 2013 and June 2024. Comorbidity burden was assessed using the modified ECI. Univariate analysis and multivariable Gamma regression models with a log-link function were performed to identify factors independently associated with LOS. Sensitivity analyses were conducted to evaluate the robustness of the findings.

Results

Among 22,626 CAD patients with a median LOS of 9 (6–13) days, ECI remained independently associated with LOS after adjustment for demographic and clinical covariates. Compared with ECI ≤ 0, patients with ECI 1–3 had 10.0% longer LOS (Exp(B) = 1.100, 95% CI = 1.064–1.138, p < 0.001), and those with ECI 4–8 had 2.9% longer LOS (Exp(B) = 1.029, 95% CI = 1.003–1.055, p = 0.026). The ECI ≥ 9 group was not statistically significant in the primary analysis but became significant after excluding patients with extremely prolonged LOS (p = 0.027). A paradoxical pattern was observed, with greater LOS prolongation in patients with mild rather than moderate-to-high comorbidity burden.

Conclusion

ECI was independently associated with LOS in CAD patients. The non-linear association between comorbidity burden and LOS suggests a more complex relationship than a simple dose-response pattern, which may have implications for risk stratification and individualized inpatient management in CAD populations.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12872-026-06195-9.

Background

Coronary artery disease (CAD), also known as coronary heart disease, is characterized by atherosclerotic narrowing or obstruction of the coronary arteries, resulting in impaired myocardial perfusion and ischemia [1]. CAD is one of the leading causes of cardiovascular morbidity and mortality worldwide [2]. Although CAD traditionally occurs in middle-aged and older adults, recent studies have reported an increasing trend toward earlier onset [3]. In addition, the incidence of CAD remains higher and the age of onset is earlier in males than in females [4]. Length of stay (LOS) is an important indicator of healthcare service quality, resource utilization efficiency, and patient prognosis. Prolonged LOS is associated with increased healthcare costs and adverse outcomes, such as hospital-acquired infections and postoperative complications [5, 6]. Identifying key factors associated with LOS in CAD patients is clinically important for improving risk stratification, optimizing inpatient management, and facilitating efficient healthcare resource allocation.

Comorbidity refers to the presence of one or more additional diseases coexisting with a primary disease [7]. It exacerbates the clinical condition and prolongs the recovery period through complex pathophysiological mechanisms. Comorbidity increases disease complexity and is associated with prolonged recovery and adverse clinical outcomes. Moreover, previous studies have demonstrated that comorbidity is associated with prolonged LOS in hospitalized patients [8–10]. CAD patients frequently present with comorbidities such as diabetes and hypertension, making comorbidity clinically prevalent [11]. These comorbidities not only increase the complexity of treatment but may also lead to a higher risk of postoperative complications and longer recovery times, thereby contributing to longer LOS [12]. The Elixhauser Comorbidity Index (ECI) is a classic tool for quantifying comorbidity, encompassing 30 clinically common chronic diseases, and was updated to the International Classification of Diseases (ICD), 10th Revision (ICD-10) version by Quan et al. [13]. Walraven et al. [14] further refined it into a weighted scoring system, using a weighting approach similar to that of the Charlson Comorbidity Index (CCI), and applied it to predict outcomes such as patient mortality and hospital readmission.

Although existing studies have confirmed the prognostic value of the ECI for diseases such as pneumonia and stroke [15–17], there is still a lack of large-sample studies focusing on the LOS in CAD patients, particularly in Chinese patients. Furthermore, existing research has predominantly focused on patient mortality, insurance types, and medical costs [17–19]. Moreover, as LOS data typically exhibit a positively skewed distribution, traditional linear regression models may fail to accurately fit the data, leading to biased results. Therefore, this study employed a retrospective cohort design, enrolling 22,626 CAD patients from the Hubei Provincial Traditional Chinese Medicine Comprehensive Statistics Platform. Using univariate analysis and multivariable Gamma regression models, it systematically investigated the association between ECI, clinical factors, and LOS. Sensitivity analysis was conducted to evaluate the robustness of the findings. The study aims to provide a scientific basis for the clinical management of CAD patients and optimize the allocation of healthcare resources.

Methods

Study population

This was a retrospective study using CAD patient data from the Hubei Provincial Traditional Chinese Medicine Comprehensive Statistics Platform between December 2013 and June 2024. Inclusion criteria were: (1) age ≥ 18 years; (2) primary diagnosis of CAD; (3) complete clinical data. Exclusion criteria were: (1) incomplete data; (2) LOS < 1 day; (3) death during hospitalization; (4) discharge against medical advice; (5) other unspecified discharge status. Patients who died in hospital were excluded because death truncates the LOS and introduces a competing risk beyond the scope of this study. Patients discharged against medical advice or with other unspecified discharge status were excluded because their LOS may be affected by non-clinical factors such as economic constraints, personal decisions, or administrative factors rather than clinical conditions alone. Ultimately, 22,626 cases were included in the study.

Data collection

Clinical data were collected from the traditional Chinese medicine (TCM) inpatient medical records, including gender, age, mode of admission, mode of discharge, surgical status, treatment category, comorbidities, and LOS. According to the national standard issued by the National Administration of Traditional Chinese Medicine (2012) [20], treatment categories were mutually exclusive and classified as follows: (1) TCM alone, defined as treatment primarily using Chinese herbal medicine or non-pharmacological TCM therapies, (2) integrated traditional Chinese and Western medicine, defined as the combined use of both modalities; and (3) Western medicine (WM) alone, defined as treatment dominated by WM. The ECI, as modified by Walraven et al. [14] (Table 1), was used to calculate weighted comorbidity scores. The ECI was calculated using only pre-existing comorbidities documented at the time of admission; new diagnoses made during hospitalization were not included. The ECI scores were categorized into four groups according to the distribution of scores in this cohort, with the 25th, 50th, and 75th percentiles corresponding to scores of 0, 3, and 8, respectively: ≤ 0, 1–3, 4–8, and ≥ 9.

Table 1.

Modified Elixhauser Comorbidity Index

No	Comorbidity Group	Score
1	Congestive heart failure	7
2	Cardiac arrhythmias	5
3	Vascular disease	-1
4	Pulmonary circulation disorders	4
5	Peripheral vascular disorders	2
6	Hypertension	0
7	Paralysis	7
8	Other neurological disorders	6
9	Chronic pulmonary disease	3
10	Diabetes, uncomplicated	0
11	Diabetes, complicated	0
12	Hypothyroidism	0
13	Renal failure	5
14	Liver disease	11
15	Peptic ulcer disease, excluding bleeding	0
16	AIDS/HIV	0
17	Lymphoma	9
18	Metastatic cancer	12
19	Solid tumor without metastasis	4
20	Rheumatoid arthritis/collagen vascular diseases	0
21	Coagulopathy	3
22	Obesity	-4
23	Weight loss	6
24	Fluid and electrolyte disorders	5
25	Blood loss anemia	-2
26	Deficiency anemia	-2
27	Alcohol abuse	0
28	Drug abuse	-7
29	Psychoses	0
30	Depression	-3

Statistical methods

Statistical analysis was performed using SPSS 26.0 and Python 3.8. The distribution of the primary outcome, LOS, was assessed using the Kolmogorov-Smirnov test, which indicated a positively skewed distribution. Continuous data are presented as M (Q₁, Q₃) and categorical data as frequency (%). A p-value < 0.05 was considered statistically significant.

Univariate analysis

All candidate variables (gender, age group, mode of admission, mode of discharge, treatment category, surgical status, and ECI score) were examined using univariate analysis, including the Mann-Whitney U test for two groups and the Kruskal-Wallis H test for comparisons among three or more groups.

Multivariable Gamma regression

A multivariable Gamma regression model with a log-link function was fitted, with LOS as the dependent variable. The model included all variables that were significant in univariate analysis, adjusting for potential confounders: gender, age group, mode of admission, mode of discharge, treatment category, and surgical status. All categorical variables were dummy-coded. The model formula is log(µ_i) = β₀ + β₁ · gender + β₂ · age_group + β₃ · mode_of_admission + β_{4 ·} mode_of_discharge + β₅ · treatment_category + β₆ · surgical_status + β₇ · ECI_score, where µ_i represents the mean LOS for an individual i, β₀ is the intercept, and β₁-β₇ are the corresponding regression coefficients for each adjusted variable.

Multicollinearity and model diagnostics

Multicollinearity among independent variables was assessed by calculating the variance inflation factor (VIF) using linear regression. Residual diagnostics for the Gamma regression model consisted of a scatterplot of deviance residuals versus fitted values and a Q-Q plot of standardized Pearson residuals. The dispersion statistics (deviance/df and Pearson χ²/df) and the Cox-Snell pseudo-R² were reported as part of the model fit evaluation.

Sensitivity analyses

A sensitivity analysis using log-normal regression (ordinary least squares on log-transformed LOS) was conducted with the same covariates, testing the robustness of the findings to alternative distributional assumptions. Results from the log-normal regression model were expressed as exponentiated coefficients (geometric mean ratios), directly comparable to the Exp(B) from the Gamma model. Additionally, because extremely prolonged LOS might be confounded by non-disease-related factors, the top 1% of patients with LOS exceeding the 99th percentile (P99) were excluded, and the Gamma model was re-fitted. Furthermore, to assess the robustness of the ECI categorization, we performed an additional sensitivity analysis using the Elixhauser classification categories adopted by Potts et al. [21] (< 0, = 0, 1–5, 6–13, ≥ 14). Finally, to address the potential bias arising from the exclusion of in-hospital deaths in the primary analysis, we performed a competing risk analysis using a Fine-Gray proportional subdistribution hazards model, with discharge alive as the event of interest, in-hospital death as the competing event, and transferred patients treated as censored. The model included the same covariates as the primary Gamma regression, excluding mode of discharge, which was used to define the event and censoring status in the competing risk framework.

Subgroup and interaction analyses

Whether the association between ECI and LOS varied by CAD subtype and treatment modality was evaluated through subgroup analyses. Patients were stratified first by CAD subtype (myocardial infarction (MI) vs. stable angina (SA), based on ICD-10 codes), and second by treatment modality (revascularization vs. non-revascularization). The revascularization group comprised patients who underwent percutaneous coronary intervention (PCI) or coronary artery bypass grafting (CABG), identified by International Classification of Diseases, 9th Revision, Clinical Modification, Volume 3 (ICD-9-CM-3) procedure codes 36.01–36.03, 36.05–36.07, 36.09, 00.66, 36.10–36.17, and 36.19; the non-revascularization group comprised patients without any coronary revascularization procedure. The same Gamma regression model was fitted in each subgroup, except that the surgical status variable (any procedure) was included only in the non-revascularization group; in the revascularization subgroup, this variable was constant (all patients underwent revascularization) and therefore omitted.

Effect modification by clinically important factors was examined using interaction analyses. Based on clinical plausibility, three potential effect modifiers were selected: age group, surgical status, and treatment category. Interaction terms between ECI group and each of these covariates were introduced one at a time into the primary Gamma regression model. The significance of each interaction term was evaluated using the Wald test, and the p-value for interaction was reported. No adjustment for multiple comparisons was applied in these exploratory analyses.

Individual comorbidity analyses

The five most prevalent Elixhauser comorbidities in the cohort were selected for evaluation of their independent effects on LOS beyond the ECI: hypertension, heart failure, diabetes, chronic pulmonary disease, and cardiac arrhythmias. For each comorbidity, a separate Gamma regression model was fitted, including the ECI score, a binary indicator for that comorbidity, and all covariates from the primary model (gender, age group, mode of admission, mode of discharge, treatment category, and surgical status). For heart failure, an interaction term between heart failure and treatment category was introduced in the same model to test effect modification. Subgroup analyses for heart failure by treatment category were also performed using the same Gamma regression model.

Results

Baseline characteristics of patients

This study included a total of 22,626 CAD patients, comprising 11,591 males (51.2%) and 11,035 females (48.8%). The age range was 22 to 100 years, with a median age of 71 (64–91) years. Elderly patients aged ≥ 65 years constituted the majority, accounting for 72.7%. Outpatient admissions accounted for 78.3%, and emergency admissions for 15.6%. The primary treatment category was integrated traditional Chinese and Western medicine (63.5%), followed by TCM alone (20.0%) and WM alone (16.4%). Discharge on medical advice accounted for 99.2% of cases, and 7.9% of patients had a history of surgery. Regarding comorbidities, 20,842 (92.1%) patients had at least one comorbidity, and 2,760 (12.2%) had seven comorbidities. The most common Elixhauser comorbidities, in descending order of prevalence, were hypertension, congestive heart failure, chronic pulmonary disease, cardiac arrhythmias, and diabetes (see Supplementary Table S1). The median ECI score was 3 (0–8), 9,740 (43.0%) patients had a score ≤ 0, 2,504 (11.1%) had scores 1–3, 4,886 (21.6%) had scores 4–8, and 5,496 (24.3%) had scores ≥ 9. The median LOS was 9 (6–13) days. Detailed data are presented in Table 2.

Table 2.

Results of univariate analysis of patients’ LOS

Variable	N (%)	Length of Stay [M (Q1, Q3)]	Mann-Whitney U / Kruskal-Wallis H	p-value
Gender			63,186,459.500	0.118
Male	11,591 (51.2)	9 (6–13)
Female	11,035 (48.8)	9 (6–13)
Age(year)			613.131	< 0.001
18–44	260 (1.1)	6.5 (4–11)
45–59	3336 (14.7)	7 (5–11)
60–74	10,519 (46.5)	9 (6–13)
≥ 75	8511 (37.6)	10 (7–14)
Mode of admission			15.106	0.001
Emergency	3523 (15.6)	9 (6–13)
Outpatient	17,721 (78.3)	9 (6–13)
Others	1382 (6.1)	9 (6–12)
Mode of discharge			137.157	0.001
Discharge on medical advice	22,434 (99.2)	9 (6–13)
Interhospital transfer on medical advice	172 (0.8)	4 (2–7)
Transfer to primary care facilities on medical advice	20 (0.1)	12.5 (5.25–18)
Treatment category			1510.879	0.001
Traditional Chinese medicine	4536 (20.0)	13 (8–16)
Integrated traditional Chinese and Western medicine	14,370 (63.5)	8 (6–12)
Western medicine	3720 (16.4)	8 (5–11)
Surgical status			16,470,870.000	0.001
No	20,843 (92.1)	9 (6–13)
Yes	1783 (7.9)	8 (5–12)
ECI score			291.185	0.001
≤ 0	9740 (43.0)	9 (6–13)
1–3	2504 (11.1)	11 (7–15)
4–8	4886 (21.6)	9 (6–13)
≥ 9	5496 (24.3)	9 (6–12)

Percentages may not add up to 100% due to rounding

Model diagnostics

The VIF for all covariates was below 1.1, indicating no problematic multicollinearity. The Gamma regression model yielded a deviance/df of 0.357 and a Pearson χ²/df of 0.495, indicating moderate underdispersion; the Cox-Snell pseudo-R² was 0.104. The residuals-versus-fitted plot (Supplementary Figure S1) demonstrated a mild funnel-shaped pattern consistent with the mean-variance relationship of the Gamma distribution, with curved lower boundaries reflecting the positive constraint of LOS. The LOWESS smoother showed no substantial systematic trend, suggesting adequate specification of the log-link functional form and providing indirect support for the proportional effect assumption of the log-link model. The normal Q-Q plot of standardized Pearson residuals (Supplementary Figure S2) followed the reference line reasonably well in the central range but showed marked deviation in the upper tail, with several residuals exceeding 30. These findings suggest that the model fit was adequate for the majority of observations but underestimated LOS for a small subset of patients with extremely prolonged hospitalization. Although underdispersion is less commonly discussed in Gamma regression, it indicates that the observed variability was smaller than expected under the assumed distribution. This condition is unlikely to bias coefficient estimates but may result in conservative standard error estimation. Importantly, sensitivity analyses using log-normal regression (Supplementary Table S2) produced nearly identical effect estimates, supporting the robustness of the primary findings despite the observed residual tail behavior and underdispersion. The heavy upper tail suggests that a small proportion of patients with extremely prolonged LOS may disproportionately affect model fit and prediction accuracy in high-LOS subgroups. To assess the robustness of the findings, we conducted sensitivity analyses excluding the top 1% of patients with the longest LOS and using a log-normal regression model (see below).

Univariate analysis of LOS

The results of univariate analysis showed statistically significant differences in LOS across the following six variables: age group (Kruskal-Wallis H = 613.131, df = 3, p < 0.001), mode of admission (Kruskal-Wallis H = 15.106, df = 2, p = 0.001), mode of discharge (Kruskal-Wallis H = 137.157, df = 2, p < 0.001), surgical status (Mann-Whitney U = 16,470,870.000, Z = -7.989, p < 0.001), treatment category (Kruskal-Wallis H = 1510.879, df = 2, p < 0.001), and ECI score (Kruskal-Wallis H = 291.185, df = 3, p < 0.001). No statistically significant difference in LOS was observed for gender (Mann-Whitney U = 63,186,459.500, Z = -1.565, p = 0.118). Specifically, LOS increased significantly with age, particularly for patients over 60 years old who had markedly longer hospital stays. Patients receiving WM alone, those who underwent surgery, and those with interhospital transfer on medical advice had relatively longer LOS. Regarding the ECI, patients with a score of 1–3 had the longest LOS, while those in all other ECI score groups had relatively shorter LOS. Detailed data are presented in Table 2.

Multivariable analysis of LOS

Using patients’ LOS as the dependent variable, variables with statistically significant associations in the univariate analysis were included in a multivariable Gamma regression model with a log-link function. The univariate analysis indicated that gender was not statistically significant in patients’ LOS. Although gender was not statistically significant in the univariate analysis, it was included in the Gamma regression model as a key demographic covariate to adjust for potential confounding.

The Omnibus test indicated that the overall model was statistically significant (Likelihood Ratio χ² = 2268.928, df = 14, p < 0.001). After adjusting for confounding factors, gender, age group, treatment category, mode of discharge, surgical status, and ECI score showed statistically significant associations with LOS in CAD patients (p < 0.001), whereas the mode of admission was not statistically significantly associated with LOS (p = 0.104). Specifically, patients with ECI scores of 1–3 (Exp(B) = 1.100, 95% CI = 1.064–1.138, p < 0.001) and 4–8 (Exp(B) = 1.029, 95% CI = 1.003–1.055, p = 0.026) were associated with longer LOS, increasing it by 10.0% and 2.9%, respectively. Age group was associated with longer LOS, with older groups showing progressively longer LOS. Females had a shorter LOS than males (Exp(B) = 0.938, 95% CI = 0.920–0.955, p < 0.001). Treatment category, particularly WM alone, was associated with shorter LOS (Exp(B) = 0.701, 95% CI = 0.682–0.721, p < 0.001). Patients who underwent surgery had a shorter LOS (Exp(B) = 0.911, 95% CI = 0.886–0.937, p < 0.001). Detailed data are presented in Table 3.

Table 3.

Results of multivariable analysis for patients’ LOS

Variable	Exp(B) (95% CI)	p-value
Gender		< 0.001
Male (Ref.)	1.000
Female	0.938 (0.920–0.955)	< 0.001
Age (year)		< 0.001
18–44 (Ref.)	1.000
45–59	1.104 (1.016–1.199)	0.019
60–74	1.252 (1.155–1.357)	< 0.001
≥ 75	1.518 (1.398–1.648)	< 0.001
Mode of admission		0.104
Emergency (Ref.)	1.000
Outpatient	1.024 (0.999–1.049)	0.063
Others	1.044 (0.996–1.094)	0.074
Mode of discharge		< 0.001
Discharged on medical advice (Ref.)	1.000
Interhospital transfer on medical advice	0.589 (0.514–0.674)	< 0.001
Transfer to primary care facilities on medical advice	1.623 (0.845–3.116)	0.146
Treatment category		< 0.001
Traditional Chinese medicine (Ref.)	1.000
Integrated traditional Chinese and Western medicine	0.760 (0.744–0.776)	< 0.001
Western medicine	0.701 (0.682–0.721)	< 0.001
Surgical status		< 0.001
No (Ref.)	1.000
Yes	0.911 (0.886–0.937)	< 0.001
ECI score		< 0.001
≤ 0 (Ref.)	1.000
1–3	1.100 (1.064–1.138)	< 0.001
4–8	1.029 (1.003–1.055)	0.026
≥ 9	1.009 (0.985–1.034)	0.479

Sensitivity analyses

As summarized in Supplementary Table S2, the log-normal regression model yielded effect estimates consistent with the primary Gamma regression model: ECI 1–3 was associated with an 11.0% longer LOS (Exp(B) = 1.110, 95% CI: 1.081–1.141, p < 0.001), ECI 4–8 with a 3.2% longer LOS (Exp(B) = 1.032, 95% CI: 1.010–1.055, p = 0.004), and ECI ≥ 9 with a 3.3% longer LOS (Exp(B) = 1.033, 95% CI: 1.012–1.055, p = 0.003). The paradoxical pattern (greater prolongation in ECI 1–3 than in ECI 4–8) persisted, and notably, the ECI ≥ 9 group became statistically significant in this sensitivity analysis. These results confirm that the association between higher comorbidity burden and prolonged LOS is robust to alternative distributional assumptions.

In a sensitivity analysis excluding the top 1% of patients with extremely long LOS (LOS > 35 days, n = 206), the Gamma regression model was re-fitted. The overall statistical significance and direction of associations for all variables remained consistent with the primary analysis. Specific changes were as follows: the ECI score ≥ 9 group changed from non-significant (p = 0.479) to significant (p = 0.027); for the age group 45–59 years, the p-value shifted from 0.019 to 0.056, attenuated to marginal significance; the outpatient admission route also changed from non-significant (p = 0.063) to significant (p = 0.034). Detailed data are presented in Supplementary Table S3.

The emergence of significance for ECI ≥ 9 upon exclusion of extreme LOS outliers prompted their characterization. As shown in Supplementary Table S4, these patients were significantly older (79.6% aged ≥ 75 years vs. 37.2% in the retained cohort), predominantly male (75.7% vs. 51.0%), and carried a higher comorbidity burden: ECI ≥ 9 in 31.6% vs. 24.2% (p = 0.005), heart failure in 39.8% vs. 27.3% (p < 0.001), cardiac arrhythmias in 15.5% vs. 7.6% (p < 0.001), and chronic pulmonary disease in 15.5% vs. 10.3% (p = 0.014). These findings indicate that the association between ECI ≥ 9 and LOS was obscured in the primary analysis by this subset of patients with extreme LOS; their exclusion revealed a significant effect.

Furthermore, we performed a sensitivity analysis using the Elixhauser classification categories adopted by Potts et al. (< 0, = 0, 1–5, 6–13, ≥ 14). As shown in Supplementary Table S5, this analysis confirmed the non-monotonic pattern: the ECI 1–5 group was associated with a 13.1% longer LOS (Exp(B) = 1.131, 95% CI: 1.009–1.268, p = 0.035), whereas the < 0, = 0, 6–13, and ≥ 14 groups were not significantly associated with LOS. These results further confirm that the paradoxical finding is robust to different ECI grouping strategies.

In the competing risk analysis that re-included in-hospital deaths (n = 116; Supplementary Table S6), higher ECI scores were associated with a significantly lower subdistribution hazard of discharge alive. Compared with ECI ≤ 0, patients with ECI 1–3 had a significantly lower subdistribution hazard (SHR = 0.890, 95% CI: 0.857–0.923, p < 0.001), indicating slower discharge and longer LOS. Similar associations were observed for ECI 4–8 (SHR = 0.947, 95% CI: 0.916–0.980, p = 0.002) and ECI ≥ 9 (SHR = 0.958, 95% CI: 0.925–0.991, p = 0.013). The ECI ≥ 9 group also reached statistical significance in the competing risk analysis, consistent with findings from other sensitivity analyses. The cumulative incidence functions (Supplementary Figure S3) showed modest differences in the cumulative incidence of discharge alive across ECI groups, whereas the cumulative incidence of in-hospital death remained very low and differed minimally between groups. These findings suggest that the observed association between higher comorbidity burden and prolonged LOS was primarily related to slower discharge rather than differential in-hospital mortality.

Subgroup analyses

Given the paradoxical finding that the ECI 1–3 group had a greater LOS prolongation than the ECI 4–8 group, we examined whether this pattern varied by CAD subtype. A forest plot of the ECI effects is provided in Supplementary Figure S4, and full regression results in Supplementary Table S7. In the MI subgroup (n = 590), the ECI 1–3 group showed a non-significant prolongation (Exp(B) = 1.142, 95% CI: 0.933–1.398, p = 0.197), whereas ECI ≥ 9 was significantly associated with longer LOS (Exp(B) = 1.163, 95% CI: 1.010–1.339, p = 0.036). In the SA subgroup (n = 773), no ECI group reached significance; point estimates for ECI 4–8 and ≥ 9 were below 1 (Exp(B) = 0.905 and 0.956, respectively). The paradoxical pattern observed in the full cohort was not consistently present in either subgroup; a significant association of comorbidity with LOS in MI patients was confined to the highest comorbidity burden category (ECI ≥ 9).

Treatment modality was next examined. Patients were stratified into revascularization (PCI/CABG, n = 366) and non-revascularization groups (n = 22,260). The same Gamma regression model was fitted in each; surgical status was omitted in the revascularization subgroup (constant) but retained in the non-revascularization subgroup. Full results are in Supplementary Table S8. In the non-revascularization group, the effect pattern mirrored the primary analysis: ECI 1–3 prolonged LOS by 10.3% (Exp(B) = 1.103, 95% CI: 1.066–1.141, p < 0.001), ECI 4–8 by 2.9% (Exp(B) = 1.029, 95% CI: 1.004–1.056, p = 0.024), and ECI ≥ 9 was non-significant (Exp(B) = 1.009, 95% CI: 0.984–1.034, p = 0.486). In the revascularization group, no ECI category was significantly associated with LOS (ECI 1–3: Exp(B) = 0.950, 95% CI: 0.834–1.082, p = 0.436; ECI 4–8: Exp(B) = 1.039, 95% CI: 0.909–1.188, p = 0.571; ECI ≥ 9: Exp(B) = 1.002, 95% CI: 0.881–1.141, p = 0.971); the overall Wald test for ECI was non-significant (p = 0.751). Revascularization attenuated the association of comorbidity burden with LOS; the paradoxical pattern was observed exclusively in patients treated without revascularization.

Interaction analyses

We tested whether the association between ECI and LOS was modified by age group, surgical status, or treatment category by adding interaction terms to the Gamma regression model. As shown in Supplementary Table S9, no statistically significant interaction was found for ECI × age group (p = 0.090), ECI × surgical status (p = 0.087), or ECI × treatment category (p = 0.424). These results indicate that the association between ECI and LOS did not differ significantly across age groups, surgical status, or treatment categories.

Individual comorbidity analyses

To evaluate whether individual Elixhauser comorbidities had an independent effect on LOS beyond the global ECI burden, the five most prevalent comorbidities in the cohort— hypertension, heart failure, diabetes, chronic pulmonary disease, and cardiac arrhythmias—were added separately to the fully adjusted Gamma regression model (each in a separate model, with ECI and other covariates retained). As summarized in Supplementary Table S10, all five comorbidities were significantly associated with LOS after adjustment for ECI. Hypertension (Exp(B) = 1.042, 95% CI: 1.023–1.062, p < 0.001), diabetes (Exp(B) = 1.101, 95% CI: 1.076–1.128, p < 0.001), chronic pulmonary disease (Exp(B) = 1.065, 95% CI: 1.031–1.101, p < 0.001), and cardiac arrhythmias (Exp(B) = 1.061, 95% CI: 1.022–1.100, p = 0.002) were associated with longer LOS, whereas heart failure showed a negative association (Exp(B) = 0.965, 95% CI: 0.939–0.991, p = 0.009).

The unexpected negative association of heart failure prompted an examination of its interaction with treatment category. A significant interaction was observed (p < 0.001). Subgroup analyses (Supplementary Table S11) revealed that among patients receiving TCM alone, heart failure was associated with significantly shorter LOS (Exp(B) = 0.836, 95% CI: 0.771–0.907, p < 0.001), whereas no significant association was found in the integrated traditional Chinese and Western medicine group (Exp(B) = 0.984, 95% CI: 0.953–1.016, p = 0.315) or the WM alone group (Exp(B) = 0.929, 95% CI: 0.863–1.001, p = 0.053). Thus, the negative association was driven primarily by patients receiving TCM alone.

Discussion

Association between ECI and LOS in CAD patients and clinical implications

This study confirms that the ECI is independently associated with LOS in CAD patients, consistent with prior research demonstrating the prognostic value of comorbidity burden across diverse clinical populations [21–23]. As a tool for assessing patient comorbidity burden, the ECI encompasses various chronic diseases closely related to CAD prognosis, such as hypertension, diabetes, and heart failure. Its score directly reflects the patient’s multi-organ functional reserve status and treatment complexity [21]. However, a paradoxical dose-response pattern emerged: patients with mild comorbidity burden (ECI 1–3) exhibited a 10.0% prolongation of LOS, whereas those with moderate-high burden (ECI 4–8) showed only a 2.9% prolongation. This counterintuitive pattern was reproduced across sensitivity analyses (log-normal regression model: 11.0% vs. 3.2%), indicating that it is not an artifact of model misspecification. Notably, this paradoxical pattern differs from the commonly observed gradient of higher comorbidity burden leading to worse outcomes in other populations. For instance, Ng et al. demonstrated that in patients undergoing inpatient procedures, LOS increased with higher Elixhauser scores, with the longest stays observed among those with ECI ≥ 6, particularly in the presence of bleeding complications [24]; similarly, Potts et al. reported that in over 6.6 million PCI patients, very high ECI scores (> 13) were associated with a fivefold increase in mortality and doubled hospitalization costs [21]. Our findings suggest that the association of comorbidity burden with LOS in CAD patients may be more complex than a simple dose-response relationship, warranting further investigation.

One plausible interpretation is that this pattern reflects differential treatment intensity across comorbidity strata. Patients with moderate-high comorbidity burden (ECI 4–8) may be more readily recognized as high-risk and therefore receive more aggressive interventions (e.g., early revascularization, intensified medication regimens, specialist consultations) that partially offset the negative impact of comorbidities on LOS [25, 26]. In contrast, patients with mild comorbidity burden (ECI 1–3) may be perceived as lower-risk and receive routine care without targeted comorbidity management, leaving their LOS more susceptible to delays from minor complications or unresolved symptoms. Although direct measures of treatment intensity (e.g., ICU admission, medication count) were unavailable in this dataset, two findings support this interpretation. First, the composition of Elixhauser comorbidities differs markedly across ECI groups (Supplementary Table S1). Patients in the ECI 4–8 group had a substantially higher prevalence of congestive heart failure (51.0%) and cardiac arrhythmias (24.1%) compared with the ECI 1–3 group (1.5% and 0.5%, respectively). In contrast, the ECI 1–3 group showed an exceptionally high prevalence of chronic pulmonary disease (72.0%), largely due to the broad Quan definition (which includes asthma, chronic bronchitis, and emphysema, not only chronic obstructive pulmonary disease), along with a high prevalence of hypertension (61.0%) and peripheral vascular disorders (25.9%). Clinical practice guidelines indicate that acute management of congestive heart failure and cardiac arrhythmias frequently requires intravenous medications, continuous monitoring, and timely specialist intervention [27, 28], whereas hospital care for chronic pulmonary disease and hypertension predominantly relies on less intensive, conventional protocols [29, 30]. These patterns suggest that the moderate-high burden group is dominated by cardiac-specific comorbidities that may trigger standardized aggressive cardiology protocols, while the mild burden group includes a high proportion of respiratory and hypertensive conditions that may lead to more diffused treatment goals. Second, formal interaction analyses revealed no significant effect modification by age group, surgical status, or treatment category. However, when examined by specific treatment modality, the paradoxical ECI-LOS pattern was observed exclusively in non-revascularized patients; in those undergoing PCI or CABG, the ECI-LOS association was completely attenuated. This offers a direct explanatory mechanism: revascularization may attenuate the impact of comorbidity burden on LOS, as the paradoxical pattern was completely absent in patients undergoing PCI or CABG.

The sensitivity analysis further clarified the effect of a very high comorbidity burden. The ECI ≥ 9 group, non-significant in the primary Gamma model (Exp(B) = 1.009, p = 0.479), became significant after excluding extreme outliers (p = 0.027) and in the log-normal regression model (Exp(B) = 1.033, p = 0.003). The excluded outlier patients were significantly older, predominantly male, and carried a markedly higher comorbidity burden; their inclusion introduced noise that diluted the ECI-LOS association. This consistency across sensitivity approaches suggests that the lack of significance in the primary analysis was likely due to moderate underdispersion and a modest effect size, rather than a true null effect.

The competing risk analysis further supported this interpretation. All ECI groups, including ECI ≥ 9, were associated with a significantly lower subdistribution hazard of discharge alive (i.e., prolonged LOS). The cumulative incidence of in-hospital death was very low and differed minimally across ECI groups, indicating that the prolonged LOS associated with higher comorbidity burden was primarily driven by slower discharge rather than by differential mortality. Thus, the exclusion of deceased patients in the primary analysis was not the primary source of the paradoxical pattern; rather, it may have masked a significant prolongation effect in the highest comorbidity group. This mitigates the concern that excluding deaths could lead to a false-negative finding for ECI ≥ 9, and confirms that the paradoxical dose-response pattern is not an artifact of survival bias.

Subgroup analyses by CAD subtype revealed additional heterogeneity. In MI patients (n = 590), the ECI 1–3 group was not significantly associated with LOS, whereas only ECI ≥ 9 was associated with longer LOS. In stable angina patients (n = 773), no ECI group reached significance. These findings indicate that the overall paradoxical pattern in the full cohort likely arose from heterogeneity between acute and stable CAD, combined with differential treatment intensities.

Beyond the global ECI score, individual comorbidities exhibited heterogeneous associations with LOS. Hypertension, diabetes, chronic pulmonary disease, and cardiac arrhythmias remained independently associated with longer LOS after ECI adjustment, suggesting that the ECI weights may not fully capture the prognostic value of these high-prevalence conditions in this Chinese CAD cohort. Heart failure showed an unexpected negative association with LOS, driven entirely by patients receiving TCM alone (Exp(B) = 0.836). This finding likely reflects unmeasured selection bias, such as less severe heart failure, earlier disease stage, or greater family support enabling earlier discharge, rather than a therapeutic effect. Given the observational design and exploratory nature of this analysis, residual confounding cannot be excluded.

When translated into absolute days using the median LOS of 9 days, the 10.0% prolongation in the ECI 1–3 group corresponds to 0.9 days, and the 2.9% prolongation in the ECI 4–8 group to 0.26 days. These per-patient increments are modest, suggesting limited clinical impact at the individual level. Nevertheless, the consistent prediction of longer LOS across the primary and sensitivity analyses supports the ECI’s value as a risk stratification tool for hospital resource planning. The predictive effect, however, varies across clinical strata: it is preserved in non-revascularized patients but confined to the highest comorbidity category in MI patients.

Regarding treatment categories, patients receiving TCM alone had the longest LOS (median 13 days), while those receiving WM alone had the shortest LOS (median 8 days), with Integrated traditional Chinese and Western medicine in between (median 8 days) (Table 2). This pattern should be interpreted with caution. The longer LOS in the TCM alone group does not imply inferior efficacy of TCM. Instead, it likely reflects differences in patient selection and treatment goals. Patients who receive TCM alone may have milder disease, prefer conservative and holistic care, or require longer functional recovery. Conversely, WM alone is often used for acute symptom control and rapid discharge. The integrated traditional Chinese and Western medicine group may represent a balanced approach, where acute problems are managed with WM while TCM supports chronic symptom relief or postoperative recovery. Due to the observational design, no causal conclusions about treatment efficacy can be drawn, and these results should not be interpreted as evidence that TCM is inferior to WM.

These findings have practical implications for clinical management. First, ECI-based risk stratification should be implemented to identify patients with mild comorbidity burden (ECI 1–3) who may be overlooked in routine care but are at risk for prolonged LOS due to insufficient targeted management. Second, for patients with moderate-high burden (ECI 4–8), our data suggest that access to revascularization was associated with attenuated LOS prolongation; however, this observational finding requires randomized validation before informing treatment policy. Third, the ECI weights may require recalibration for Chinese CAD populations to better reflect the independent association of high-prevalence conditions such as hypertension and diabetes.

Limitations of this study

Several limitations should be acknowledged. First, this was a retrospective, single-region study using administrative data from a provincial TCM statistics platform, which limits generalizability to other geographic regions and healthcare systems with differing practice patterns. Second, the Gamma regression exhibited moderate underdispersion (deviance/df = 0.357), and the Q-Q plot indicated substantial right-tail deviation, suggesting that the Gamma distribution may not perfectly fit the LOS data for extreme outliers. However, a sensitivity analysis using log-normal regression produced nearly identical effect estimates, supporting the robustness of our conclusions. Third, despite adjusting for several covariates, we could not account for key confounders such as CAD severity (e.g., SYNTAX score, GRACE score), timing of revascularization, or detailed medication use. These unmeasured factors might have influenced the paradoxical dose-response pattern, particularly if patients with moderate-high comorbidity burden were more likely to receive timely and intensive interventions. If patients with higher CAD severity or earlier revascularization were overrepresented in the moderate-high ECI groups, this could attenuate the observed comorbidity-LOS association, potentially contributing to the paradoxical finding. Fourth, we excluded patients who died in hospital, those discharged against medical advice, and those with unspecified discharge status. The exclusion of deceased patients may introduce selection bias, because death is more frequent among individuals with higher comorbidity burden and truncates LOS, potentially masking the true effect of high ECI on LOS. Likewise, the exclusion of patients discharged against medical advice or with unspecified discharge status could limit the generalizability of our findings to those who complete a medically indicated hospitalization. Nevertheless, the competing risk sensitivity analysis that re-included in-hospital deaths confirmed the robustness of the paradoxical ECI-LOS pattern, with all ECI groups showing significantly prolonged LOS (Supplementary Table S6). The cumulative incidence curves (Supplementary Figure S3) further illustrated that higher ECI scores were associated with a persistently lower probability of discharge alive, mitigating concern about bias from the exclusion of in-hospital deaths. Fifth, the ECI groups were defined by quartiles; using clinically predefined cutoffs or treating ECI as a continuous variable with restricted cubic splines could provide additional insights. Sixth, the exploratory analyses of individual comorbidities, including the interaction between heart failure and treatment category, were not adjusted for multiple comparisons; therefore, these significant associations should be interpreted with caution and require independent validation. Seventh, regarding the high proportion of patients who received TCM in this cohort (83.5%), the treatment modality findings (e.g., differences in LOS among TCM alone, integrated, and WM alone) should be interpreted with caution when extrapolating to settings where WM predominates. Nonetheless, the core ECI-LOS association, which was consistently observed across multiple sensitivity analyses and subgroups, is expected to be robust across treatment settings.

Conclusion

The results of this study indicate that the ECI can serve as an independent and effective assessment tool for predicting LOS in CAD patients. In clinical practice, utilizing the ECI for patient risk assessment facilitates the implementation of individualized management for CAD patients and the optimization of healthcare resource allocation. This holds significant practical importance and application value for promoting refined hospital operational management and enhancing the quality of healthcare services. Future prospective, multi-center studies encompassing institutions with varying treatment modalities are needed to validate the ECI-LOS association, delineate the association of specific comorbidity combinations with LOS, and formally test the treatment intensity hypothesis through detailed data on intervention timing, intensity, and patient-level clinical characteristics such as CAD severity.

Abbreviations

CABG

Coronary Artery Bypass Grafting

CAD

Coronary Artery Disease

CCI

Charlson Comorbidity Index

ECI

Elixhauser Comorbidity Index

ICD-10

International Classification of Diseases, 10th Revision

ICD-9-CM-3

International Classification of Diseases, 9th Revision, Clinical Modification, Volume 3

LOS

Length of Stay

MI

Myocardial Infarction

PCI

Percutaneous Coronary Intervention

SA

Stable Angina

TCM

Traditional Chinese Medicine

VIF

Variance Inflation Factor

WM

Western Medicine

Authors’ contributions

Wenting Yan: Data curation, data analysis, and original draft writing. Shuanggui Tian: Conceptualization, reviewing and editing the original draft. Yong Xiao: Conceptualization. Guodong Xu: Data collection. All authors read and approved the final manuscript.

Funding

This work was supported by Research on Establishing a Knowledge Base for Traditional Chinese Medicine Clinical Diagnosis, Treatment, and Prevention of Common Geriatric Diseases (Number: 2025AFD585).

Data availability

The data generated or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the relevant ethical principles of the Declaration of Helsinki. Ethical approval was granted by the Medical Research Ethics Committee of Hubei University of Chinese Medicine (Approval No. 2026005). This study was a retrospective analysis using anonymized and de-identified existing medical records and did not involve direct contact with participants or any prospective intervention. The requirement for informed consent was waived by the Medical Research Ethics Committee of Hubei University of Chinese Medicine due to the retrospective nature of the study and the use of de-identified data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.