Diagnostic coding combination patterns, CHS-DRG severity stratification, and hospitalization expenditure in pneumonia with coexisting hypertension or diabetes mellitus: a single-center cautionary case study of definitional circularity

Abstract

Background

China’s Healthcare Security Diagnosis-Related Groups (CHS-DRG) framework anchors inpatient payment to a relative weight (RW) set by each hospitalization’s severity stratum. That stratum depends on the complications and comorbidities (CC) and major complications and comorbidities (MCC) identified among recorded secondary diagnoses, so ICD-10 documentation carries direct payment consequences. Evidence on diagnostic coding combination patterns in clinically complex populations, and on the methodological hazards of analyzing an exposure defined by grouper logic, remains limited.

Methods

This single-center retrospective study used medical record front page and CHS-DRG grouping data (locally implemented national grouper, v2.0) for pneumonia hospitalizations discharged from Anhui Chest Hospital in 2025. The hospitalization was the unit of analysis; 131 met pre-specified criteria and were classified by recorded ICD-10 secondary diagnoses into three mutually exclusive coding patterns (basic-disease-only; basic disease with CC; enhanced complication). Because the exposure shares the grouper’s own logic, its correspondence with the severity stratum is treated as definitional and the regressions as exploratory. Reporting follows STROBE and RECORD.

Results

Among 131 hospitalizations (mean age 70.4 ± 10.7 years; 91 of 131 male, 69.5%), 15 (11.45%), 20 (15.27%), and 96 (73.28%) fell into the basic-disease-only, basic disease with CC, and enhanced complication patterns. By construction each pattern mapped one-to-one onto its severity stratum. Total hospitalization expenditure, the only non-definitional outcome, rose across patterns (H = 43.57, P < 0.001): medians 5,938.62, 6,769.91, and 25,501.36 yuan (approximately US$827, US$943, and US$3,552); RW followed a parallel gradient (H = 42.96, P < 0.001). In exploratory regression, length of stay and the count of recorded chronic conditions co-varied with MCC entry (adjusted OR 1.234, 95% CI 1.101 to 1.384, and 1.411, 95% CI 1.144 to 1.741, respectively), both contemporaneous and partially coding-dependent rather than independent predictors.

Conclusions

The structure of recorded ICD-10 diagnoses corresponded to CHS-DRG severity stratification and RW by definition rather than through an independent association, while realized expenditure rose in parallel. The study is best read as a cautionary illustration of definitional circularity: rule-conformant ICD-10 coding remains necessary for accurate stratification and equitable payment, but coding-pattern data alone cannot establish independent clinical or economic effects.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12913-026-15020-0.

Introduction

Diagnosis-Related Groups (DRG) constitute a case-mix payment framework in which hospitalizations are categorized by clinical similarity and resource intensity, and each resulting group is assigned a relative weight (RW) that anchors prospective payment. China has progressively implemented a national DRG system, the China Healthcare Security DRG (CHS-DRG), as a central instrument of inpatient payment reform. A 2023 systematic review and meta-analysis encompassing more than 36 million patients reported that DRG-based prospective payment was associated with shorter length of stay relative to fee-for-service models, although the pooled effects on 30-day readmission and in-hospital mortality did not reach statistical significance [1]. Chinese interrupted time-series and quasi-experimental analyses have similarly shown that DRG implementation is correlated with reductions in per-case cost, length of stay, and selected utilization metrics, accompanied by changes in case-mix index (CMI), patient out-of-pocket share, and cost structure [2–5]. Together, these analyses suggest that the policy effects of DRG reform are multidimensional rather than unidirectional.

CHS-DRG is the national DRG payment standard in China, comprising 26 Major Diagnostic Categories (MDCs) and 409 Adjacent DRG (ADRG) groups, with the technical specifications and grouping scheme published by the National Healthcare Security Administration in 2024 [6]. Within this architecture, severity stratification of an ADRG is determined by the identification of complications and comorbidities (CC) and major complications and comorbidities (MCC) recorded among the secondary diagnoses on the medical record front page. The completeness, granularity, and rule-conformity of ICD-10 coding therefore have direct implications for the assigned subgroup and RW [7]. Empirical evaluations indicate that documentation quality, coder competence, and rule-based identification of omitted diagnoses are central determinants of CHS-DRG accuracy and CMI stability [8–10]. International data show that variation in clinical coding practice is associated with measurable differences in case-level financial classification and may give rise to discharge-level shifts in resource intensity that are not entirely accounted for by changes in clinical complexity [11, 12]. Documentation-focused improvement initiatives have correspondingly been shown to enhance the accuracy with which patient complexity is captured in administrative data [13].

Pneumonia is among the most frequently encountered hospital diagnoses worldwide, and its clinical course is shaped by coexisting chronic disease. Current guidelines on community-acquired pneumonia recognize hypertension, diabetes mellitus, and cardiovascular disease as major comorbidities that influence severity assessment, site-of-care decisions, and outcome [14]. Large cohort studies further indicate that hospitalization for pneumonia is associated with elevated short-term and long-term cardiovascular risk, suggesting that pneumonia interacts with the chronic disease burden in clinically meaningful ways [15]. Within the CHS-DRG architecture, whether the documentation of chronic comorbidities and their attendant target organ damage or acute decompensation is captured in detail (such as hypertensive heart disease, chronic kidney disease, cerebrovascular sequelae, hyperosmolar states, or acute kidney injury) determines the CC/MCC determination path and therefore corresponds to severity stratification, RW, and total expenditure.

Despite this conceptual coherence, granular evidence on how diagnostic coding combination patterns correspond to CHS-DRG severity stratification among hospitalizations for pneumonia with coexisting hypertension and/or diabetes mellitus remains limited, and the methodological consequences of studying an exposure that is itself constructed from grouper logic have received little attention. This distinction is consequential for health services audiences. Payers, hospital administrators, and audit programs increasingly rely on coding-derived severity to set prospective payment, monitor case-mix index, and flag potential upcoding; yet an analysis in which the exposure and the severity outcome are governed by the same rules can readily mistake a definitional identity for an empirical association, with the attendant risk of misdirected audit or payment policy. The objective of the present single-center retrospective study was therefore two-fold. The first was to characterize three pre-specified diagnostic coding combination patterns within this clinical population and to describe their correspondence with CHS-DRG severity stratification, RW, total hospitalization expenditure, and entry into the MCC severity level. The second, and the principal methodological contribution, was to make explicit the definitional circularity that constrains such analyses, so that the one outcome not fixed by the grouper, realized total hospitalization expenditure, can be interpreted on its own terms. The analysis adheres to the structure of the medical record front page and the operative CHS-DRG grouping rules and is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement [16] and the RECORD extension for studies using routinely-collected health data [17]. The correspondence between the exposure (coding pattern) and the primary outcome (severity stratum) is, by construction, definitional, and every component of the analysis is framed and interpreted accordingly.

Methods

Study design and setting

This was a single-center retrospective observational study conducted at Anhui Chest Hospital, a tertiary specialist hospital in Hefei, Anhui Province, China. The hospital operates under the CHS-DRG payment framework and routinely submits standardized medical record front page data to the regional medical insurance authority. The grouping engine in operation at the institution during the study period was the locally implemented national grouper based on the National Healthcare Security Administration CHS-DRG technical specification (the 2024 national specification [6], version 2.0, as adopted by the Anhui provincial medical insurance authority). The CHS-DRG relative weights analyzed here are the regional grouper outputs returned to the institution for settlement rather than hospital-specific or purely national weights. Reporting follows the STROBE statement for cohort studies [16] and the RECORD extension for studies that use routinely-collected health data [17]; the completed STROBE and RECORD checklists are provided as separate checklist files accompanying this submission. The study design and analytic plan were specified prior to data extraction.

Unit of analysis

The unit of analysis was the hospitalization. Where the same patient was admitted on more than one occasion during the study period, each hospitalization was treated as an independent record, consistent with the CHS-DRG payment unit. Within the analytic cohort of 131 hospitalizations, each hospitalization corresponded to a distinct medical record number, and no patient contributed more than one eligible hospitalization during the study period (0 of 131, 0%). Within-patient clustering therefore did not arise in the analytic cohort, and a clustered (random-intercept) sensitivity analysis was accordingly not required. Throughout the manuscript, all counts, proportions, and medians refer to hospitalizations rather than to unique patients.

Data sources

Data were extracted from the institutional Hospital Information System (HIS) and the medical record front page management system. The variables retrieved included demographic information (age, sex), administrative data (admission and discharge dates, length of stay, discharge department, payer type, discharge mode), the principal diagnosis with its ICD-10 code and clinical descriptor, all recorded secondary diagnoses with their ICD-10 codes and descriptors, all recorded procedures, total hospitalization expenditure and its component cost categories, the assigned CHS-DRG group code and descriptor, and the corresponding CHS-DRG RW. Data extraction was performed by two members of the medical record management office, with cross-checking of all front-page fields used in the analysis. The dataset was anonymized at the point of extraction, with the medical record number serving as the only linkage identifier. No direct personal identifiers were retained in the analytic dataset. To characterize the reliability of front-page abstraction, agreement between the two extractors on the fields used in the analysis was assessed on the full analytic cohort. All front-page fields used in the analysis were independently abstracted by the two extractors and reconciled to a single consensus value against the source medical record before the dataset was locked. Because abstraction was reconciled to consensus rather than retained as two independents parallel codings, a formal Cohen kappa was not computed; discrepancies identified during cross-checking were infrequent and were all resolved by consensus prior to analysis.

Study population, time frame, and cohort assembly

Hospitalizations discharged from January 1 to December 31, 2025 with a principal diagnosis of pneumonia were screened. A total of 1,101 hospitalizations met this initial screening criterion. The cohort was then assembled in a stepwise manner with the following pre-specified inclusion and exclusion rules.

Eligibility required all of the following: (1) a principal diagnosis of pneumonia within the ICD-10 code list specified in section “Principal diagnosis (pneumonia)”; (2) age 18 years or older; (3) length of stay of 24 h or longer; (4) complete recording of the front-page fields required for grouping, defined as a non-missing principal diagnosis, fully populated mandatory secondary-diagnosis and procedure fields in accordance with the national medical record front-page specification, and the absence of any internal logical inconsistency flagged during the routine front-page quality audit; and (5) successful entry into a CHS-DRG grouping path with a valid RW.

Pre-specified exclusion criteria, applied in sequence, were: (1) absence of coexisting hypertension and/or diabetes mellitus among the recorded ICD-10 secondary diagnoses; (2) in-hospital death; (3) discharge against medical advice; (4) assignment by the grouper to CHS-DRG group EQY (operation unrelated to the principal diagnosis), reflecting an atypical clinical pathway with incomplete medical insurance settlement information; (5) assignment by the grouper to CHS-DRG group GDTC (special excluded-payment cases); and (6) recorded metastatic malignancy (ICD-10 codes C77, C78, or C79) with a length of stay of one day or fewer, indicating end-stage malignancy with palliative-only treatment, given that such hospitalizations are managed under different clinical and reimbursement pathways.

Of the 1,101 screened hospitalizations, 138 had coexisting hypertension and/or diabetes mellitus identified among recorded secondary diagnoses and survived to live discharge other than against medical advice. Application of the remaining exclusion criteria removed 7 further hospitalizations: 4 grouped to EQY, 2 grouped to GDTC, and 1 with recorded metastatic malignancy and a length of stay of one day. The final analytic cohort therefore comprised 131 hospitalizations. The cohort assembly process is summarized in Fig. 1.

Fig. 1.

STROBE-style flow of cohort assembly. Of 1,101 hospitalizations with a principal diagnosis of pneumonia between January 1 and December 31, 2025, 138 had coexisting hypertension and/or diabetes mellitus identified among recorded secondary ICD-10 codes (I10 to I15 and/or E10 to E14, including subtypes), after exclusion of in-hospital deaths and discharges against medical advice. Application of the remaining exclusion criteria removed 7 hospitalizations: 4 grouped to CHS-DRG group EQY, 2 grouped to GDTC, and 1 with a recorded metastatic malignancy code and a length of stay of one day. The final analytic cohort comprised 131 hospitalizations, classified into three diagnostic coding combination patterns and three CHS-DRG severity strata. The exposure (coding pattern) is rule-based on recorded ICD-10 codes and overlaps definitionally with the CHS-DRG severity stratum. CC, complication or comorbidity; MCC, major complication or comorbidity. Among the screened hospitalizations, 38 ended in in-hospital death and 44 were discharged against medical advice; the remaining excluded hospitalizations either lacked coexisting hypertension or diabetes mellitus among the recorded secondary diagnoses or did not enter an analyzable severity-stratified grouping path. All counts denote case-level hospitalizations rather than unique patients; the analytic cohort comprised 131 unique patients (Section “Unit of analysis”)

Among the 1,101 screened hospitalizations, 38 ended in in-hospital death and 44 were discharged against medical advice; the remaining excluded hospitalizations either did not carry coexisting hypertension or diabetes mellitus among the recorded secondary diagnoses or did not enter an analyzable severity-stratified CHS-DRG grouping path. The in-hospital-death and against-medical-advice counts are displayed in Fig. 1. In-hospital deaths and discharges against medical advice were excluded a priori because both interrupt the documentation and grouping process that the present analysis examines: a hospitalization that ends in death or premature self-discharge frequently carries an incomplete front-page record and a non-representative resource trajectory, which would confound the description of the coding-to-severity correspondence. We acknowledge that excluding in-hospital deaths introduces the potential for survivor bias, since fatal hospitalizations are plausibly enriched for high-severity coding. To gauge the sensitivity of the principal findings to this decision, a sensitivity analysis was undertaken in which the 36 in-hospital deaths with codable front-page records were reintroduced and the description of the expenditure and RW gradients was repeated (Section “Cohort assembly and flow”).

Variables and definitions

Principal diagnosis (pneumonia)

The principal diagnosis was defined as the ICD-10 code recorded as the principal diagnosis on the medical record front page. The pneumonia code list comprised: J12 (viral pneumonia, not elsewhere classified), J13 (pneumonia due to Streptococcus pneumoniae), J14 (pneumonia due to Haemophilus influenzae), J15 (bacterial pneumonia, not elsewhere classified), J16 (pneumonia due to other infectious organisms, not elsewhere classified), J17 (pneumonia in diseases classified elsewhere), J18 (pneumonia, unspecified organism, including J18.0 to J18.9 and the institutional sub-codes J18.900 to J18.903 for severe pneumonia), J69 (pneumonitis due to solids and liquids), and J70 (respiratory conditions due to other external agents, including radiation pneumonitis at J70.001). Hospitalizations whose principal diagnosis fell outside this list were not eligible.

Coexisting hypertension and diabetes mellitus

Coexisting chronic disease was defined by the presence of relevant ICD-10 codes among the recorded secondary diagnoses. Hypertension included essential hypertension (I10 and its institutional sub-codes such as I10.x00 × 002 to I10.x00 × 032, reflecting recorded hypertension grade and risk stratification), hypertensive heart disease (I11), hypertensive chronic kidney disease (I12), hypertensive heart and chronic kidney disease (I13), and secondary hypertension (I15). Diabetes mellitus comprised type 1 diabetes mellitus (E10), type 2 diabetes mellitus (E11), other specified diabetes mellitus (E13), and unspecified diabetes mellitus (E14), each with their full set of complication-coded sub-codes. Hospitalizations were further categorized as having hypertension only, diabetes mellitus only, or hypertension combined with diabetes mellitus.

Target organ damage and acute severe complications

Target organ damage and acute severe complications relevant to hypertension or diabetes mellitus were identified from secondary ICD-10 codes. Target organ damage codes included hypertensive heart disease and cardiac insufficiency (I11, I50), chronic kidney disease (N18, N19), cerebrovascular sequelae (I63, I69), and hypertensive retinopathy (H35.0). Acute severe complications were defined as the recorded presence of any of the following ICD-10 codes: type I or type II respiratory failure (J96.900 × 002, J96.900 × 003, J96.9), acute kidney injury (N17, N19.x00 × 002), hyperosmolar hyperglycemic state (E11.0, E13.0, E14.0), diabetic ketoacidosis (E10.1, E11.1, E13.1), heart failure (I50), respiratory or metabolic acidosis (E87.2, E87.3), severe electrolyte disturbance (E87.5, E87.8), or shock (R57). The presence variable was binary (any code recorded versus none).

Number of recorded chronic conditions

The number of recorded chronic conditions was defined as the count of distinct ICD-10 codes recorded among the secondary diagnoses on the medical record front page. This variable is therefore coding-dependent: it captures the number of conditions that were both clinically present and documented at the level of an ICD-10 code, and it is not a clinically validated comorbidity index. To answer this point directly, the count comprised all distinct ICD-10 codes recorded among the secondary diagnoses, encompassing chronic comorbidities, target organ damage, and any recorded acute complication codes; it was not restricted to a pre-specified chronic subset. The recorded presence of acute severe complications was, in addition, captured separately as a distinct binary indicator (Section “Target organ damage and acute severe complications”) for use as a covariate, so that the count and the binary acute indicator, although both derived from the secondary-diagnosis field, serve different analytic roles. This count is therefore best understood as a documentation-volume measure rather than a clinically validated comorbidity index.

CHS-DRG severity stratum

The CHS-DRG severity stratum was determined by the operative CHS-DRG grouper. Each ADRG that admits severity stratification is partitioned into three mutually exclusive subgroups: no CC/MCC, CC, and MCC. The grouper identifies CC and MCC by mapping recorded secondary diagnoses against the CHS-DRG CC/MCC reference list [6]. Hospitalizations grouped to ADRGs that do not admit severity stratification (such as ES36 for severe pneumonia, which carries an integrated severity tier) were classified, for the purpose of the present analysis, as MCC-tier hospitalizations consistent with the published CHS-DRG technical specification [6]. The grouping scheme corresponded to the 2024 national CHS-DRG technical specification; the precise grouper version applied at the study site is stated in Section “Study design and setting”. The handling of ES36 (severe pneumonia) and of the ADRG-specific CC and MCC reference lists follows the grouping rules and the CC/MCC exclusion tables of that specification [6].

CHS-DRG relative weight and total hospitalization expenditure

The CHS-DRG RW was extracted directly from the grouper output and represents the relative resource intensity assigned to the case. Total hospitalization expenditure was extracted from the medical record front page, expressed in Chinese yuan, and reflects the total billed amount, inclusive of insurance-covered and out-of-pocket components. Because all hospitalizations were discharged within a single calendar year (January to December 2025), no intra-period adjustment for price inflation was applied; reported amounts represent the settled expenditure recorded at discharge. To aid interpretation for an international readership, yuan figures are accompanied at their first occurrence in the Results by approximate US dollar equivalents converted at the 2025 average exchange rate (US$1 is approximately equal to 7.18 yuan).

Exposure: Diagnostic coding combination pattern

The exposure of interest was the diagnostic coding combination pattern, derived from the structure of recorded ICD-10 codes for hypertension and/or diabetes mellitus and their associated complications. Three mutually exclusive patterns were pre-specified prior to data analysis:

1. Basic-disease-only coding pattern. Only the base diagnosis code for hypertension and/or diabetes mellitus was recorded among the secondary diagnoses, with no associated target organ damage or acute complication code. The hospitalization was not identified as carrying CC or MCC by the grouper.

2. Basic disease with CC coding pattern. The chronic disease code was recorded together with at least one secondary diagnosis identified by the grouper as a CC, but the recorded codes did not meet the MCC threshold.

3. Enhanced complication coding pattern. In addition to the chronic disease code, target organ damage codes (such as cardiac insufficiency, chronic renal insufficiency, or cerebrovascular sequelae) or acute metabolic or organ-specific complication codes (such as hyperosmolar hyperglycemic state or acute kidney injury) were recorded, and the hospitalization was identified as carrying MCC by the grouper, including assignment to ADRGs with an integrated MCC tier.

This exposure construction is, by design, rule-based and therefore overlaps definitionally with the CHS-DRG severity stratum. The classification step in this study reflects this overlap rather than an empirical association. The substantive analyses of the present study are the description of the gradient in RW and total expenditure across patterns (with formal Kruskal-Wallis tests and Bonferroni-adjusted pairwise comparisons; Section “CHS-DRG relative weight and total hospitalization expenditure across coding patterns”) and the multivariable logistic regression of MCC entry on covariates that are not definitionally tied to the exposure (Section “Multivariable associations with entry into the MCC stratum”). The interpretation of these analyses, including the contemporaneous and partially coding-dependent nature of length of stay and the count of recorded chronic conditions, is addressed in the Discussion.

Outcomes

Three outcomes were specified: (1) the CHS-DRG severity stratum (no CC/MCC, CC, or MCC); (2) the CHS-DRG RW; and (3) the total hospitalization expenditure. The dependent variable in the regression analyses was entry into the MCC severity stratum (yes = 1, no = 0).

Covariates

Covariates were extracted from the medical record front page and selected for clinical relevance to severity stratification: age (years, continuous), sex (male / female), length of stay (days, continuous), the number of recorded chronic conditions (count, continuous), the recorded presence of acute severe complications (binary; defined in Section “Target organ damage and acute severe complications”), and the combination of hypertension and diabetes mellitus versus either alone (binary). It is acknowledged in advance that length of stay is a contemporaneous resource and severity marker that accumulates during the same hospitalization in which the codes are recorded, and that the count of recorded chronic conditions is partially coding-dependent. These properties are addressed in the Discussion.

Statistical analysis

Analyses were performed with SPSS version 26.0 (IBM Corp., Armonk, NY, USA) and corroborated in Python 3.12 with scipy version 1.13.1 for the Kruskal-Wallis and Mann-Whitney pairwise tests. The distribution of continuous variables was first examined visually and tested with the Shapiro-Wilk test. Continuous variables with an approximately normal distribution were summarized as mean ± standard deviation (SD) and compared with one-way analysis of variance (ANOVA). Continuous variables with non-normal distributions, including length of stay, the number of recorded chronic conditions, CHS-DRG RW, and total hospitalization expenditure, were summarized as median (P25, P75) and compared with the Kruskal-Wallis H test. Pairwise post-hoc comparisons among the three coding patterns were performed with the Mann-Whitney U test, with Bonferroni correction for three comparisons. Categorical variables were summarized as frequencies (percentages). For Table 1, the Pearson chi-squared test or, where indicated by expected cell counts, the Fisher exact test was used. For the cross-tabulation of coding pattern by CHS-DRG severity stratum (Table 2), an inferential test was not performed, because the relationship between exposure and outcome is rule-based and not an empirical discovery (Section “Exposure: Diagnostic coding combination pattern”); the cross-tabulation is presented descriptively.

Table 1.

Baseline characteristics of hospitalizations across diagnostic coding combination patterns (n = 131)

Variable	Basic-disease-only (n = 15)	Basic disease with CC (n = 20)	Enhanced complication (n = 96)	F / χ² / H	P value
Age, years (mean ± SD)	64.47 ± 9.13	68.05 ± 7.47	72.08 ± 11.44	3.952	0.022
Sex, n (%)				3.049	0.218
Male	11 (73.33)	17 (85.00)	63 (65.63)
Female	4 (26.67)	3 (15.00)	33 (34.38)
Length of stay, days, median (P25, P75)	7.00 (5.00, 9.00)	8.00 (5.00, 11.25)	17.00 (11.75, 24.00)	36.822	< 0.001
Number of recorded chronic conditions, median (P25, P75)	4.00 (3.00, 5.00)	6.00 (4.75, 7.50)	10.00 (8.00, 12.00)	41.633	< 0.001
Records with acute severe complication codes, n (%)	2 (13.33)	3 (15.00)	67 (69.79)	31.933	< 0.001

Note. Continuous variables are reported as mean ± SD or median (P25, P75) as appropriate. Categorical variables are reported as n (%). Continuous variables were compared using one-way ANOVA (F) or the Kruskal-Wallis H test; categorical variables were compared using the Pearson chi-squared test (χ²), with the Fisher exact test substituted where any expected cell count was below 5. The baseline characteristics constitute a descriptive family of tests; P values were additionally examined under Benjamini-Hochberg false discovery rate control as a sensitivity check, and no characteristic changed in interpretation. CC, complication or comorbidity

Table 2.

Cross-tabulation of diagnostic coding combination pattern and CHS-DRG severity stratum (n = 131)

Coding pattern	No CC/MCC n (%)	CC n (%)	MCC n (%)	Total n
Basic-disease-only	15 (100.00)	0 (0.00)	0 (0.00)	15
Basic disease with CC	0 (0.00)	20 (100.00)	0 (0.00)	20
Enhanced complication	0 (0.00)	0 (0.00)	96 (100.00)	96
Total	15 (11.45)	20 (15.27)	96 (73.28)	131

Note. No inferential test is reported for this cross-tabulation because the relationship between exposure and outcome is rule-based (Section “Exposure: Diagnostic coding combination pattern”) and not an empirical discovery. The cross-tabulation is presented descriptively. CHS-DRG, China Healthcare Security Diagnosis-Related Groups; CC, complication or comorbidity; MCC, major complication or comorbidity

Univariable binary logistic regression was performed with entry into the MCC stratum as the dependent variable. Variables of interest were age, sex, length of stay, number of recorded chronic conditions, recorded presence of acute severe complications, and the combined hypertension–diabetes status. Variables that reached statistical significance in univariable analysis (P < 0.05) were entered into a multivariable logistic regression model. Effect estimates are reported as odds ratios (OR) with 95% confidence intervals (CI). The overall multivariable model was assessed using the likelihood-ratio test and the Hosmer-Lemeshow goodness-of-fit test. A two-sided P value of less than 0.05 was considered statistically significant.

The multivariable logistic regression was specified, and is reported, as an exploratory and descriptive analysis of correlates of MCC entry rather than as a confirmatory or predictive model. The selection of covariates on the basis of univariable significance is recognized as a hypothesis-generating strategy that does not establish independent or causal effects. With 35 hospitalizations outside the MCC stratum constituting the limiting outcome events and four covariates entered, the events-per-variable ratio was approximately 8.75, which is below the conventional threshold of 10; the model is therefore susceptible to small-sample bias and to unstable estimation. To address these concerns, collinearity among the covariates was examined using variance inflation factors and the condition number of the design matrix; the maximum-likelihood estimates were accompanied by Firth penalized-likelihood estimates as a sensitivity analysis for sparse-data and separation effects; and 95% confidence intervals for the adjusted odds ratios were additionally derived from 2,000 bootstrap resamples. Model calibration was summarized by the Hosmer-Lemeshow statistic, the Brier score, and the bootstrap-corrected area under the receiver operating characteristic curve.

For the between-pattern comparisons of CHS-DRG RW and of total hospitalization expenditure, the magnitude of the differences was quantified in addition to the omnibus and pairwise tests. Hodges-Lehmann median differences, with 95% confidence intervals derived from 2,000 bootstrap resamples, were computed for each pairwise contrast, and the rank-biserial correlation was reported as a distribution-free measure of effect size.

Multiplicity was managed by pre-specifying the families of tests and the correction applied within each family. The pairwise post-hoc comparisons of RW and of total expenditure each constituted a family of three comparisons controlled by the Bonferroni method, as described above. The baseline characteristics in Table 1 were treated as a descriptive family, and the Benjamini-Hochberg false discovery rate procedure was applied within that family as a sensitivity check; any characteristic whose interpretation changed under false-discovery-rate control is identified in the footnote to Table 1. The univariable regression screen was treated as a hypothesis-generating step and was not subjected to family-wise correction, consistent with its exploratory purpose.

Hospitalizations with logical inconsistency or substantial missing data on the medical record front page were excluded a priori during cohort construction (Section “Study population, time frame, and cohort assembly”). For the 131 retained hospitalizations, all variables included in the analysis were complete on audit by the medical record management office; no imputation was applied.

Ethical considerations

This study was reviewed and approved by the Ethics Committee of Anhui Chest Hospital prior to data extraction. Owing to the retrospective study design and the exclusive use of de-identified administrative data, the same Ethics Committee of Anhui Chest Hospital that approved the study also formally waived the requirement for individual informed consent to participate. All methods were performed in accordance with the principles of the Declaration of Helsinki.

Results

Cohort assembly and flow

Of 1,101 hospitalizations with a principal diagnosis of pneumonia recorded between January 1 and December 31, 2025, 138 had coexisting hypertension and/or diabetes mellitus identified among the recorded secondary ICD-10 codes after exclusion of in-hospital deaths and discharges against medical advice. Application of the remaining exclusion criteria removed 7 additional hospitalizations: 4 grouped by the CHS-DRG grouper to EQY (operation unrelated to the principal diagnosis), 2 grouped to GDTC (special excluded-payment cases), and 1 with a recorded metastatic malignancy code (C77) and a length of stay of one day, indicating end-stage malignancy with palliative-only treatment. The final analytic cohort comprised 131 hospitalizations (Fig. 1). The discharge departments most frequently represented were the intensive care/RICU service, the respiratory medicine wards, and the interventional pulmonology service, consistent with the patient mix of a tertiary chest hospital.

Baseline characteristics by coding pattern

Of the 131 hospitalizations, 15 (11.45%) were classified into the basic-disease-only coding pattern, 20 (15.27%) into the basic disease with CC coding pattern, and 96 (73.28%) into the enhanced complication coding pattern. The mean age across the cohort was 70.4 ± 10.7 years; 91 hospitalizations (69.5%) were of male patients. Across the three coding patterns, statistically significant differences were observed for age (P = 0.022), length of stay (P < 0.001), the number of recorded chronic conditions (P < 0.001), and the recorded proportion of hospitalizations with acute severe complication codes (P < 0.001). The distribution of sex did not differ significantly across the coding patterns (P = 0.218). Hospitalizations in the enhanced complication coding pattern were of older patients, had longer stays, and carried more recorded chronic conditions and acute complications than hospitalizations in the other two patterns. Detailed group-wise characteristics are presented in Table 1. In the sensitivity analysis that reintroduced the 36 in-hospital deaths with codable front-page records (Section “Study population, time frame, and cohort assembly”), the direction and statistical significance of the gradients in total hospitalization expenditure and CHS-DRG RW across the three coding patterns were unchanged (expenditure, Kruskal-Wallis H = 50.81, P < 0.001; RW, H = 52.70, P < 0.001), indicating that the principal descriptive findings were robust to the exclusion of fatal hospitalizations.

Mapping of coding pattern to chs-drg severity stratum

Across the analytic cohort, 15 hospitalizations (11.45%) were assigned to the no CC/MCC stratum, 20 hospitalizations (15.27%) to the CC stratum, and 96 hospitalizations (73.28%) to the MCC stratum. Each coding pattern mapped one-to-one to a CHS-DRG severity stratum: all basic-disease-only hospitalizations entered the no CC/MCC stratum, all basic disease with CC hospitalizations entered the CC stratum, and all enhanced complication hospitalizations entered the MCC stratum (Table 2). This mapping reflects the rule-based construction of the exposure (Section “Exposure: Diagnostic coding combination pattern”) and is presented descriptively rather than as an inferential association. The substantive findings of this analysis are presented in Section “CHS-DRG relative weight and total hospitalization expenditure across coding patterns” to “ Multivariable associations with entry into the MCC stratum”.

CHS-DRG relative weight and total hospitalization expenditure across coding patterns

Median CHS-DRG RW differed significantly across the three coding patterns (Kruskal-Wallis H = 42.96, P < 0.001), rising from 0.3056 (P25, P75: 0.3056, 0.9173) in the basic-disease-only pattern, to 0.5455 (0.5455, 0.6954) in the basic disease with CC pattern, and to 1.3742 (1.0420, 2.1228) in the enhanced complication pattern. In Bonferroni-adjusted pairwise Mann-Whitney comparisons, all three pairwise differences in RW reached statistical significance (basic-disease-only versus basic disease with CC, P_adj = 0.032; basic-disease-only versus enhanced complication, P_adj < 0.001; basic disease with CC versus enhanced complication, P_adj < 0.001). The coincidence of the lower quartile and the median at 0.3056 in the basic-disease-only pattern is not a distributional anomaly; it reflects the discrete structure of the RW scale, in which the small number of low-severity grouping paths available to that pattern map onto a limited set of fixed weight values, so that a single value occupies both the 25th percentile and the median. The corresponding Hodges-Lehmann median differences and the rank-biserial correlations for each pairwise contrast, with bootstrap 95% confidence intervals, are reported in Table 3a. The pairwise Hodges-Lehmann median difference was 0.2399 (95% CI 0.228 to 0.240; rank-biserial correlation 0.49) for the basic disease with CC versus basic-disease-only contrast, 0.4965 (0.497 to 0.829; rank-biserial 0.68) for the enhanced complication versus basic disease with CC contrast, and 0.7364 (0.457 to 1.069; rank-biserial 0.80) for the enhanced complication versus basic-disease-only contrast; the narrow intervals for the lower-tier contrasts reflect the discrete RW scale described above.

Table 4.

b. Distribution of total hospitalization expenditure (yuan) across diagnostic coding combination patterns

Coding pattern	n	Total expenditure (yuan) median (P25, P75)	P (vs. ref.)	Pairwise (Bonferroni P_adj)
Basic-disease-only (ref.)	15	5,938.62 (3,589.16, 8,056.41)	—	1 vs. 2: P_adj = 0.634
Basic disease with CC	20	6,769.91 (4,620.89, 12,727.75)	0.634	1 vs. 3: P_adj < 0.001
Enhanced complication	96	25,501.36 (12,243.33, 57,969.54)	< 0.001	2 vs. 3: P_adj < 0.001
Overall Kruskal-Wallis H = 43.57, P < 0.001

Note. Pattern numbering in pairwise comparisons: 1 = basic-disease-only; 2 = basic disease with CC; 3 = enhanced complication. The Kruskal-Wallis H test was used for the overall comparison; pairwise post-hoc comparisons used the Mann-Whitney U test with Bonferroni correction for three comparisons. Median expenditures converted to US dollars at the 2025 average exchange rate (US827, US$943, and US$3,552 for the basic-disease-only, basic disease with CC, and enhanced complication patterns, respectively. Hodges-Lehmann median differences, bootstrap 95% confidence intervals, and rank-biserial correlations for each pairwise contrast are: basic disease with CC versus basic-disease-only, 1,415 yuan (95% CI − 873 to 5,105), rank-biserial 0.25; enhanced complication versus basic disease with CC, 15,882 yuan (9,115 to 24,498), rank-biserial 0.67; and enhanced complication versus basic-disease-only, 19,630 yuan (12,422 to 31,569), rank-biserial 0.85. These between-pattern differences are unadjusted and descriptive and should not be interpreted as the causal effect of a coding pattern on expenditure

Table 3.

a. Distribution of CHS-DRG relative weight across diagnostic coding combination patterns

Coding pattern	n	CHS-DRG RW median (P25, P75)	P (vs. ref.)	Pairwise (Bonferroni P_adj)
Basic-disease-only (ref.)	15	0.3056 (0.3056, 0.9173)	—	1 vs. 2: P_adj = 0.032
Basic disease with CC	20	0.5455 (0.5455, 0.6954)	0.032	1 vs. 3: P_adj < 0.001
Enhanced complication	96	1.3742 (1.0420, 2.1228)	< 0.001	2 vs. 3: P_adj < 0.001
Overall Kruskal-Wallis H = 42.96, P < 0.001

Note. Pattern numbering in pairwise comparisons: 1 = basic-disease-only; 2 = basic disease with CC; 3 = enhanced complication. The Kruskal-Wallis H test was used for the overall comparison; pairwise post-hoc comparisons used the Mann-Whitney U test with Bonferroni correction for three comparisons. RW, relative weight. Hodges-Lehmann median differences, bootstrap 95% confidence intervals, and rank-biserial correlations for each pairwise contrast are: basic disease with CC versus basic-disease-only, 0.2399 (95% CI 0.228 to 0.240), rank-biserial 0.49; enhanced complication versus basic disease with CC, 0.4965 (0.497 to 0.829), rank-biserial 0.68; and enhanced complication versus basic-disease-only, 0.7364 (0.457 to 1.069), rank-biserial 0.80. In the basic-disease-only pattern the lower quartile and the median coincide at 0.3056; this reflects the discrete structure of the RW scale, in which the limited set of low-severity grouping paths available to that pattern maps onto a small number of fixed weight values, and is not a distributional anomaly

Median total hospitalization expenditure differed similarly across the three coding patterns (Kruskal-Wallis H = 43.57, P < 0.001), increasing from 5,938.62 yuan (P25, P75: 3,589.16, 8,056.41) in the basic-disease-only pattern, to 6,769.91 yuan (4,620.89, 12,727.75) in the basic disease with CC pattern, and to 25,501.36 yuan (12,243.33, 57,969.54) in the enhanced complication pattern. In Bonferroni-adjusted pairwise comparisons, total expenditure in the enhanced complication pattern was significantly higher than in both the basic-disease-only pattern (P_adj < 0.001) and the basic disease with CC pattern (P_adj < 0.001). The pairwise difference between the basic-disease-only and basic disease with CC patterns did not reach statistical significance after Bonferroni adjustment (P_adj = 0.634), suggesting that the gradient in expenditure across patterns is driven primarily by the enhanced complication pattern. The pattern is summarized numerically in Table 3 and graphically in Fig. 2. Hodges-Lehmann median differences with bootstrap 95% confidence intervals and rank-biserial correlations for the pairwise expenditure contrasts are presented in Table 3b. The median expenditure difference was 1,415 yuan (95% CI − 873 to 5,105; rank-biserial 0.25) for the basic disease with CC versus basic-disease-only contrast, an interval that includes zero and is consistent with the non-significant Bonferroni-adjusted comparison for that pair, 15,882 yuan (9,115 to 24,498; rank-biserial 0.67) for the enhanced complication versus basic disease with CC contrast, and 19,630 yuan (12,422 to 31,569; rank-biserial 0.85) for the enhanced complication versus basic-disease-only contrast. These between-pattern differences in expenditure are unadjusted and descriptive; they should not be read as the causal effect of a coding pattern on cost. The gradient is equally compatible with underlying differences in clinical severity, care processes, length of stay, and the completeness of front-page documentation, all of which co-vary with the recorded coding pattern.

Fig. 2.

Distribution of CHS-DRG relative weight (panel a) and total hospitalization expenditure (panel b, log scale) across the three diagnostic coding combination patterns. Box-and-whisker elements summarize the median, interquartile range, and full observed range; overlaid points are observed case-level values for the 131 hospitalizations in the analytic cohort. The Kruskal-Wallis H statistic and overall P value are shown for each panel

Univariable associations with entry into the MCC stratum

Univariable logistic regression with entry into the MCC stratum as the dependent variable identified four covariates significantly associated with higher odds of MCC entry: age (OR per year 1.048, 95% CI 1.011 to 1.087, P = 0.012); length of stay (OR per day 1.261, 95% CI 1.144 to 1.391, P < 0.001); number of recorded chronic conditions (OR per condition 1.617, 95% CI 1.351 to 1.936, P < 0.001); and recorded presence of acute severe complications (OR 10.222, 95% CI 3.992 to 26.173, P < 0.001). Sex (OR for male versus female 0.477, 95% CI 0.188 to 1.209, P = 0.119) and the combination of hypertension and diabetes mellitus (OR 1.100, 95% CI 0.488 to 2.478, P = 0.819) were not significantly associated with MCC entry. Univariable estimates are presented in Table 4.

Table 5.

Univariable logistic regression of factors associated with entry into the MCC severity stratum

Variable	β	SE	Wald χ²	OR	95% CI	P value
Age (per year)	0.047	0.019	6.368	1.048	1.011 to 1.087	0.012
Sex (male = 1, female = 0)	-0.740	0.474	2.434	0.477	0.188 to 1.209	0.119
Length of stay (per day)	0.232	0.050	21.656	1.261	1.144 to 1.391	< 0.001
Number of recorded chronic conditions (per condition)	0.481	0.092	27.375	1.617	1.351 to 1.936	< 0.001
Recorded acute severe complication code (yes = 1, no = 0)	2.325	0.480	23.484	10.222	3.992 to 26.173	< 0.001
Hypertension and diabetes (yes = 1, no = 0)	0.095	0.414	0.053	1.100	0.488 to 2.478	0.819

Note. The dependent variable is entry into the MCC severity stratum (yes = 1, no = 0). β, regression coefficient; SE, standard error; OR, odds ratio; CI, confidence interval. Univariable estimates are presented as a hypothesis-generating screen; selection of covariates for the multivariable model on the basis of univariable significance is an exploratory strategy and does not establish independent or causal associations

Multivariable associations with entry into the MCC stratum

The four univariable-significant covariates (age, length of stay, number of recorded chronic conditions, and recorded presence of acute severe complications) were entered into a multivariable logistic regression model. Length of stay (adjusted OR per day 1.234, 95% CI 1.101 to 1.384, P < 0.001) and the number of recorded chronic conditions (adjusted OR per condition 1.411, 95% CI 1.144 to 1.741, P = 0.001) were associated with entry into the MCC stratum after mutual adjustment. Age (adjusted OR 0.998, 95% CI 0.949 to 1.049, P = 0.934) and the recorded presence of acute severe complications (adjusted OR 2.507, 95% CI 0.696 to 9.034, P = 0.160) were no longer significantly associated with MCC entry after adjustment. The overall multivariable model was statistically significant (likelihood-ratio chi-squared = 72.500, P < 0.001). Consistent with the events-per-variable ratio of approximately 8.75 (Section “Statistical analysis”), these estimates should be regarded as exploratory. The Hosmer-Lemeshow goodness-of-fit test was non-significant (chi-squared = 5.09, df = 8, P = 0.75), the Brier score was 0.09, and the bootstrap-corrected area under the receiver operating characteristic curve was 0.92. Variance inflation factors for the four covariates ranged from 1.16 to 1.54 and the condition number of the design matrix was 2.13, indicating only modest collinearity; the attenuation of the acute-complication odds ratio after adjustment is therefore better understood as shared explanatory variance with length of stay and the recorded chronic-condition count than as severe multicollinearity. The Firth penalized-likelihood estimates were materially consistent with the maximum-likelihood estimates for length of stay and the number of recorded chronic conditions, and bootstrap confidence intervals for the adjusted odds ratios are reported in Table 5. The multivariable estimates are presented in Table 5 and visualized in Fig. 3. Both covariates that retained statistical significance are contemporaneous with the hospitalization and partially coding-dependent: length of stay accumulates during the same hospitalization in which the codes are recorded, and the count of recorded chronic conditions is itself a function of front-page documentation. These observations are central to the interpretation of the multivariable analysis (Section “Discussion”) and are not consistent with a pre-admission predictive use of the model.

Table 6.

Multivariable logistic regression of factors associated with entry into the MCC severity stratum

Variable	β	SE	Wald χ²	Adjusted OR	95% CI	P value
Length of stay (per day)	0.210	0.058	13.005	1.234	1.101 to 1.384	< 0.001
Number of recorded chronic conditions (per condition)	0.344	0.107	10.337	1.411	1.144 to 1.741	0.001
Age (per year)	-0.002	0.026	0.007	0.998	0.949 to 1.049	0.934
Recorded acute severe complication code (yes = 1, no = 0)	0.919	0.654	1.976	2.507	0.696 to 9.034	0.160

Note. The dependent variable is entry into the MCC severity stratum (yes = 1, no = 0). Variables entered in the multivariable model were those that reached statistical significance in the univariable analysis. Overall model likelihood-ratio chi-squared = 72.500, P < 0.001. Length of stay and the number of recorded chronic conditions are contemporaneous and partially coding-dependent variables (Section “Exposure: Diagnostic coding combination pattern”); the model is descriptive of correlates of MCC entry rather than predictive. The events-per-variable ratio was approximately 8.75 (35 non-MCC events, 4 covariates), below the conventional threshold of 10, and the estimates are correspondingly exploratory. Model diagnostics: Hosmer-Lemeshow chi-squared = 5.09 (df = 8), P = 0.75; Brier score = 0.09; bootstrap-corrected area under the receiver operating characteristic curve = 0.92; variance inflation factors 1.16 to 1.54; condition number = 2.13. Bootstrap 95% confidence intervals for the adjusted odds ratios were 1.12 to 1.57 for length of stay and 1.14 to 1.86 for the number of recorded chronic conditions; the Firth penalized-likelihood adjusted odds ratios (1.22 and 1.32, respectively) were materially consistent with the maximum-likelihood estimates. OR, odds ratio; CI, confidence interval

Fig. 3.

Forest plot of multivariable logistic regression for entry into the MCC severity level. Statistically significant covariates (length of stay, number of recorded chronic conditions) are shown in dark blue; non-significant covariates are shown in gray. Length of stay and the number of recorded chronic conditions are contemporaneous and partially coding-dependent variables; the model should be interpreted as descriptive of correlates of MCC entry rather than as a pre-admission predictive model

Discussion

In this single-center retrospective observational study of 131 hospitalizations for pneumonia with coexisting hypertension and/or diabetes mellitus, more granular recording of chronic comorbidities and acute complications corresponded directly to higher CHS-DRG severity stratification, higher CHS-DRG RW, and higher total hospitalization expenditure. The mapping between coding pattern and CHS-DRG severity stratum is, by construction, rule-based: the exposure was defined using the same ICD-10 code logic that the grouper applies to determine CC and MCC, and the corresponding cross-tabulation (Table 2) reflects this definitional overlap rather than an empirical discovery. The substantive contributions of the present analysis are therefore three-fold: a description of the gradient in CHS-DRG RW and total hospitalization expenditure across patterns, the identification of contemporaneous correlates of MCC entry that are not definitionally tied to the exposure, and a methodological articulation of the circularity inherent in coding-pattern analyses of CHS-DRG severity.

The gradient in CHS-DRG RW and total expenditure observed across the three coding patterns was confirmed by Kruskal-Wallis tests with Bonferroni-adjusted pairwise comparisons (Section “CHS-DRG relative weight and total hospitalization expenditure across coding patterns”). For RW, all three pairwise differences reached statistical significance, including the difference between the basic-disease-only and basic disease with CC patterns. For total expenditure, the gradient was driven primarily by the enhanced complication pattern, while the difference between the basic-disease-only and basic disease with CC patterns did not reach significance after Bonferroni correction. This dissociation suggests that the addition of CC-tier codes alone has a smaller effect on realized expenditure than the addition of MCC-tier codes, which corresponds to the more pronounced step in CHS-DRG RW between the CC and MCC subgroups. Recent Chinese policy evaluations have similarly reported that DRG implementation is correlated with shifts in CMI, cost structure, and patient out-of-pocket share rather than with uniform reductions in per-case expenditure [2, 4, 5, 7].

In multivariable analysis of MCC entry, length of stay and the number of recorded chronic conditions retained statistical significance after mutual adjustment for age and the recorded presence of acute severe complications. Both of these covariates are contemporaneous with the hospitalization rather than pre-admission risk markers, and both are partially coding-dependent. Length of stay accumulates during the same hospitalization in which the secondary diagnoses are recorded, and the appearance of additional ICD-10 codes (and, in turn, eligibility for CC or MCC tiers) is plausibly downstream of an evolving clinical course rather than determinative of it. Reverse causation between length of stay and recorded comorbidity burden cannot be excluded in a retrospective design [9, 10]. The number of recorded chronic conditions, although correlated with the underlying clinical comorbidity burden of the hospitalization [14, 15], is itself a documentation artifact: it captures conditions that were both clinically present and documented at the level of an ICD-10 code on the medical record front page. The multivariable analysis should therefore be read as a description of contemporaneous and coding-dependent correlates of MCC entry, not as a predictive model.

The recorded presence of acute severe complications, although strongly associated with MCC entry in univariable analysis (OR 10.222), did not retain statistical significance after adjustment. With 96 hospitalizations at the MCC level and four covariates in the multivariable model, the analysis is sensitive to collinearity between length of stay, the number of recorded chronic conditions, and the binary acute-complication indicator. The apparent attenuation should be interpreted as consistent with such collinearity rather than as evidence of an absent association. Larger multicenter cohorts would be required to disentangle the partial effects of these intercorrelated documentation-derived variables.

The methodological articulation of circularity is the principal contribution of the present study to the existing CHS-DRG literature. Empirical evaluations have repeatedly emphasized that documentation quality, coder competence, and rule-based identification of omitted diagnoses are central determinants of CHS-DRG accuracy and CMI stability [8–10], and international data show that variation in clinical coding practice can give rise to discharge-level shifts in administrative classification [11, 12]. Documentation-improvement programs in other settings have been associated with measurable gains in the capture of patient complexity, without requiring changes in clinical practice [13]. The gradient observed in the present analysis is therefore consistent with the operational logic of the grouper: the documentation of clinically meaningful complications and target organ damage, when undertaken in accordance with ICD-10 rules, corresponds to substantial differences in administrative classification and payment-relevant indicators. The findings do not constitute evidence of inappropriate coding behavior. Rather, they illustrate that, within the CHS-DRG framework, the structure of recorded ICD-10 codes is a primary mechanical determinant of severity assignment and RW.

The practical implications of these findings are several. First, standardized documentation of chronic comorbidity, target organ damage, and acute complications, in conformance with ICD-10 rules and the CHS-DRG CC/MCC reference list, is essential for accurate severity stratification and equitable payment under CHS-DRG. Second, internal hospital quality control should focus not only on whether codes are entered but on whether the combination of codes meets the rule-based requirements of the grouper [8–10]. Third, payers and regulators should triangulate severity stratification with clinical data such as length of stay, organ-specific markers, and treatment intensity before drawing inferences about coding behavior at the institutional level [9, 12]. Audit programs that assess agreement between recorded ICD-10 codes and underlying medical record content provide a more direct route to such triangulation than payment data alone.

The proportion of hospitalizations assigned to the MCC tier in this cohort, 96 of 131 (73.28%), is high, and two non-exclusive interpretations warrant consideration. The first is that the case mix of a tertiary chest hospital is genuinely skewed toward severe pneumonia accompanied by multiple documented complications, so that a high MCC proportion reflects authentic clinical complexity. The second is that more complete or more assiduous documentation of secondary diagnoses, rather than a difference in underlying illness, shifts hospitalizations into higher severity tiers; international experience demonstrates that changes in clinical coding practice alone can move discharges across administrative severity boundaries [11, 12]. The present design cannot adjudicate between these interpretations, because admission-level measures of clinical severity were not available in the analytic dataset. We were moreover unable to benchmark the observed MCC proportion against a provincial or national CHS-DRG ES-series reference distribution for pneumonia, because a comparable published distribution for the relevant ADRG was not accessible to us; this absence is acknowledged as a limitation, and we therefore refrain from interpreting the 73.28% figure as evidence either of appropriate or of inappropriate coding [2, 5, 7].

The exclusion of in-hospital deaths merits explicit comment in the interpretation of the gradients. Because fatal hospitalizations are plausibly enriched both for high-severity coding and for high resource use, their exclusion would be expected to attenuate rather than to inflate the observed gradients. The sensitivity analysis that reintroduced codable fatal hospitalizations (Section “Cohort assembly and flow”) did not alter the direction or the statistical significance of the findings, which mitigates, although it does not wholly eliminate, the concern of survivor bias.

This study has several limitations. First, it is a single-center retrospective analysis conducted at a tertiary chest hospital; generalizability across regions, hospital tiers, and CHS-DRG versions is therefore limited. Second, the construction of the exposure overlaps definitionally with the outcome of severity stratification, as discussed above; this is the central methodological caveat and is reflected throughout the framing of the analysis. Third, no clinical-process variables (laboratory values, imaging findings, treatment intensity, or microbiological results) were available in the analytic dataset, and unmeasured confounding cannot be excluded. Fourth, length of stay and the count of recorded chronic conditions are contemporaneous and partially coding-dependent rather than pre-admission predictors, and the multivariable model should be interpreted accordingly. Fifth, the modest sample size (131 hospitalizations, of which 96 entered the MCC stratum) limits the statistical resolution of the multivariable model. Multicenter studies with prospective audit of coding decisions, integration of clinical-process data, and external validation of grouping outcomes are required before broader inferences can be drawn.

Conclusion

In this single-center retrospective cohort of 131 hospitalizations for pneumonia with coexisting hypertension and/or diabetes mellitus, more detailed coding of chronic comorbidities, target organ damage, and acute complications corresponded directly to higher CHS-DRG severity stratification, higher CHS-DRG RW, and higher total hospitalization expenditure, reflecting the rule-based logic of the grouper. Length of stay and the number of recorded chronic conditions were associated with entry into the MCC stratum after multivariable adjustment, but both are contemporaneous and partially coding-dependent variables rather than pre-admission predictors. Standardized documentation and rule-conformant ICD-10 coding of chronic comorbidities and their complications are necessary conditions for accurate severity stratification and equitable payment under CHS-DRG. Multicenter studies are required to confirm these correspondences and to clarify their implications for payment governance.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

TL conceived the study, designed the analysis, performed the data extraction and statistical analyses, and drafted the manuscript. YL contributed to the design of the coding pattern definitions and to the interpretation of CHS-DRG payment implications. YL (Yueyue Liang) and XL contributed to data extraction, cross-checking of front-page fields, and revision of the manuscript. NX supervised the study, contributed to interpretation, and critically revised the manuscript. All authors read and approved the final manuscript. All named authors meet the four authorship criteria of the International Committee of Medical Journal Editors, having made substantial contributions to the conception or design of the work or to the acquisition, analysis, or interpretation of the data; having drafted the work or revised it critically for important intellectual content; having approved the final version to be published; and having agreed to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Funding

This study did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

The de-identified case-level analytic dataset that supports the findings of this study is provided as Supplementary Material accompanying this article. To support reproducibility, the analysis code (Python and SPSS syntax) and the ICD-10 code-to-CC/MCC mapping table (Supplementary Table S1) are also provided as Supplementary Material; individual-level records beyond the de-identified analytic variables cannot be shared because of institutional governance and patient-confidentiality restrictions.

Declarations

Ethics approval and consent to participate

The study was reviewed and approved by the Ethics Committee of Anhui Chest Hospital prior to data extraction. Owing to the retrospective study design and the exclusive use of de-identified administrative data, the same Ethics Committee of Anhui Chest Hospital that approved the study also formally waived the requirement for individual informed consent to participate. The study was conducted in accordance with the principles of the Declaration of Helsinki.

Consent for publication

Not applicable. This manuscript does not contain any individual person’s identifying images or any other personal or clinical details that could compromise anonymity; all data were analyzed and are reported in de-identified, aggregate form.

Competing interests

The authors declare no competing interests.

Use of generative artificial intelligence

Not Applicable.