Association between hospital case volume and mortality in pediatric sepsis: A retrospective observational study using a Japanese nationwide inpatient database

This nationwide retrospective observational study was conducted in Japan between April 2014 and March 2018. Data from the Diagnosis Procedure Combination (DPC) database, a Japanese nationwide inpatient registry, was used. Details of the DPC database have been described previously [10,11]. The following information is available from the DPC database [6,11]: baseline patient characteristics (e.g., age at admission, sex, referral from another hospital, and transport by ambulance); diagnoses recorded using both International Classification of Diseases, 10th revision (ICD-10) codes and Japanese texts; dates of medication use, medical procedures, and medical devices; discharge status (discharge to home, discharge to a nursing facility, transfer to another hospital, or death in hospital); hospital length of stay; and total medical costs. The DPC database categorizes diagnoses as follows [12]: main diagnosis, admission-precipitating diagnosis, most resource-consuming diagnosis, second-most resource-consuming diagnosis, comorbidities present at admission, and conditions arising after admission. During the study period, data from >1,000 acute care hospitals, including 80 university and 16 children's hospitals, were included in the DPC database. The database covered >90% of tertiary care emergency hospitals in Japan [6]. According to a previous validation study, the sensitivity and specificity of the main diagnosis (defined as main diagnosis or admission-precipitating diagnosis) in the DPC database were 78.9% and 93.2%, respectively [13].

This study was approved by the Institutional Review Board of Hiroshima University (approval number: E2020-2332; study title: evaluation of the volume–outcome relationship in pediatric sepsis using the DPC database; approval date: January 1, 2021). Database access was granted by the University of Occupational and Environmental Health. All procedures adhered to the institutional ethical standards on human experimentation and with the Declaration of Helsinki established in 1975. The requirement for informed consent was waived due to the study's retrospective nature and the use of anonymized patient data in the DPC database.

This study evaluated patients who met the following inclusion criteria: 1) aged 28 days to 17 years; 2) discharged from the hospital between April 2014 and March 2018; 3) had a sepsis diagnosis coded under the International Classification of Diseases, 10th revision; 4) underwent blood cultures on hospital admission day (day 0) or day 1; 5) received antimicrobial agents on day 0 or 1; and 6) required at least one of the following organ support measures on day 0 or 1: mechanical ventilation, inotropes/vasopressors (norepinephrine, epinephrine, dopamine, dobutamine, milrinone, olprinone, and vasopressin), cardiopulmonary resuscitation (CPR), mechanical circulatory support (MCS) (intra-aortic balloon pump or extracorporeal membrane oxygenation), and blood purification (intermittent or continuous renal replacement therapy, hemoadsorption, and plasma exchange) [14,15]. The ICD-10 codes for sepsis are listed in Table 1. Patients with missing data on referral from another hospital, date of diagnostic test or treatment, hospital discharge status, or hospital identification code were excluded.

We categorized hospitals by case volume during the study period. HVHs were defined as those in the highest quartile, while low-volume hospitals (LVHs) were those in the remaining quartiles.

The following variables were extracted from the DPC database: age at hospital day 0; sex; referral from another hospital; transport by ambulance; diagnosis; diagnostic test or treatment on day 0 or 1 (blood cultures, antimicrobial agents, mechanical ventilation, inotropes/vasopressors, CPR, MCS, blood purification, blood transfusion, intravenous immunoglobulin, albumin, recombinant thrombomodulin, antithrombin, corticosteroids [hydrocortisone, methylprednisolone, prednisolone, dexamethasone, and betamethasone], and enteral nutrition); hospital discharge status (dead or alive); and hospital identification code. In-hospital mortality rates were compared between patients treated at HVHs (HVH group) and those treated at LVHs (LVH group).

Qualitative variables were expressed as frequencies and percentages. Quantitative variables were reported as means and standard deviations (SDs). We initially planned to use multiple imputation to address potential missing values. However, the final dataset was complete, with no missing data. Consequently, multiple imputation was unnecessary [16,17]. In the unmatched cohort, crude in-hospital mortality rates of the HVH and LVH groups were compared using Fisher's exact test.

Propensity score (PS) matching was used in the main analysis to balance the potential confounders between the HVH and LVH groups [18]. The PS for being treated at HVHs was calculated using a multivariable logistic regression model. To adjust for organ failure status, already-known prognostic factors in pediatric sepsis, and factors of our interest, the following were selected as independent variables for the PS model [19,20,21,22,23]: age at hospital day 0, sex, referral from another hospital, transport by ambulance, and treatment on day 0 or 1 (mechanical ventilation, inotropes/vasopressors, CPR, MCS, blood purification, blood transfusion, intravenous immunoglobulin, albumin, recombinant thrombomodulin, antithrombin, corticosteroids, and enteral nutrition). The discriminative ability of the PS model was evaluated using the C-statistic.

Thereafter, one-to-one nearest-neighbor matching was performed without replacement. The caliper width of the pooled PSs was set to 0.2 SD. Variables were compared before and after PS matching, and those with an absolute standardized mean difference of <0.1 were considered well-balanced. Additionally, Kernel density plots for the distribution of the PS were created before and after matching. The odds of in-hospital mortality were compared using a mixed-effects logistic regression analysis, accounting for patient clustering within each hospital.

For sensitivity analyses, the following mixed-effects logistic regression analyses were performed: 1) an analysis with inverse probability of treatment weighting (IPTW) [18], 2) an analysis with PS matching excluding patients referred from another hospital, and 3) an analysis with PS matching excluding patients who received CPR on day 0 or 1.

All statistical tests were two-sided, and P values <0.05 were considered statistically significant. Stata/MP 15 (Stata Corp. 2017. Stata Statistical Software: Release 15. College Station, TX, StataCorp LLC) was used for all statistical analyses.

The C-statistic for the PS model was 0.75. After PS matching between the HVH and LVH groups, 234 pairs of patients were included for further analysis. In the PS-matched cohort, the absolute standardized mean differences of all variables except for referral from another hospital were <0.1 and well-balanced between the groups (Table 2). Kernel density plots for the distribution of the PS before and after matching are shown in Figures. 2 and 3.

Fig. 2.

Distribution of the propensity score before matching. HVH, high-volume hospital: LVH, low-volume hospital.

Fig. 3.

Distribution of the propensity score after matching. HVH, high-volume hospital: LVH, low-volume hospital.

The odds of in-hospital mortality were significantly lower for the HVH group compared with the LVH group (odds ratio [OR], 0.42; 95% confidence interval [CI], 0.22–0.80; P = 0.008) (Table 3).

All prespecified sensitivity analyses demonstrated significantly lower odds of in-hospital mortality in the HVH group than in the LVH group (Table 3).