Hospitals’ Diversity of Diagnosis Groups and Associated Costs of Care

This is a retrospective analysis of hospitalizations for children age 0 to 21 years in the Agency for Healthcare Research and Quality, Healthcare Cost and Utilization Project Kids’ Inpatient Database (KID) 2016. The 2016 KID database contains ∼3.1 million hospitalizations in 2016, representing an estimated 6.3 million hospital discharges nationally; the KID draws on data from 4200 hospitals in 46 states and Washington DC.¹²  The study cohort was derived with the following discharge exclusions: (1) missing diagnosis or cost data (n = 40 034, 0.6%), (2) healthy newborn care (n = 1 044 084, 16.7%), and (3) labor and delivery care (n = 243 505, 3.9%). In addition, hospitals, and all of their accompanying discharges, were excluded if they (1) had a low volume (<10 annually) of pediatric discharges (n = 922 hospitals [2995 discharges]) or (2) were specialty hospitals that predominately admitted 1 type of health problem (n = 412 hospitals [121 122 discharges]) (Supplemental Fig 3). This analysis of publicly available deidentified data did not constitute human subjects research.

Hospital cost per hospitalization was the main outcome measure. Hospital costs were estimated from charges by using hospital-specific cost-to-charge ratios provided by Healthcare Cost and Utilization Project as a supplement to the 2016 KID. Cost estimates associated with a specific hospitalization were estimated by multiplying the total charges reported in the data by the appropriate hospital-specific cost-to-charge ratio and then adjusting for price factors beyond a hospital’s control by using their area wage index. Costs per discharge were Winsorized at the fifth and 95th percentiles (ie, costs lower than the fifth percentile were set to the fifth percentile, and costs over the 95th percentile were set to the 95th percentile) and then log-transformed to mitigate the effect of outliers and nonnormality, respectively. The geometric mean (ie, the back-transformed mean from the log state) of each hospital’s cost per hospitalization (across all hospitalized children in each hospital) was then calculated.

Cost per hospitalization was compared across 3 types of acute-care hospital: freestanding children’s hospital (FCHs), nonfreestanding children’s hospital (NFCHs), and NCHs. FCHs were (1) labeled FCH by the KID data element, KID_STRATUM and (2) had ≥2550 annual volume of pediatric admissions. All acute-care, general CHs had ≥2550 admissions in 2016 (correspondence Children’s Hospital Association). There were 16 KID_STRATUM-identified FCHs that did not reach this admission threshold. Those 16 hospitals were small, specialty CHs that admitted children with a select group of patients (eg, orthopedic rehabilitation patients). The discharges (n = 10 804) from those hospitals were therefore removed from analysis. NFCHs were hospitals with ≥2550 annual pediatric admissions and no freestanding designation by the KID_STRATUM data element. The remaining hospitals were designated NCHs.

The main independent variable was hospitals’ diagnosis diversity experienced by their hospitalized patients. To assess this diversity, we used 3M Health System’s All Patient Refined Diagnosis Related Groups (APR-DRG). APR-DRGs assess the array the International Classification of Diseases, 10th Revision, Clinical Modification codes (up to 25 per admission) assigned to each hospitalization at discharge and categorize the hospitalization into 1 of 316 mutually exclusive diagnosis groups (128 surgical and 188 medical). APR-DRGs encompass the spectrum of health problems necessitating hospitalization, including ones that are common, rare, simple, and complex to treat. For reach diagnosis group, the APR-DRGs also assign 1 of 4 subclasses (minor, moderate, major, and extreme) that indicate SOI.

Using the APR-DRGs, we calculated a hospital diagnosis diversity index (HDDI) for each hospital by using the Shannon–Wiener entropy index, a concept adapted from ecology that measures the complexity of an ecosystem.^13,14  We used the D-measure of diversity based on Shannon–Wiener entropy index H.¹⁴  First, we used the possible APR-DRG diagnosis groups (n = 316) and SOI (n = 4) to calculate the effective number (maximum = 1264) of total diagnosis variety treated by each hospital. HDDI increased with increasing variety. Second, we assessed the proportion of total bed days for all inpatient admissions attributable to each diagnosis. HDDI increased with decreasing proportions of bed days attributable to each diagnosis. The HDDI formula is:

where = proportion of bed days attributed to i^th diagnosis, categorized by APR-DRG SOI, and = the number of APR-DRG SOI combinations (maximum = 1254).

An example of a low-diversity hospital is one that treated 4 diagnoses that each accounted for 25% of that hospital’s bed days (HDDI = 4). In contrast, a high-diversity hospital is one that treated 400 diagnoses that each accounted for 0.25% of the hospital’s bed days (HDDI = 400).

We assessed patient demographic characteristics, including age (neonatal [0–28 days], infant [29–364 days], preschool [1–4 years], primary school [5–12 years], adolescence [13–18 years], and young adulthood [19–21 years]), sex (male, female), race and ethnicity (non-Hispanic white, non-Hispanic Black, Hispanic, Asian American or Pacific Islander, other), payer (public, private, other), median family income by zip code, elective versus nonelective admission, and transfer status (ie, transferred versus not from another hospital at admission). We assessed patient’s clinical characteristics, including case-mix index (using APR-DRGs¹⁵ ) and the presence of complex chronic conditions (CCCs).^16,17  Discharges of children with a CCC were classified by using Feudtner’s International Classification of Diseases, 10th Revision, Clinical Modification diagnosis classification scheme.¹⁶  CCCs represent defined diagnosis groupings expected to last longer than 12 months and involve either a single organ system severely enough to require specialty pediatric care and hospitalization or multiple organ systems.¹⁷  We also assessed hospitals’ teaching status, assigned from membership in the Council of Teaching Hospitals or having residency training approved by the Accreditation Council for Graduate Medical Education.

In bivariable analysis, we used χ² tests to assess differences in characteristics by hospital type (FCH, NFCH, NCH). We used Wilcoxon rank tests to assess differences in median HDDI by hospital type. To assess the bivariable association of HDDI and cost per hospitalization, we regressed HDDI on the geometric, unadjusted, log-transformed mean cost per hospitalization across all hospitals. We used a cubic model for this analysis because model fit statistics and visual inspection (Fig 1) revealed that the HDDI-cost relationship was nonlinear.

To assess variation in cost per hospitalization by hospital type, we derived a series of nested log-normal regression models. In the first model, we regressed hospital type (FCH, NFCH, and NCH) on the geometric mean cost per hospitalization without case-mix adjustments. In the second model, we adjusted for patient-level demographic (eg, age, sex, race and ethnicity, insurance, income), clinical characteristics (eg, presence of CCCs), and hospital teaching status. We then retransformed adjusted cost estimates by hospital type using Duan’s smearing estimator.¹⁸  This resource intensity adjustment was applied to all patients, irrespective of the hospital at which they were treated. In the third model, we estimated the mean of the log-transformed patient-level, resource intensity–adjusted costs for each hospital, with additional hospital-level adjustment for HDDI, assuming a cubic relationship between HDDI and log-costs. None of the hospital-level models were fit with both hospital-level HDDI and hospital-level resource intensity covariates. Covariance estimates from the log-normal models were used to assess how much variation in cost per hospitalization was explained independently by all of the covariates included in each model. The statistical significance threshold was P < .05. All analyses were performed by using SAS, Version 9.4 (SAS Institute, Inc, Cary, NC).

The analytic sample included 2816 hospitals: 1.6% (n = 46) were in FCHs, 5.3% (n = 150) were in NFCHs, and 93.0% (n = 2620) were in NCHs (Table 1). There were 1 654 869 discharges across all hospitals, including 383 789 (23.2%) in FCHs, 588 463 (35.6%) in NFCHs, and 682 617 (41.2%) in NCHs.

Median age at admission was 5 years (interquartile range [IQR]: 0–15). Age at admission varied across hospital type. For example, the percentage of neonatal discharges (0–28 days old) was 31.1%, 16.8%, and 8.9% and in FCHs, NCHs, and NFCHs, respectively (P < .001) (Table 1). Of all discharges, 47.7% were female, 45.8% were non-Hispanic white, and 52.9% were for children with public insurance. Although demographic characteristics (eg, sex and payer) varied significantly (P < .001 for all) by hospital type, the absolute differences in percentages were small across subcategories for these demographic attributes. FCHs had a larger percentage of discharges for individuals with CCCs (49.0%) than NFCHs (38.7%) and NCHs (23.7%), P < .001 (Table 2)

The median number of diagnosis groups (ie, APR-DRG SOI groups) was higher in FCHs (252; IQR: 241–262) and NFCHs (224; IQR: 197–246) than in NCHs (38; IQR: 19–72), P < .001 (Table 3). The most dramatic differences were observed for the number of surgical diagnosis groups: 97 (IQR 86–102) for FCHs, 79 (IQR 63–93) for NFCHs, and 7 (IQR 3–15) for NCHs. The median HDDI varied significantly (P < .001) by hospital type: 249 (IQR: 213–276) for FCHs, 173 (IQR: 139–210) for NFCHs, and 27 (IQR: 15–46) for NCHs. From these findings, the diversity of diagnosis groups was a median 9.2 times greater in FCHs versus NCHs (Table 3).

Across all hospitals, greater HDDI was associated (P = .002) with increased cost per hospitalization (Fig 1). Although the HDDI-cost association was nonlinear, in general greater diagnosis diversity was associated with higher cost per hospitalization. For example, HDDI of 25 and 250 was associated with cost per hospitalization of $3266 and $8977, respectively (Fig 1).

Without adjusting for patient-level or hospital-level characteristics, the geometric mean cost per hospitalization varied significantly (P < .001) by hospital type: $10 757 (95% confidence interval [CI]: $9451 to $12 243) for FCHs, $6264 (95% CI: $5830 to $6729) for NFCHs, and $4192 (95% CI: $4121 to $4265) for NCHs. Hospital type explained 11.6% of the total variance in cost per hospitalization (Fig 2).

After those adjustments, the mean cost per hospitalization was attenuated by 9.0% in FCHs ($−967; 95% CI: −$691 to −$1302). In contrast, the adjustment increased (33.2% and 25.4%, respectively) the estimated cost per hospitalization in NFCHs (+$2803; 95% CI: $2018 to $2148) and in NCHs ($+1067; 95% CI: $1061 to $1072). Together, patient-level clinical and demographic characteristics and hospital teaching status explained 33.3% of the total variance in cost per hospitalization (Fig 2).

After adding HDDI to the adjustment of patient-level characteristics and hospital teaching status, the adjusted cost per hospitalization decreased by 34.8% in FCHs (−$3406; 95% CI: −$3313 to −$3406) This adjustment significantly decreased the cost per hospitalization by an additional 32.7% in NFCHs (−$2727; 95% CI: −$2702 to −$2745) and had a modest effect (3.1% increase) on the cost per hospitalization in NCHs (−$161; 95% CI: −$158 to −$164). This adjustment resulted in a nonsignificant (P = .1) difference in cost by hospital type. Together, hospital type, the patient-level clinical and demographic characteristics, hospital teaching status, and HDDI explained 44.0% of the total variance in cost per hospitalization (Fig 2).

The current study advances understanding of the relationship between hospitals’ diagnosis diversity and associated costs for pediatric patients. This diversity contributed significantly to hospitals’ costs, adding explanatory power about the variance of cost beyond that provided by patient-level characteristics alone. Hospital-level diagnosis diversity had the largest influence on cost per hospitalization in CHs. Diversity adjustment at the hospital level lowered the patient-level adjusted costs in CHs by one-third. With hundreds of diagnosis groups treated, CHs had substantially higher diagnosis diversity than NCHs.

Further investigation is needed to understand the reasons why hospitals’ diagnosis diversity and cost per hospitalization are related. In a post hoc analysis, we excluded admissions from the top 10% of APR-DRG SOI combinations by mean cost (eg, solid-organ transplant, spinal fusion) and the results did not change: increased diversity remained associated with higher costs (P < .001). Therefore, the diversity itself appears to be an independent driver of cost, even when diagnoses that are associated with the most expensive hospitalizations are excluded. Hospitals striving to be full-service for the treatment of diverse pediatric diagnoses may require a broad and complex network of expensive, customized expenses related to highly specialized infrastructure, supplies, treatments, and staff. Clinical staff, in particular, are believed to be one of the largest contributors of hospitals’ operating costs.⁹  A variety of clinical personnel (eg, specialized physicians, nurses, therapists, technicians) on hospital staff is required to effectively care for patients with a diverse array of diagnoses. In addition, the continuous demand for hospitals to be equipped to treat a vast array of diagnoses may have necessitated increased supply and overhead costs, especially with the purchasing and maintenance of specialized medications, high-technology equipment (surgical and diagnostic), and therapies. This demand may have been particularly pertinent when no other local or regional hospital in proximity to the high-diversity hospital offered such specialized items.

The low complexity of patients and low diversity of diagnosis groups in NCHs may help explain why their costs were minimally influenced by adjustments for patient complexity and diversity. In another post hoc analysis, we remeasured diversity in NCHs when excluding admissions for infants age 0 to 29 days of age, who accounted for nearly one-third of admissions in NCHs and could have had low diagnosis diversity. This exclusion actually lowered the diagnosis diversity of NCHs. Complexity and diversity had a larger impact on costs in CHs, where the complexity and diversity of patients was much higher. Treatment of hundreds of rare conditions, both acute and chronic, was likely required to achieve the magnitude of diagnosis diversity measured across CHs in the current study. Further evaluation of the contribution of rare diagnoses to HDDI and to increased cost is warranted. It is possible that hospitals specializing in treating rare conditions may inevitably incur higher cost of care.

The framework used by hospitals to direct their clinical operations may also have influenced the relationship between diagnosis diversity and cost per hospitalization.¹⁹  Hospitals treating a limited set of diagnoses may be better positioned to operate as a “focused factory,” characterized by a uniform, engineered, evidence-based, and highly efficient approach to each diagnosis encountered.¹⁰  Hospitals treating a diverse array of diagnoses may operate more like a manufacturing “solution shop” in which creativity and calculated deviation from traditional approaches to treatment are essential to solve the rare and unusual health problems experienced by their patients.²⁰  In-depth case studies that compare care processes, operation approaches, and cost of care in hospitals with low and high diagnosis diversity are warranted.

Hospitals’ diagnosis diversity may be an important attribute to consider when assessing other outcomes beyond cost per hospitalization, including length of stay and hospital readmission. Most traditional case-mix methods used to assess hospital performance focus predominately, and appropriately, on patient-level factors, including type and number of chronic conditions.²¹  Although useful, these methods are perceived to not fully capture the full spectrum of complexity and SOI experienced across a hospital’s population of patients.²²  Hospital-level diagnosis diversity may help account for these attributes. In the current study, hospital-level diagnosis diversity did not behave as a “double adjuster” of hospital costs. Rather, it made a significant, independent contribution to the explanation of costs beyond that of patient-level factors. Additional study is necessary to understand the meaning and value of including hospital-level factors, such as diagnosis diversity, in case-mix adjustment, especially when comparing performance across different types of hospitals that care for children.

This study has several limitations. The nationally representative but limited administrative data in this study cannot be used to precisely determine if one type of hospital provides more or less expensive care than another. Because KID does not contain information on itemized cost, we could not assess the cost of specific aspects of clinical care (eg, pharmacy, medical equipment, personnel). Detailed clinical data (eg, electronic health data) may be preferable to administrative billing data to measure the type and number of patients’ chronic conditions as well as their reason for hospitalization. KID data cannot be used to fully account for SOI, functional status, social and family factors, and other attributes that may influence hospital cost. Specific information on hospital location (beyond US region) was not available in KID, and therefore, we could not assess how each hospital’s local market position may have been associated with hospital cost. Although we used a common definition of CHs in KID, there is no universally accepted standard or definition for a CH, and a different definition may yield different results. We excluded hospitals that appeared to be specialty hospitals, focused on a particular type or small set of admission diagnoses. Future investigations should explore diversity and cost in children’s specialty hospitals.

Despite these limitations, the findings from the current study may be useful in several ways. First, hospitals and payers may leverage the findings on diagnosis diversity to generate discussions about expected cost of inpatient care for children. Second, the methods and findings may prompt use of diagnosis diversity by quality improvement specialists and health services researchers for case-mix adjustment with additional, important aspects of inpatient care, beyond hospital cost, including efficiency, timeliness, and safety. Third, the findings may prompt hospitals with high diagnosis diversity to explore cross-cutting opportunities for cost containment through their approach to clinical operations. Fourth, pediatric health systems may benefit from exploring regional variation in the complexity and diversity of pediatric inpatient care across hospitals to understand the supply and demand for expensive, highly complex and diverse care.

APR-DRG

All Patient Refined Diagnosis Related Groups

CCC

complex chronic condition

CH

children’s hospital

FCH

freestanding children’s hospital

HDDI

hospital diagnosis diversity index

IQR

interquartile range

KID

Kids’ Inpatient Database

NCH

nonchildren’s hospital

NFCH

nonfreestanding children’s hospital

SOI

severity of illness