Evaluating hospital performance within a national public health framework has long challenged health economists, clinical leaders, and policymakers. Traditional metrics—such as crude mortality rates, mean lengths of stay, and simple financial throughput measures—often fail to capture the complex web of institutional specialization, varying clinical workloads, and resource constraints inherent in modern healthcare delivery.
Addressing this analytical blind spot, a new study published in Frontiers in Health Services explores the use of Diagnosis-Related Group (DRG)-based normalized indicators to evaluate inter-hospital variation across selected cardiovascular services within the Chilean public system. Analyzing over 141,000 adult discharges from 54 public hospitals between 2019 and 2023, the research team—comprising J. Rodríguez, M. Vargas, P. Cos, and Andrés Viveros from institutions including the University of Santiago of Chile (USACH) and the University of Lleida—developed an exploratory framework using dimensionless indicators.
Rather than validating hospital efficiency or serving as a stand-alone basis for resource allocation, this research offers a pioneering structural digest. By applying principal component analysis (PCA) and K-means clustering to routinely collected administrative data from the Chilean National Health Fund (FONASA), the study successfully segments the public hospital network into four distinct, highly stable operational profiles.
Detailed Chronology & Methodology: Mapping 141,893 Discharges
The methodological framework deployed by the research team bridges traditional healthcare economics with engineering-inspired dimensionless coefficients. To eliminate direct dependence on absolute hospital size and variable case-mixes, the investigators constructed three core normalized indicators from administrative DRG records:
- Relative Efficiency (ER): Calculated as the ratio of diagnosis relative weight to the length of stay (in days). This metric serves as a descriptive proxy for throughput, where higher values indicate a combination of higher coded complexity and fewer bed-days.
- Relative Severity (SR): Computed as the clinical severity level divided by the diagnosis relative weight. This allows institutions treating severely ill cohorts to be compared proportionally against assigned case complexity.
- Relative Mortality (MR): Defined as the number of in-hospital deaths divided by the diagnosis relative weight, capturing complexity-weighted mortality outcomes.
Data Cleansing and Analytical Pipeline
The primary dataset derived from FONASA initially encompassed over 4.7 million discharge records across all specialties and age groups between 2019 and 2023. Through rigorous filtering—restricting the sample to adult patients, focusing exclusively on Cardiology, Cardiovascular Surgery, and Peripheral Vascular Surgery, and omitting records with zero or missing lengths of stay—the analytical sample was refined to 141,893 discharges across 54 public hospitals.
Using Python 3.10.11 and libraries such as pandas, NumPy, and Scikit-learn, the hospital-level means of these standardized indicators (z-scores) underwent dimensionality reduction via PCA. The first two principal components accounted for an overwhelming 86.26% of total variance (PC1: 58.27%; PC2: 27.98%). Sampling adequacy was confirmed via the Kaiser-Meyer-Olkin (KMO) statistic ($KMO = 0.6251$) and Bartlett’s test of sphericity ($p = 1.00 times 10^-10$).
To categorize the institutions, K-means clustering was executed on the principal components ($k=4$, $n_init=50$, $textrandom_state=42$). The resulting four-cluster partition demonstrated exceptional algorithmic stability, boasting an average Adjusted Rand Index (ARI) of 0.9957 across 1,000 randomized initializations.
Supporting Context & Metrics: The Four Hospital Profiles
The cluster analysis revealed profound structural and operational heterogeneity across the 54 public hospitals examined. The system displayed a massive coefficient of variation (CV) of 126.1% for Relative Efficiency, highlighting stark operational disparities across the network.
The K-means segmentation categorized the 54 hospitals into four distinct clusters:
- Cluster 0 (8 hospitals, 14.8%): Characterized by low relative throughput/efficiency ($ER = 0.2360$), high clinical severity ($SR = 2.4178$), and low relative mortality ($MR = 0.0085$).
- Cluster 1 (8 hospitals, 14.8%): Exhibited the highest relative throughput ($ER = 0.9633$), lowest clinical complexity relative to weight ($SR = 1.0720$), and minimal relative mortality ($MR = 0.0016$). Top-tier efficiency hospitals clustered predominantly here.
- Cluster 2 (26 hospitals, 48.1%): The largest segment, presenting intermediate efficiency ($ER = 0.4709$) and severity ($SR = 1.3332$) with low-to-moderate mortality ($MR = 0.0151$).
- Cluster 3 (12 hospitals, 22.2%): Combined low relative throughput ($ER = 0.2465$), high clinical severity ($SR = 2.3286$), and the highest complexity-weighted mortality rate in the system ($MR = 0.0375$).
Furthermore, Kruskal-Wallis tests and Dunn’s post-hoc analyses identified significant differences in relative efficiency across medical specialties and cluster assignments ($p < 0.001$). Notably, an evaluation of temporal stability revealed that hospital throughput experienced a statistically significant dip during the COVID-19 pandemic ($-text12.68%$, $p < 0.001$), followed by a partial post-pandemic recovery ($+text5.88%$, $p < 0.001$).
Official Statements and Expert Perspectives
The authors emphasize that these indicators must be interpreted with extreme caution. In a policy statement accompanying the study, corresponding author Andrés Viveros and his colleagues underscored the study’s exploratory nature:
"The outputs should be interpreted as descriptive and hypothesis-generating rather than as validated measures of hospital efficiency or stand-alone evidence for resource-allocation decisions. While DRG systems provide structural transparency, raw administrative ratios do not replace context-aware clinical audits or account for unobserved institutional constraints."
The research team noted that although statistically significant differences in throughput emerged between male and female cohorts, the practical effect size was negligible ($eta^2 = 0.00024$), cautioning against premature conclusions regarding gender bias or clinical inequities within the public network. Instead, organizational structures, referral networks, and coding behaviors are posited as the primary drivers of observed variation.
Future Outlook: Implications for Policy and Healthcare Management
As health systems across Latin America and the globe grapple with post-pandemic fiscal tightening and rising demand, the deployment of transparent, administrative data-driven tools is increasingly vital.
This study demonstrates that routinely collected DRG data can be successfully leveraged to construct stable, reproducible hospital profiles without requiring complex, data-heavy frontier modeling like Data Envelopment Analysis (DEA). However, the path forward requires several critical developments:
- External Validation: The proposed ratios (ER, SR, MR) must be tested against external clinical registries and risk-adjusted mortality models to establish true construct validity.
- Integration of Contextual Covariates: Future iterations of the model must incorporate institutional variables such as geographic location (urban vs. rural settings), teaching hospital status, bed capacity, and specialist density.
- Granular Within-Hospital Analysis: Shifting from hospital-level aggregations to service-level or team-level evaluations will help isolate whether variation stems from specific clinical departments or systemic administrative bottlenecks.
Ultimately, the findings indicate that a generalized, undifferentiated evaluation framework for national public health systems risks overlooking vital institutional nuances. By utilizing transparent, DRG-based normalized indicators as screening tools, health authorities can better target institutions requiring deep-dive managerial reviews, laying the groundwork for more resilient, data-informed healthcare governance.
