Predicting Postoperative Complications and Mortality After Acetabular Surgery in the Elderly

A Comparison of Risk Stratification Models

Chang-Yeon Kim, MD; Nikunj N. Trivedi, MD; Lakshmanan Sivasundaram, MD; George Ochenjele, MD; Raymond W. Liu, MD; Heather Vallier, MD


Curr Orthop Pract. 2020;31(2):162-167. 

In This Article

Materials and Methods

Ethical Review and Study Design

This retrospective review of a national database was exempted from institutional review board approval and patient informed consent.

Data Search

The National Inpatient Sample (NIS) was retrospectively reviewed from 2002 to 2014. A product of the Healthcare Cost and Utilization Project, the NIS is a sample of discharges from all domestic hospitals; it remains the largest inpatient database in the United States.[10] The NIS was queried via International Classification of Diseases, 9th edition (ICD-9) diagnosis coding for patients with a primary diagnosis of acetabulum fracture (80.80, 80.81) and surgical intervention (open reduction internal fixation, total hip arthroplasty, or percutaneous fixation). Data on the following demographic variables were collected: age, sex, and race.

Outcome Measures

Our primary outcome was mortality. Our secondary outcomes were: (1) extended length of stay (LOS [rounded to the day] in top 25% of our cohort), (2) discharge to a post-acute care facility (e.g. skilled nursing facility or acute rehabilitation institute), and (3) postoperative complications. Postoperative complications were stratified into surgical, cardiac, pulmonary, renal, and postoperative intubation/transfusion (within 2 days of surgery). Missing data were addressed through multiple imputation by chained equation (MICE).

Risk Models and Statistical Analysis

We quantified patient comorbidity burden using validated Elixhauser and Charlson algorithms available for ICD-9-CM codes. Of note, we only included 29 of the 30 original ECM (cardiac arrhythmia excluded due to concerns about reliability). The Charlson comorbidities were collected via custom algorithm.[11] The Charlson weights assigned to each comorbidity range from +1 to +6 and the Elixhauser weights range from −7 to +14 (Table 1A).[6,12] We also measured comorbidity using the CCS, with weights assigned according to the model from Gagne et al. (Table 1B).[9]

Demographic variables and the comorbidity indices were placed into logistic regression models. We first created the following models: (1) only demographic variables (the "base" model), (2–4) a simple count of the Elixhauser/Charlson/Combined variables ("unweighted score" model), (5–7) the Elixhauser/Charlson/Combined weighted scores ("weighted score" model), (8–10) all Elixhauser/Charlson/Combined comorbidities as separate binary variables ("complete" model). These comorbidity models were combined with the base model to construct an additional nine additional models ("base+ECM/CCI/CCM complete", "base+ECM/CCI/CCM weighted score", "base+ECM/CCI/CCM unweighted score"), for a total of 19 logistic regression models for each outcome.

We plotted receiver operating characteristic (ROC) curves to compare our regression models based on their calculated C-statistics (area under the curve [AUC]). C statistic values range from 0.50 to 1.0, with 0.50 indicating no ability to discriminate and 1.0 indicating perfect discrimination. In general, values less than 0.70 are considered to show poor discrimination, values between 0.70 and 0.80 can be considered acceptable, and between 0.80 and 0.90 excellent.[13] Data processing and statistics were conducted using Python programming language (Python Software Foundation, and R (R Foundation for Statistical Computing, Vienna, Austria.