Postoperative 30-Day Readmission: Time to Focus on What Happens Outside the Hospital

Melanie S. Morris, MD; Laura A. Graham, MPH; Joshua S. Richman, MD, PhD; Robert H. Hollis, MD; Caroline E. Jones, MD; Tyler Wahl, MD; Kamal M. F. Itani, MD; Hillary J. Mull, PhD; Amy K. Rosen, PhD; Laurel Copeland, PhD; Edith Burns, MD; Gordon Telford, MD; Jeffery Whittle, MD, MPH; Mark Wilson, MD; Sara J. Knight, PhD; Mary T. Hawn, MD, MPH

Disclosures

Annals of Surgery. 2016;264(4):621-631. 

In This Article

Methods

This is a retrospective cohort study with 30 days of postdischarge follow-up using existing administrative and clinical databases within the Veterans Healthcare System. The study protocol was reviewed and approved by the VA Central Institutional Review Board with a waiver of informed consent.

Study Sample

Patients were enrolled in the study if they had a qualifying general, vascular, or orthopedic surgery assessed by the Veterans Affairs Surgical Quality Improvement Program (VASQIP) between October 1, 2007 and September 30, 2014 (the index event). Patients were excluded from the study if they died during hospitalization or if their index hospital length of stay was less than 2 days or longer than 30 days. The unit of analysis was the hospital stay. In the event of multiple surgical procedures during a hospital stay, only operative information from the first surgical procedure was included in the analysis.

Data Sources

The VASQIP dataset was used to identify potential surgeries meeting inclusion criteria. Additional preoperative, operative, and postoperative variables abstracted by trained VASQIP nurses were also obtained. The VA Corporate Data Warehouse (CDW) was then queried for laboratory and vital signs in the 30 days before admission through hospital discharge using the Lab Chemistry and Vital Signs domains. The CDW Inpatient and Outpatient domains were queried for preadmission healthcare utilization, mental health diagnoses, and postdischarge inpatient admission. The patient domains in CDW were queried for additional sociodemographic factors such as marital status and insurance coverage.

Variables

The main outcome of interest was 30-day unplanned readmission following an index hospitalization, defined by ICD-9-CM diagnosis and procedure codes in the current CMS algorithm (2016).[3] A readmission was defined as unplanned if it did not involve any of the following: bone marrow transplant, kidney transplant, organ transplant, maintenance chemotherapy, rehabilitation, or a CMS-identified potentially planned procedure. In the event of a potentially planned procedure, the readmission was defined as unplanned if the primary diagnosis was acute or for a complication of care.[3]

Predischarge complications were defined as any occurrence of a VASQIP-identified postoperative complication before hospital discharge during the index hospitalization. Postdischarge complications were defined as any occurrence of VASQIP-identified postoperative complication after hospital discharge but before the first inpatient readmission within 30 days of hospital discharge. Postdischarge emergency room (ER) utilization within 30 days postdischarge was also included as a postdischarge complication.

In addition to VASQIP clinical comorbidities, mental health comorbidities (depression, post-traumatic stress disorder, bipolar disorder, and schizophrenia/schizoaffective disorder) were identified from ICD-9-CM diagnosis codes using the CDW Inpatient and Outpatient domains. All laboratory, vital signs, and pain scores from CDW were analyzed as the closest preoperative results within 30 days, and the highest and lowest postoperative result during the postoperative stay. To obtain a better assessment of risk across surgical specialties, the Clinical Classifications Software from the Healthcare Cost and Utilization Project was used to categorize principal Current Procedure Terminology (CPT) codes.[4] For all laboratory tests, observations without data were included in the analysis using a substituted "0" and missing data indicator; they were also used to examine whether a lab was ordered and whether this affected risk of readmission.[5]

For all patients experiencing an unplanned 30-day readmission, reasons for readmission were categorized by review of the primary and secondary ICD-9-CM diagnosis codes. Categories for readmission were identified a priori following a thorough literature review; combinations of primary and secondary ICD-9-CM codes were then independently reviewed by 3 clinicians to determine the appropriate category. Discrepancies were discussed among the clinicians until consensus was met.

Statistical Analyses

Univariate and bivariate statistics examined the factors associated with a 30-day unplanned readmission. Differences in readmission reason by specialty were also examined using bivariate statistics. Chi-square test statistics and t tests were used to compare results along with clinically meaningful differences in proportions or means.

Survival analysis estimated the hazard of readmission at each day postdischarge, both overall and stratified by specialty and reason for readmission. Multivariable logistic regression was used to model significant predictors of 30-day unplanned readmission during the hospital stay. Backwards stepwise selection at P ≤ 0.001 was initially used to identify the subset of potential variables to include in the model and forward stepwise selection was then used on the subset of potential variables to avoid overfitting the model. The relative contribution of variables in the final model was calculated as the difference between the ANOVA χ2 for each variable and its degrees of freedom (χ2-df).[6] The final model for predicting readmission at the time of discharge included 36 predictors plus 10 indicators for missing laboratory values. Procedure-specific models for each specialty were also developed in a similar fashion.

Sensitivity analyses included complete case analysis, excluding missing laboratory data, to examine the effect of including the missing data indicator. Predictive performance of logistic models was validated using cross-validation where models were fit on 90% of the data and tested on the remaining 10%. All bivariate analyses and logistic models were completed using SAS v9.4 (SAS Institute, Cary, NC).

Comments

3090D553-9492-4563-8681-AD288FA52ACE

processing....