The HALOS (Hospital Admission Length of Stay) Model: a New Tool for Predicting Hospital

The HALOS modelPage | 1

The HALOS (Hospital Admission Length Of Stay) model: A new tool for predicting hospital length of stay after liver transplantation

Kartik R Krishnan, MD1, Renuka Bhattacharya, MD2, AremaPerreira, MD2, Robert L Carithers, Jr., MD2, Jorge Reyes, MD3and James D Perkins, MD3.1Internal Medicine, University of Washington, Seattle, WA, United States;2Gastroenterology, University of Washington, Seattle, WA, United States and3Transplant Surgery, University of Washington, Seattle, WA, United States.

ABSTRACT

Background: Liver transplant centers are under increasing pressure to understand and optimize transplant-related costs. Hospital admission length of stay (LOS) is known to be a significant driver and useful surrogate for cost. We sought to develop and evaluate mathematical models predictive of LOS through retrospective analysis of the large, multi-institutional United Network for Organ Sharing Standard Transplant Analysis and Research files.

Methods: Of 45,256 adults (≥18 years) who underwent liver transplantation for any indication between January 1, 2003 and November 30, 2010, 3,705 were excluded for missing LOS data or death less than 19 days into the hospital stay, leaving 41,551 recipients in the study. For predictive model building internal split-group validation was employed, with 75% and 25% of the study population randomly assigned to the training and validation sets, respectively. Multiple linear regression was used to generate four models predicting LOS, correlating to the stages of the transplantation process: (1) using only recipient factors; (2) using a combination of recipient factors and payment source; (3) using a combination of recipient factors, payment source, and donor factors; and (4) using a combination of recipient factors, payment source, donor factors, and one posttransplant factor from the transplant admission (acute cellular rejection episode). Use of the models projected the validation patients as low (≤8 days), moderate (>8 to 18 days), or high risk (>18 days) for prolonged LOS. Using receiver operating characteristic curve analysis from logistic regression of the three groups, we measured discrimination of our models by calculating the area under each curve (AUC).

Results: LOS ranged from 1 to 528 days, with a mean of 17.38 days (SD of 22). Overall, xx recipient factors, primary payment sources, x donor factors, and postoperative acute cellular rejection during the transplant admission were identified as predictors of LOS. For prediction of the highrisk group, the AUC was 0.73, 0.73, 0.74, and 0.75 for models (1) through (4), respectively. For prediction of the moderate risk group, the AUC was 0.55, 0.55, 0.56, and 0.57 for models (1) through (4), respectively. For prediction of the low risk group, the AUC was 0.68, 0.68, 0.70, and 0.71 for models (1) through (4), respectively.

Conclusion: This study is the first to identify mathematical models predictive of the risk for prolonged LOS, with each predictor variably weighted based on the data. Most importantly, the models accurately identify patients at high risk for prolonged LOS, and can serve as valuable tools to transplant centers striving to understand transplant-related costs.

INTRODUCTION

As their operating costs continue to rise, there is increasing pressure on liver transplant (LT) centers to understand and optimize transplant-associated costs. Patients who are deemed medically “high risk” are sometimes delayed for transplant listing by transplant centers to obtain better insurance coverage, out of concern that their care will be prohibitively expensive to the transplant center. Conversely, financial approval is easier to obtain for those patients thought to be at low risk for an expensive, long hospitalization after LT. Unfortunately, to date these costs have been largely unpredictable at the time of listing, resulting in unnecessary delays in listing while waiting for insurance coverage.

Transplant admission hospital length of stay (LOS) is a well-known driver and useful surrogate of transplant-associated costs [1]. While the model for end-stage liver disease (MELD) and donor risk index (DRI) have each been found to be predictors of LOS [2-7], no model currently exists to more directly identify potential LT recipients at highest risk for a prolonged LOS. To aid the difficult decisions involved in recipient listing, donor selection, and postoperative management, there remains an important need to understand the factors that affect costs following LT.

Our study uses retrospective analysis of a large, multi-institutional database to construct mathematical models predictive of LOS. Factors examined in the database include recipient factors, payment factors, donor factors, and one postoperative factor (acute cellular rejection). Identification of recipient factors affecting LOS may improve the pretransplant evaluation and listing process, while an understanding of donor factors may assist surgeons in graft selection.

In our experience, postoperative events such as acute cellular rejection (ACR), biliarystenosis, hepatic artery thrombosis and portal vein thrombosis lengthen hospital LOS, as time is spent managing these complications. Of these events, data on ACR are carefully recorded in the multi-institutional database, making this important complication accessible for our retrospective analysis. ACR was evaluated for any potential correlation with LOS, and to determine its influence on recipient and donor selection factors.

PATIENTS AND METHODS

Protocol, Design, Data Sources, and Inclusion Criteria

The expedited review process of the institutional review board of the University of Washington was used to approve this project. A retrospective cohort study was conducted, including data from LT recipients during the MELD era obtained from the United Network for Organ Sharing Standard Transplant Analysis and Research (UNOS STAR) files. The UNOS STAR files include data submitted by the members of the Organ Procurement and Transplantation Network (OPTN) on all waitlisted candidates, transplant recipients, and donors [8]. The Health Resources and Services Administration within the U.S. Department of Health and Human Services provides oversight for the activities of the OPTN/UNOS contractor.

For the present study, inclusion was restricted to adult patients (≥ 18 years) who underwent liver transplantation between January 1, 2003 and November 30, 2010 for any indication. Patients with multiple organ transplants were included, while those lacking length of stay data were excluded. Length of stay was defined within the UNOS STAR files as the number of consecutive hospitalization days from 24 hours prior to transplantation to the day of discharge.

Patients who died within 18 days of transplant during the initial posttransplant hospital stay were excluded, and we refer to this group as the early death patients. This was done to avoid skewing the data. Our concern was that patients dying posttransplant without first enduring a long hospital stay would share more predictors with those patients with prolonged LOS than with those surviving patients with short or moderate LOS. Therefore, we feared that inclusion of these early death patients would obscure the characteristics of the surviving patients who also did not have prolonged hospital stays. LOS for the entire available population was divided into quartiles, and then the cutoff number of 18 days was chosen because this was the LOS of the 75th percentile of patients. This cutoff point was chosen to include in the study those patients who died only after a prolonged (falling within the uppermost quartile) hospital stay, while excluding those who died earlier.

Missing categorical data was recorded as “unknown”. For missing continuous data, the mean of the existing data for that variable was substituted. No continuous data variable had data missing for greater than 1% of patients.

This work was supported in part by Health Resources and Services Administration contract 234-2005-370011C. The content is the responsibility of the authors alone and does not necessarily reflect the views or policies of the Department of Health and Human Services, nor does mention of trade names, commercial products, or organizations imply endorsement by the U.S. Government.

Clinical Outcomes and Covariate Definitions

The primary outcome was LOS, a useful marker of cost. The secondary outcome was recipient mortality. Recipient covariates included age, gender, body mass index (BMI), self-identified ethnicity, liver disease diagnosis, blood type, medical condition at the time of transplant (living at home, living in a non-intensive care unit (ICU) hospital room, or living in the ICU), primary mode of payment, secondary mode of payment, dialysis in the week before transplant, encephalopathy prior to transplant, ascites prior to transplant, use of any type of life support, use of mechanical ventilation, diabetes mellitus (as recorded in the UNOS STAR files), peripheral vascular disease, portal vein thrombosis, history of previous abdominal surgery, transjugularintrahepaticportosystemic shunt (TIPS) present at the time of transplant, any history of variceal bleeding prior to transplant, serum albumin, laboratory MELD score, receiving exception MELD points for hepatocellular carcinoma (HCC), history of previous transplant, number of previous transplants, days on transplant list, receipt of multiple transplanted organs, allograft type, and transplant procedure (whole or reduced/split). The UNOS database defines life support as mechanical ventilation, intraaortic balloon pump, or a mechanical heart device. Other definitions used by the UNOS database have been described elsewhere [8]. Recipient bilirubin, INR, and serum creatinine were not included in the statistical analysis, given their close correlation with the laboratory MELD score.

Model-building was conducted in four broad stages, involving (1) recipient factor analysis, (2) payment factor analysis, (3) donor factor analysis, and finally (4)postoperative factor analysis. The staged approach was taken to simulate the transplantation process through its different stages, including recipient listing (at which time payment source is determined), donor selection, and postoperative management, with the goal of maximizing the clinical utility of this study at each of these transplantation process stages.

For the purposes of statistical analysis, liver disease diagnoses were grouped into larger umbrella diagnoses. Specifically, the umbrella diagnosis liver cancer was used for all cases with the recorded diagnoses of hepatocellular carcinoma (HCC); HCC and cirrhosis; fibrolamellar carcinoma; cholangiocarcinoma; hepatoblastoma; hemangioendothelioma, hemangiosarcoma, or angiosarcoma; and bile duct cancer. The umbrella diagnosis of cholestatic disease includes recorded diagnoses of primary biliary cirrhosis and all types of secondary biliary cirrhosis. The umbrella diagnosis of viral cirrhosis was assigned to all cases with a hepatitis A, B, C, D, or non-A, non-B diagnosis. The umbrella diagnosis of metabolic disease was assigned to all cases with a recorded diagnosis of α-1 antitrypsin deficiency; Wilson’s disease or other copper metabolism disorder; hemochromatosis or hemosiderosis; glycogen storage disease type I or II; hyperlipidemia or homozygous hypercholesterolemia; tyrosinemia; primary oxalosis/oxaluria or hyperoxaluria; maple syrup urine disease; or metabolic disease: other. The umbrella diagnosis of benign tumor was assigned to all cases with a recorded diagnosis of hepatic adenoma, polycystic liver disease, or benign tumor: other.

Also for the purposes of statistical analysis, American Indian, Pacific Islander, and multiracial patients were combined into one group.

Donor covariates included age, gender, blood type, cause of death, ethnicity, donor after cardiac death (DCD), extended donor criteria (renal transplant definition) [9], weight, height, BMI, serum creatinine at death, aspartateaminotransferase, alanineaminotransferase, total bilirubin, diabetes, chronic hypertension, hepatitis C virus (HCV) infection, hepatitis B virus (HBV) infection, donor clinical infection, organ sharing region, distance between donor and recipient, and cold ischemia time (CIT). Donor and recipient ABO blood type match was also included in this analysis as a covariate.

One postoperative variable, ACR during the transplant admission, was included in separate analysis.

Statistical Analysis

We used descriptive statistics of mean ± SD for the continuous data and percentages for categorical data. Linear regression with standard least squares was used to perform univariable and multivariable analysis to determine factors associated with increasing LOS. A P-value of 0.05 was used to determine statistical significance. Survival curves were generated using Kaplan-Meier analysis with log-rank test for significance. Analyses were performed using JMP version 9.0.2 (SAS Institute, Cary, NC).

To build a predictive model, internal split-group validation was employed; 75% and 25% of the study population were randomly assigned to the training and validation sets, respectively. Multiple linear regression was used to generate four models predicting LOS, correlating to the stages of the transplantation process: (1) a model using only recipient factors; (2) one using a combination of recipient factors and payment source; (3) one using a combination of recipient factors, payment source, and donor factors; (4) and one using a combination of recipient factors, payment source, donor factors, and one posttransplant factor from the transplant admission (acute cellular rejection episode).

Model-building was approached in stages, as described above, to simulate the transplantation process through its different stages (recipient listing, including payment/financial analysis; donor selection; and postoperative management) and to maximize the clinical utility of the study at each of these transplantation stages. For the purposes of our analysis, living donor was included as a covariate in the recipient factors analysis, rather than in the subsequent donor factors analysis. This approach was taken because the presence of a potential living donor is information available around the time of listing, and is therefore information present at the same time other recipient factors become known to the transplant team. This approach is consistent with our goal of maximizing the clinical utility of the study.

The four models were then used to predict which of three categories each validation patient would fall into: low (≤8 days), moderate (>8 to 18 days), or high risk (>18 days) for prolonged LOS. The parameters of these categories are based on the quartiles of the entire study cohort’s LOS distribution, with an LOS of 8 days representing the 25th percentile and an LOS of 18 days representing the 75th percentile. Using receiver operating characteristic curve analysis from logistic regression of the three groups, we measured discrimination of our models by calculating the area under each curve (AUC).

For multivariable analysis, both backward and forward analyses were performed, including variables with a p-value of less than 0.25 and setting a p-value of 0.1 for initial placement in the model. To be kept in the model, a significance of p0.05 was required.

Primary source of payment was analyzed separately, after the main univariable and multivariable recipient factor analysis. This separate treatment of the payment source variable was done with the goal of better assisting transplant centers in financial analysis. Multivariable analysis was performed on the primary payment variables, controlling for thesignificant recipient predictors identified earlier.

Donor predictors of LOS were likewise identified through univariable and multivariable analysis while controlling for the recipient predictors and payment source already identified. The postoperative variable of transplant admission ACR episode was then analyzed through the same process, now controlling for the significant recipient, payment, and donor predictors already identified.

Linear regression, a parametric test, was chosen despite the non-Gaussian distribution of the LOS data. This approach is based on the central limit theorem, which allows the application of linear regression to a non-Gaussian distribution of greater than 100 data points [10].

RESULTS

Recipient demographics

Between January 1, 2003 and November 30, 2010, a total of 45,256 adult liver transplants were performed and recorded in the UNOS STAR files. Of these transplants, 3,705 cases were excluded due to either missing LOS data or recipient death less than 19 days after transplantation. The mean LOS was 17.38 days (SD of 22), with a minimum LOS of 1 day for a surviving patient and a maximum of 528 days. The mean follow-up period was 1,032 days (764).

Of the 41,551 patients included in our analysis, 67.2% were male (table 1). The mean BMI was 28.0 (SD of 5.7) and the mean age was 52.9 years (10.2). A total of 72.6% self-identified as white, 12.7% as Hispanic, and 9.3% as black. The primary liver disease diagnosis was viral cirrhosis for 29.9% of recipients. A majority of recipients, 71.7%, were living at home immediately prior to transplant, while 16.9% were confined to a general hospital floor bed and 11.4% were ICU-bound. Most (61.1%) relied on private insurance as a primary source of payment, while Medicare/Medicaid was the primary source for 34.5% of recipients. A large majority of recipients, 65.9%, were reported as exhibiting hepatic encephalopathy prior to transplant, while 77.4% were reported as exhibiting ascites. There were 6,440 cases involving recipients with diabetes, 364 involving recipients with peripheral vascular disease, and 2,287involving recipients with portal vein thrombosis. A large minority of cases, 40.4%, involved recipients with a history of prior abdominal surgery, with 2,926 cases involving a recipient with a history of previous liver transplant. The mean serum albumin was 2.9 mg/dl (0.7) and mean MELD score was 21.2, with a large SD (9.9).

Donor demographics

The mean donor age was 40.9 (17), while 60% of donors were male (table 2). Stroke (40.5% of deceased donors) or head trauma (37.5% of deceased donors) were the leading causes of death among deceased donors, while anoxia was the cause of death for 15.5% of donors. A large majority of donors, 68.4%, were white, while 12.8% and 15.4% of donors were Hispanic and black, respectively. A total of 1,729 donors were DCD donors, while 24.9% met extended donor criteria. The mean donor was overweight (BMI 26.7 (5.8)) and had renal insufficiency at the time of death (mean serum creatinine 1.5 (1.6)). A total of 7.1% and 30.9% of donors were known to have diabetes and hypertension, respectively. A small portion of donors were HCV-positive (2.8% with serum HCV-antibody) and HBV-positive (0.2% with serum HBV surface antigen), while a significant portion of donors, 35.6%, had a clinically-significant infection at the time of death.