Care Fragmentation and Mortality in Readmission after Surgery for Hepatopancreatobiliary and Gastric Cancer: A Patient-Level and Hospital-Level Analysis of the Healthcare Cost and Utilization Project Administrative Database

Abstract

BACKGROUND

Hepatopancreatobiliary (HPB) and gastric oncologic surgeries are frequently performed at referral centers. Postoperatively, many patients experience care fragmentation including readmission to ‘outside hospitals’ (OSH), which is associated with increased mortality. Little is known about patient-level and hospital-level variables associated with this mortality difference.

STUDY DESIGN

Patients undergoing HPB or gastric oncologic surgery were identified from select states within the Healthcare Cost and Utilization Project database (2006–2014). Follow-up was 90 days after discharge. Analyses utilized Kruskal-Wallis test, Youden index, and multilevel modeling at the hospital level.

RESULTS

7,536 patients were readmitted within 90 days of HPB or gastric oncologic surgery to 636 hospitals. 28% of readmissions (n=2,123) were to an OSH, where 90-day readmission mortality was significantly higher: 8.0% vs. 5.4% (p<0.01). Patients readmitted to an OSH lived farther from the index surgical hospital (median 24 miles vs 10; p<0.01) and were readmitted later (median 25 days after discharge vs 12; p<0.01). These variables were not associated with readmission mortality. Surgical complications managed at an OSH were associated with greater readmission mortality: 8.4% vs 5.7% (p<0.01). Hospitals with <100 annual HPB and gastric surgeries for benign or malignant indications had higher readmission mortality (6.4% vs 4.7%, p=0.01), although this was not significant after risk-adjustment (p=0.226).

CONCLUSION

For readmissions following HPB and gastric oncologic surgery, travel distance and timing are major determinants of care fragmentation. However, these variables are not associated with mortality, nor is annual hospital surgical volume after risk-adjustment. This information could be used to determine safe sites of care for readmissions following HPB and gastric surgery. Further analysis is needed to explore the relationship between complications, the site of care, and readmission mortality.

Precis

Readmission to a hospital other than the index surgical hospital is associated with increased mortality. In this analysis, we report on risk factors for readmission to other hospitals and for readmission mortality after hepatopancreatobiliary and gastric oncologic operation using patientlevel, diagnosis-level, and hospital-level analyses.

Graphical Abstract

INTRODUCTION

Hepatopancreatobiliary (HPB) and gastric oncologic surgeries are complex procedures associated with a high rate of complications and readmissions¹. These are regularly performed at major regional referral centers, particularly since higher surgical volume has been shown to be associated with improved outcomes^2–5. As a result, patients travel great distances to undergo these procedures, despite a general preference for patients to receive care locally^6–8. The combination of these challenging issues – complex procedures with high rates of complications and increasing centralization of care within geographic areas or within healthcare systems requiring patients to travel – raises a significant concern: where should patients seek care for complications after discharge following complex procedures? Although an ideal answer to this question would be to have the patients always return to the index surgical hospital to be managed by their surgeon, readmission to other hospitals is not avoidable: prior studies have shown that more than 20% of patients after HPB and gastric oncologic surgery experience postoperative care fragmentation – that is, care at a site (or sites) other than the index surgical hospital^9–12. Care fragmentation in this patient population has been associated with poor outcomes including increased periprocedural mortality and decreased oncologic survival^9–15.

Although a body of literature supports these findings, the research community has not yet fully examined what actionable variables are contributing to these adverse outcomes. Certainly patient-level factors such as travel distance contribute to determining the site of care a patient chooses^6–8,13, but few prior studies have thoroughly examined socioeconomic disparities in care fragmentation in this patient population¹⁶. Additionally, the relationship between a hospital’s surgical volume and a number of quality measures in oncologic surgery is well-established, particularly for postoperative complications and mortality^17–19. To date, no studies have examined this volume-to-outcome relationship in the context of postoperative care fragmentation. Therefore, we hypothesized that, for patients readmitted after HPB and gastric oncologic surgery, patient-level and hospital-level variables are associated with readmission to outside hospitals and with the mortality associated with this form of care fragmentation. Specific variables to be tested included race, income, distance from the index hospital, and hospital size as measured by bed count or annual surgical volume. The primary objective of this study was to identify and disseminate specific and actionable patient-level and hospital-level variables that can be used by providers, patients, caregivers, and payers to identify at-risk patients and determine safe sites for readmission following HPB and gastric oncologic surgery.

METHODS

This project was approved as exempt by the Institutional Review Board of Washington University in St. Louis.

Data Source and Study Population

All patients 18 years or older undergoing HPB or gastric surgeries for primary or secondary malignancies (as recorded using ICD-9 diagnosis and procedure codes²⁰; Table 1) were identified within the Agency for Healthcare Research and Quality (AHRQ) Healthcare Cost and Utilization Project (HCUP)²¹ State Inpatient Databases for the following states and years: California (2006–2011), Florida (2006–2014), and New York (2006–2013). Follow-up occurred to the earliest of 90 days after discharge from the index surgical hospitalization, discharge to hospice, or first recorded adjuvant therapy (using ICD-9 codes for chemotherapy (V58.11 or 99.25), immunotherapy (V58.12), and radiation (V58.0)). No new patients were abstracted from the 4^th quarter of the final year to allow for follow-up. All eligible patients were residents of the state in which they received care and eligible readmissions were limited to that state. Patients were excluded if their initial surgery was not elective or took place more than two days after their admission. Patients with postoperative LOS less than 2 days or greater than 21 days were excluded. Less than 2 days was chosen to reduce or eliminate the potential inclusion of patients with incorrectly or over-coded procedures, assuming that the anticipated LOS for all procedures of interest be at least 2 days. Eliminating patients with LOS greater than 21 days has previously been performed in surgical population-level analyses²², with rationale provided in the Discussion section. Data was sourced from subsequent inpatient stays and emergency room visits (HCUP State Emergency Department (ED) Databases). Inter-facility transfers were accounted for if they occurred within one day of the first readmission, at which point a patient’s entire 90-day readmission ED and inpatient mortality, cost, and LOS were categorized as either a readmission to the index surgical hospital or a readmission to an outside hospital (any hospital other than the index surgical hospital). Emergency room visits that did not result in admission were included for cumulative documentation of mortality and cost but were not used to define the site of readmission or length of stay.

Table 1.

Included Diagnosis and Procedure Codes (ICD-9, Clinical Modification²⁰)

Malignancy location	Diagnosis Code	Procedure Code	Example procedure
Gastric	151.0/1/2/3/4/5/6/8/9;209.23*, 209.63/5*	43.5/6/7/9, 43.81/2/9	Partial or total gastrectomy
Pancreatic	157.0/1/2/3/4/8/9	52.51/2/3/9, 52.6/7	Distal pancreatectomy,Whipple
Hepatobiliary
Liver	155.0/1/2, 197.7,209.72*	50.22, 50.3, 51.69	Partial hepatectomy,lobectomy
Gallbladder	156	51.2/3	Cholecystectomy, including revision of earliercholecystectomy
Bile duct	156.1/2/8/9	51.63/9	Excision of commonduct

^*ICD-9 code for malignant carcinoid (ICD-9 terminology) tumor; there is no specific code for neuroendocrine tumor of the pancreas

Patients undergoing multivisceral resection were coded and analyzed in all applicable surgical subgroups. Hospital surgical volume was calculated as a sum of all eligible procedures performed at each hospital in the specific year a patient was readmitted. Threshold analyses for annual hospital surgical volume were performed for each subgroup of surgeries (gastric, hepatobiliary, and pancreatic) using thresholds at the subgroup median from this dataset and at previously published thresholds defining high-volume surgery. Thresholds were: liver 10 surgeries, 20 surgeries, and group median 36^23,24; pancreas 16, 20, and group median 33^25,26; and gastric 10 and group median 32^25,27. Although many of these publications examined surgical volume for only oncologic cases, we assumed that a combined volume of oncologic and nononcologic cases more accurately represents a hospital’s experience in the postoperative care of these patients; therefore, we incorporated both benign and oncologic procedures into the calculation of a hospital’s annual surgical volume.

Complications were coded from any ICD-9 diagnosis code listed during the index surgical hospitalization or subsequent hospitalizations according to a coding and categorization schema that incorporated standardized complication coding systems (eTable 1). Comorbidities were coded from diagnoses documented in inpatient records 6 months prior to and including the index surgical hospitalization using Elixhauser Comorbidity measures from ICD-9 codes converted into pooled weighted continuous indices for mortality and readmission^28–30. Hospital variables including size by bed count and teaching hospital status were obtained from the American Hospital Association Annual Survey Database™³¹ for even-numbered years from 2006 – 2014, with odd-numbered years reflecting data from the prior year. For multilevel models, hospital-level variables were converted into an average across all years of the inclusion period weighted by the number of patient readmissions to that hospital each year. Cost was calculated using HCUP-provided cost-to-charge ratios³² for 2005 – 2014 and converted into 2014 $USD using the Medical Component of the Consumer Price Index³³.

Statistical Analysis

All analyses were performed using SAS 9.4 (SAS Institute, Inc.; Cary, NC). Group differences were examined via Chi-square tests for categorical variables and Kruskal-Wallis test for continuous variables. Logistic regression was used to compare the odds of mortality by various risk factors. Missing data were assumed to be missing at random and were excluded from analyses by pairwise deletion. Less than 2% of the cohort had missing data for the variables used in model building. Distances were calculated as the straight-line distance between centroids of a patient’s home ZIP code and the ZIP code of either the index surgical hospital or the readmission hospital using ArcMap 10.6 (Environmental Systems Research Institute; Redlands, CA). An optimal threshold of annual HPB and gastric surgery volume for mortality was identified using the Youden index (i.e., the difference between true positive and false positive), which reflects the area under the receiver operating characteristic curve³⁴. To evaluate outcomes including mortality, cost, and LOS, similar methods to those of the American College of Surgeons National Surgical Quality Improvement Program were employed³⁵: stepwise multivariate logistic regression was used for the variable (risk factor) selection, with an entry threshold of p=0.3 and removal threshold of p=0.1. The selected variables were examined as the fixed effects in a multilevel model, with the hospital as the random effect. Cost and LOS were log-transformed in the model due to their skewed distributions. Hospitals with 10 or fewer readmissions were included in model development but not in the hospital-level figures, in accordance with data use agreements.

RESULTS

31,526 patients met inclusion criteria and were discharged from the hospital following their index surgery (Figure 1). Included diagnoses and procedure codes are listed in Table 1. Demographic data is presented in Table 2. The 90-day readmission rate was 24% (n = 7,536). 90-day mortality for readmitted patients was 6% (n = 450). Readmissions accounted for a median of 6 additional days in-hospital within the 90-days after discharge (IQR 3 – 12). The cumulative median per-patient readmission cost was $13,443 (IQR $7,021 - $27,849), more than half of the initial surgical hospitalization costs (median $23,910, IQR $16,757 - $33,957).

Figure 1.

Patient inclusion schema.

Table 2.

Demographic Variables and Statistical Comparison Between Patients Readmitted to the Index Surgical Hospital and Patients Readmitted to an Outside (Non-Index) Hospital

Variable	Total (n = 7,336)	Readmitted to Index Hospital (n = 5,273)	Readmitted to Outside Hospital (n = 2,063)	p Value
Demographic
Age, y, median (IQR)	66 (57 – 74)	65 (56 – 74)	67 (58 – 75)	<0.001
Sex, f, n (%)	3,191 (43.5)	2,300 (43.6)	891 (43.3)	0.814
Race, n (%)				<0.001
White	4,675 (64.7)	3,269 (62.0)	1,406 (68.2)
Black	636 (8.8)	490 (9.3)	146 (7.1)
Hispanic	997 (13.8)	739 (14.0)	258 (12.5)
Asian or Pacific Islander	631 (8.7)	506 (9.6)	125 (6.1)
Other or missing	397 (5.4)	269 (5.1)	128 (6.2)
Elixhauser Comorbidity Indexfor readmission, median (IQR)	29 (21 – 40)	29 (21 – 39)	30 (22 – 42)	<0.001
Primary payer, n (%)				<0.001
Medicare	3,769 (51.4)	2,603 (49.4)	1,166 (56.5)
Medicaid	632 (8.6)	489 (9.3)	143 (6.9)
Private	2,705 (36.9)	2,001 (38.0)	704 (34.1)
Self-pay or other	230 (3.1)	180 (3.4)	50 (2.4)
Median household income forpatient ZIP code, n (%)				0.659
1 - Lowest quartile, by state	1,501 (21.0)	1,096 (21.3)	405 (20.1)
2	1,713 (23.9)	1,218 (23.7)	495 (24.5)
3	1,893 (26.5)	1,357 (26.4)	536 (26.6)
4 - Highest Quartile, by state	2,047 (28.6)	1,464 (28.5)	583 (28.9)
Patient residence: urban-rural,n (%)				<0.001
Urban (population center ≥250,000)	6,758 (92.1)	4,928 (93.5)	1,830 (88.7)
Rural (population center <250,000)	576 (7.9)	344 (6.5)	232 (11.3)
Index surgical hospitalizationvariable
Operation*, n (%)				<0.001
Gastric malignancy	2,679 (36.5)	2,033 (75.9)	646 (24.1)
Hepatobiliary malignancy	2,444 (33.3)	1,704 (69.7)	740 (30.1)
Pancreatic malignancy	2,557 (34.9)	1,785 (69.8)	772 (30.2)
Length of stay during index hospitalization, d, median(IQR)	8 (6 – 11)	8 (6 –11)	8 (6 – 11)	0.001
Discharge destination, n (%)				0.051
Home	4,208 (57.4)	2,980 (56.5)	1,228 (59.5)
Home with home health	2,502 (34.1)	1,841 (34.9)	661 (32.0)
Skilled nursing or other intermediate care facility	626 (8.5)	452 (8.6)	174 (8.4)
Days to readmission, median (IQR)	14 (6 – 41)	12 (5 – 33)	25 (8 – 54)	<0.001

^*Totals exceed 100 due to multivisceral resection at the index operation IQR, interquartile range

Of all readmitted patients, 28% (n = 2,123) were readmitted to an ‘outside hospital’ (OSH), defined as any hospital other than the index surgical hospital. Mortality was significantly higher for patients readmitted to OSH: 8.0% vs. 5.4% (OR 1.5, 95% CI 1.3 – 1.9; p<0.001).

Demographic Variables

Several demographic variables were independently associated with readmission to OSH, including race, insurance status, and urban-rural residence (Table 2). There is considerable overlap within these data across age, race, insurance status, and urban-rural residence. When applying multivariable regression, many demographic variables were no longer associated with the site of readmission but increasing age and rural residence remained significantly associated with readmission to OSH (Table 3). On multivariable regression modeling, readmission to OSH was significantly associated with 90-day readmission mortality (p<0.001; OR 1.50, 95% CI 1.22 – 1.84) (eTable 2). Socioeconomic variables were not significantly associated with mortality in this model.

Table 3.

Multivariable Model for Variables Associated with 90-Day Postoperative Readmission to an Outside Hospital after Hepatopancreatobiliary and Gastric Oncologic Operation

Variable	p Value	OR	95% CI
			Low	High
Age, y	0.002	1.01	1.00	1.02
Race	<0.001
White		1.00	-	-
Black		0.75	0.61	0.93
Hispanic		0.94	0.80	1.11
Asian or Pacific Islander		0.67	0.54	0.84
Other or missing		1.25	1.00	1.02
Elixhauser Comorbidity Index forReadmission	<0.001	1.01	1.00	1.01
Payer	0.028
Medicare		1.00	-	-
Medicaid		0.81	0.64	1.03
Private		0.86	0.74	0.99
Other		0.64	0.45	0.90
Urban-rural residence	<0.001
Urban		1.00	-	-
Rural		1.72	1.43	2.08
Operation	<0.001
Liver		1.00	-	-
Gastric		0.71	0.62	0.81
Pancreas		0.98	0.86	1.12
Discharge status	<0.001
Home		1.00	-	-
Home with home health		0.81	0.72	0.91
Facility (skilled nursing, rehabilitation,etc)		0.77	0.63	0.95
Time to readmission, d	<0.001	1.02	1.01	1.02

c = 0.641

OR, odds ratio

Time and Distance

Table 2 and Table 3 both demonstrate a significant difference in the timing of initial readmission between patients readmitted to the index hospital versus those readmitted to OSH: patients in the OSH group presented much later (median post-discharge day 25) than the patients readmitted to the index hospital (median post-discharge day 12; p<0.001). We hypothesized that this time differential could be associated with disparities in access to care or in the clinical indications for readmission. However, there was no significant difference in median household income between groups (Table 2). We then analyzed distance data, which. is not available for patients from the California database, leaving 4,951 patients (67%). The median distance between a patient’s home ZIP code and the ZIP code of the index surgical hospital was 11 miles (IQR 5 – 29). Patients readmitted to an OSH lived much closer to the OSH than to the index surgical hospital: median 5 miles to the OSH vs median 23 miles to the index surgical hospital (p<0.001). The likelihood of readmission to OSH increased significantly as distance from the index surgical hospital increased, particularly for patients living more than 100 miles from the index hospital (eFigure 1). Significant racial disparities were noted in the distance data. Caucasian patients were more likely to be readmitted to an OSH than non-white patients (30% vs 25%, OR 1.3, 95% CI 1.2 – 1.5; p<0.001), perhaps because Caucasian patients lived more than twice as far from the index hospital as non-Caucasian patients (median 16 miles vs 7 miles; p<0.001). Racial disparities in distance travelled and the site of readmission persisted in subgroup analyses (Figure 2). Distance from the index hospital was incorporated into the multivariable model initially presented in Table 3 and resulted in significantly improved model fit for modeling readmission site (c = 0.734 from 0.641) (Table 4).

Figure 2.

Travel distances between a patient’s home ZIP code and the ZIP code of the index surgical hospital or the non-index (“outside”) hospital. The box is bound by the 1^st and 3^rd quartiles, with the group median represented by the middle bar. Whiskers represent the 10^th and 90^th percentiles.

Table 4.

Multivariable Model for 90-Day Readmission to an Outside Hospital Following Hepatopancreatobiliary and Gastric Oncologic Operation, Adding the Distance Between the Patient’s Home ZIP Code and the ZIP Code of the Index Surgical Hospital

Variable	p Value	OR	95% CI
			Low	High
Distance from index hospital, mi	<0.001	1.02	1.01	1.02
Age, y	<0.001	1.02	1.01	1.02
Race	0.012
White		1.00	-	-
Black		0.92	0.72	1.17
Hispanic		1.32	1.06	1.63
Asian or Pacific Islander		0.84	0.59	1.19
Other or missing		1.36	1.01	1.81
Operation	0.046
Hepatobiliary		1.00	-	-
Gastric		1.05	0.89	1.25
Pancreatic		1.22	1.03	1.44
Time to readmission, d	<0.001	1.02	1.02	1.02

c = 0.734

OR, odds ratio

Distance between the patient’s home ZIP code and the ZIP code of the index surgical hospital was not associated with mortality (p=0.202), nor was longer time from discharge to readmission (p=0.119).

Patient-Level Clinical Variables

Patients experiencing any surgical complication (e.g., wound infections, sepsis, bowel obstructions - see eTable 1 for complete ICD-9 listings; n=5,935, 79.7% of entire readmitted cohort) had higher mortality if they were cared for at the OSH (8.4% vs 5.7%; p<0.001). 548 patients (7%) underwent surgery during a readmission, which took place more often in the cohort readmitted to the index hospital (8% vs 5%, p<0.01). 1,356 patients (18%) underwent endoscopy during a readmission, with no significant difference in the rate of endoscopy between the sites of care (p=0.07). 90-day mortality for patients undergoing surgery or endoscopy during a readmission was 8.4% (146 deaths out of 1,741 cases); this was not significantly between the OSH and index groups (10.3% vs 7.8%; p=0.099). Patients experiencing medical complications as their primary readmission diagnosis (e.g., pneumonia, cardiac events, or stroke; n=235, 3.2% of entire cohort) also had higher mortality if they were cared for at the OSH (39.1% vs 23.8%; p=0.017).

Hospital-Level Variables

Readmissions took place at 636 hospitals. The median number of patients readmitted per hospital during the inclusion period was 5 (IQR 2 – 11). 112 hospitals (18%) had only one readmission of an eligible postoperative patient during the inclusion period. The median hospital size by bed count was 424 beds (IQR 233 – 665). The median annual hospital volume of oncologic and nononcologic HPB and gastric surgeries for any indication was 59 cases (IQR 16 – 219). When patients presented to OSH, these hospitals were significantly smaller than the index surgical hospital by bed count (median 291 beds vs 469; p<0.001) and by annual volume of HPB and gastric surgeries (median 14 vs 105; p<0.001) (eFigures 2 and 3). 66% of the readmissions occurred at teaching hospitals. Readmissions to teaching hospitals resulted in lower rates of 90-day readmission mortality: 7.3% vs 5.5% (p=0.003). Teaching hospitals were significantly larger by bed count (median 509 vs 248, p<0.001).

Due to data use agreements, we excluded hospitals with less than or equal to 10 cumulative readmissions from the hospital-level analyses, leaving 75% of all readmitted patients (n=5,670) and 188 hospitals (30%). In this cohort, hospital size measured by bed count was not associated with readmission mortality, cost, or LOS (data not shown). Annual volume of HPB and gastric surgeries at the hospital where a patient was readmitted was associated with readmission mortality: higher surgical volume was associated with lower mortality on univariate analysis (p=0.001). Using the maximal Youden index, a threshold of 104 annual HPB and gastric surgeries was identified as the optimal threshold above which mortality decreased (Youden index = 0.088). For pragmatic application, this was rounded down to 100 annual HPB and gastric surgeries. 56% (n=4,114) of all index surgeries in this cohort took place at index hospitals with 100 or more annual HPB or gastric surgeries. 90-day postoperative readmission mortality was 36% higher for patients readmitted to a hospital with less than 100 annual HPB or gastric procedures (6.4% vs 4.7%, p=0.007; OR 1.37, 95% CI 1.1 – 1.7). Only 38 (20%) out of the 188 hospitals had, on average, 100 or more annual HPB or gastric surgeries, but these 38 hospitals accounted for 51% (n=2,899) of all readmitted patients. When controlling for the site of readmission, the threshold of 100 annual surgeries remained significantly associated with mortality (p=0.034; OR 1.30, 95% CI 1.02 – 1.66). Surgical volume at the index hospital was not significantly associated with 90-day readmission mortality using this threshold (6.2% at lower volume hospitals vs 5.0%, p=0.055; OR 1.25, 95% CI 0.996 – 1.58).

Total 90-day readmission LOS was no different between the small and large volume hospitals (median 6 days for both; p=0.82). Total 90-day readmission inpatient costs were higher at large volume hospitals (median $15,000 vs $12,000, +25%; p<0.001).

Annual hospital surgical volume was not associated with 90-day readmission mortality for each surgical subgroup – liver, pancreas, and gastric surgery – at any predefined thresholds of ‘high-volume hospitals’ (all p>0.2; eTable 3).

Finally, we evaluated hospital-level performance using multilevel modeling. Small hospitals with 10 or fewer readmissions were reinstated into this analysis. A multivariable model for 90-day readmission mortality was developed using variables and criteria listed in eTable 4. Separate models were developed similarly for cumulative readmission LOS and cumulative readmission cost. Observed-to-expected ratios for 90-day readmission mortality were calculated per hospital and are presented in Figure 3. Contrary to the univariate analyses, annual hospital surgical volume was not significantly associated with 90-day readmission mortality in the multilevel model (p=0.226), although the site of readmission remained significant (p=0.011) (Table 5).

Figure 3.

Adjusted observed-to-expected 90-day readmission mortality by hospital. Under the multilevel model, observed values are the risk-adjusted best linear unbiased predictions including both the fixed and random effects. Expected values are the marginal linear predictions including only fixed effects. Each circle represents a hospital, sorted along the X-axis by observed-to-expected mortality (left-sided Y-axis). The color of that circle corresponds to a categorical variable for the readmission hospital’s annual volume of hepatopancreatobiliary and gastric operation (both benign and oncologic; right-sided Y-axis). No association exists between mortality and the readmission hospital’s annual volume of hepatopancreatobiliary and gastric operation.

[CE: Please set the following text as a footnote for Figure 3]: Hospitals with 10 or fewer readmissions are excluded from this figure but were included in the model. 95% CIs extend above the Y-axis due to modest model fit statistics, partially due to low event rates. The Y-axis values were chosen so that hospital differences from the O:E benchmark of 1 could be more easily visualized. A figure with a comprehensive Y-axis is available in eFigure 6.

Table 5.

Multilevel Model for 90-Day Readmission Mortality

Variable	p Value	OR	95% CI
			Lower	Upper
Age, y*	<0.001	1.03	1.02	1.04
Length of stay during index surgicalhospitalization, d*	<0.001	1.05	1.03	1.08
Elixhauser Comorbidity Index formortality, points*	<0.001	1.02	1.01	1.03
Operation	0.007
Hepatobiliary malignancy		1.00	-	-
Pancreatic malignancy		0.75	0.57	0.98
Gastric malignancy		1.09	0.85	1.41
Sex	0.010
Female		1.00	-	-
Male		1.31	1.07	1.61
Site of readmission	0.011
Index surgical hospital		1.00	-	-
Outside hospital		1.34	1.07	1.69
Discharge destination	0.013
Home		1.00	-	-
Home with home health		1.18	0.94	1.48
Skilled nursing or other facility		1.61	1.17	2.22
Annual volume of HPB and gastricoperation at the readmission hospital	0.226	1.00	0.99	1.00

^*Effects of continuous variables were assessed as 1-unit offsets from the population mean HPB, hepatopancreatobiliary; OR, odds ratio

Hospital annual surgical volume was not associated with hospital-level performance for cumulative 90-day readmission LOS (p>0.1). Hospital annual surgical volume was significantly associated with cumulative 90-day readmission cost (p=0.003): larger volume hospitals had greater cumulative readmission cost (eFigures 4 and 5).

DISCUSSION

Here, in a multi-state cohort sourced from administrative data, we have presented a comprehensive evaluation of care fragmentation – as indicated by readmissions to hospitals other than the index surgical hospital – in readmissions following HPB and gastric oncologic surgery, which has previously been associated with increased mortality^10–15. Patient-level variables associated with care fragmentation included living farther away from the index hospital and presenting later, but these variables were not associated with mortality, suggesting other factors were contributing to the mortality difference associated with care fragmentation. Patients with postoperative complications diagnosed during a readmission experienced greater mortality when these complications were treated at the outside hospital. Finally, an analysis of hospital-level variables found that annual hospital surgical volume of 100 or more benign and oncologic HPB and gastric surgeries was associated with decreased mortality, but this finding was non-significant after the application of risk-adjusted multilevel models.

This study was inspired by an existing gap in the literature exploring readmissions after complex oncologic surgery. Studies have repeatedly shown significant increases in mortality for patients readmitted to outside hospitals, but we wanted to identify variables that could be acted on to prevent patients from experiencing this adverse outcome. Centralization is one essential component of HPB and gastric surgery that adds complexity to the study of readmissions and care fragmentation: these cases are regularly performed at high-volume referral centers^36,37 yet patients have strong preferences for receiving care locally^8,38,39. Socioeconomic variables further add to the complexity by amplifying the existing limitations on access to care^40–42.

Care fragmentation and the associated mortality seems like an easy problem to solve by always transferring the patient back to the index surgical hospital, but that has a number of challenges and limitations including distance, time, payment, medical necessity, and the burden on the patient and caregivers when transitioning care further away from home. In the analyses presented here, we uncovered a number of observations and associations but ultimately found that readmissions after HPB and gastric oncologic surgery is a complex process that is difficult to accurately and comprehensively model by examining only a few variables of interest. As a result, our data has mixed and complex messages depending on the outcome examined and the methods used. Overall, our analyses identified risk-factors for readmission to an outside hospitals and associations between hospital-level variables and mortality on certain models, but mortality remains associated with variables largely inherent to the patient’s baseline performance status and the details of their initial surgical hospitalization (Table 5).

Our initial primary hypothesis was that hospital size, either by bed count or by annual surgical volume, would be associated with mortality, and that patients should be readmitted to hospitals of a particular size but not necessarily the index surgical hospital. This association was not significant in this cohort. Hospital size is used as an inexact surrogate variable to represent hospital resources or clinical expertise necessary for rescue. It appears that additional variables – likely more granular than those available in the dataset used here – would need to be analyzed to further explore why patients readmitted to an outside hospital have increased mortality. One opportunity for further analysis is exploring patient-level diagnosis data. We discovered that the management of any complications – surgical or medical – resulted in higher mortality if treated at an outside hospital. We purposefully studied a large number of complications using just these two categories for two reasons. First, the number of patients who died with specific complications was very small, which was even smaller when making between-group comparisons. Second, we must acknowledge the limited reliability of diagnoses within administrative data^43–46. Further analysis of specific complications might be better suited for a different database, particularly when studying a very specific population of patients.

We believe our results are generalizable to many HPB and gastric surgery practices. The specific states we selected represent large populations with geographic variability. The HCUP database includes hospitals regardless of size, academic affiliation, or affiliation with existing quality reporting organizations that might exclude hospitals from other available databases frequently used to study postoperative outcomes. We did exclude patients with a postoperative LOS greater than 21 days a recognize this may be a subjective limitation of the generalizability. We have done this in a prior population-level analysis of readmissions after GI surgery with the intent of improving, not restricting, generalizability of population-level research²². In our clinical practice, patients with LOS beyond a subjective threshold – around 21 days – become a unique cohort of patients that diverge from a generalizable risk profile for post-discharge events and risk-skew data like length of stay or cost. The outlier outcomes of these patients is certainly multifactorial but may be due to rare complications, the interaction of pre-existing comorbidities and postoperative complications, or extensive surgeries (e.g., multivisceral, vascular resection) with accompanying risk that is not well-captured using administrative coding. In addition, these outlier patients are managed differently after discharge in our practice through managed care (higher rates of discharge to a care facility or with home health, which has been associated with readmissions⁴⁷) and closer clinical observation (repeat visits for diagnostic procedures such as drain studies or more frequent in-person or telehealth follow-up⁴⁸), introducing bias in follow-up data.

There are several limitations to our work. First and foremost is the challenge of applying models to processes as complex as readmissions and care fragmentation. Our group has previously shown that, even with surgery-specific variables and existing validated models, achieving adequate model fit when modeling readmissions using administrative data is challenging when applying models to specific patient populations²². Limitations of this dataset include the inability to track patients across state lines, the aforementioned limitation of the reliability of coding from administrative data, and the inexact estimation of cost from administrative charges using a conversion that varies between hospitals and years. The states selected may not be representative of states with lower population density, and we do intend to test our models using data from less populated states and data from our own institution. We did not incorporate outpatient data or costs in our analysis. A final limitation is the problem of small numbers. Data use agreements required us to exclude hospitals with 10 or fewer observations for most of our hospital-level analyses, which complicated the presentation of our data and skewed the data toward larger hospitals, likely increasing the possibility of Type II error for outcomes such as hospital size and annual surgical volume. When we reintroduced small hospitals into our multilevel model, shrinkage toward the mean clearly impacted our results and may explain why hospital surgical volume was non-significant in this model. Shrinkage, which has been well-reviewed in prior publications of the methods we replicated from the American College of Surgeon’s National Surgical Quality Improvement Program (ACS-NSQIP) ^35,36, particularly affects hospitals with low event numbers, of which there are many in the cohort reported here. To address this, we also analyzed unadjusted mortality rates against modeled expected mortality, which yielded results similar to those in Figure 3 (data not shown).

CONCLUSION

Following HPB and gastric oncologic surgery, readmission to an outside hospital is a frequent occurrence and is associated with worse mortality. Patient-level risk-factors for readmission to an outside hospital include living further away from the index hospital and presenting later following discharge, but these were not associated with 90-day readmission mortality. Complications managed at the outside hospital were associated with increased mortality, and hospitals with larger annual surgical volume had lower mortality, although this association was non-significant after risk-adjusted multilevel modeling. Readmissions following complex oncologic surgeries is a complex process, made more challenging by the effects of centralization. Further investigation is needed to explore the effect of specific complications and the site of care in determining readmission mortality. Providers and support staff should use these data to identify at-risk patients and design patient-level and systems-level interventions to guide postoperative care to safe sites following HPB and gastric oncologic surgery.

Abbreviations:

ACS

American College of Surgeons

CI

Confidence Interval

ED

Emergency Department

HCUP

Healthcare Cost and Utilization Project

HPB

Hepatopancreatobiliary

IQR

Interquartile Range

LOS

Length of stay

OR

Odds ratio

OSH

Outside hospital (any hospital other than the index surgical hospital