Annals of Surgical Oncology

, Volume 19, Issue 1, pp 309–317

Development and Validation of a Reference Table for Prediction of Postoperative Mortality Rate in Patients Treated with Radical Cystectomy: A Population-based Study

  • Firas Abdollah
  • Maxine Sun
  • Jan Schmitges
  • Rodolphe Thuret
  • Orchidee Djahangirian
  • Claudio Jeldres
  • Zhe Tian
  • Shahrokh F. Shariat
  • Paul Perrotte
  • Francesco Montorsi
  • Pierre I. Karakiewicz
Urologic Oncology

DOI: 10.1245/s10434-011-1852-7

Cite this article as:
Abdollah, F., Sun, M., Schmitges, J. et al. Ann Surg Oncol (2012) 19: 309. doi:10.1245/s10434-011-1852-7

Abstract

Purpose

The existing literature suggests that the postoperative mortality (POM) rate in radical cystectomy (RC) patients does not exceed 3%. We sought to develop and externally validate a reference table that quantifies POM after RC.

Methods

We identified 12,274 patients treated with RC, between 1998 and 2007, within the Nationwide Inpatient Sample database. A total of 6188 (50.4%) randomly selected patients was used as the development cohort. Logistic regression analysis for prediction of POM adjusted for: age, sex, race, Charlson comorbidity index (CCI), urinary diversion type, year of surgery, annual hospital caseload, location/teaching status of hospital, region and bed size of hospital. The reference table was developed by using stepwise variable removal to identify the most accurate and parsimonious model. The model was externally validated in 6086 (49.6%) patients.

Results

POM occurred in 2.4% of patients. POM proportion increased with increasing age (≤59: 0.6% vs. 60–69: 1.6% vs. 70–79: 3.1% vs. ≥80: 4.6%, P < 0.001), and higher CCI (CCI 0: 1.7% vs. CCI 1: 3.0% vs. CCI 2: 4.2% vs. CCI 3: 4.3% vs. CCI ≥ 4: 12.1%, P < 0.001). In multivariable analyses, only age and CCI remained as independent predictors of POM, after stepwise variable removal. The discrimination accuracy of the reference table in predicting POM was 70%.

Conclusions

Age and CCI represent the foremost determinants of POM after RC. The developed reference table is capable of predicting POM after RC, in an individualized fashion. The accuracy of the model is good (70%), and it is highly generalizable.

Bladder cancer is the second most common genitourinary malignancy after prostate cancer. In the year 2010, bladder cancer is estimated to account for more than 14,000 deaths within the United States.1 In patients with nonmetastatic muscle-invasive bladder cancer, radical cystectomy (RC) is the treatment of choice.2

The existing literature suggests that the rate of postoperative mortality (POM) after RC is ≤3%.3, 4, 5, 6, 7, 8, 9, 10 However, postoperative outcome may vary according to several factors. For example, previous studies suggest that POM is associated with hospital characteristics, such as hospital volume.8 Moreover, POM may differ according to baseline patient characteristics. For example, more advanced age represents a determinant of increased risk of POM after surgery.6,8 In consequence, the consideration of POM after RC is important in the context of several pretreatment characteristics. As such, the risk of POM cannot be simply represented by an average rate.

Currently, it remains unknown to what extent POM after RC differs according to pretreatment hospital and patient characteristics. In consequence, we set to assess POM, defined in our study as in-hospital mortality rate, in a large population-based data. Moreover, we decided to develop a model that may help clinicians in quantifying the risk of POM preoperatively.

Materials and Methods

Data Source

We relied on the Nationwide Inpatient Sample (NIS) discharge records. The NIS is a part of the Healthcare Cost and Utilization Project (HCUP), sponsored by the Agency for Healthcare Research and Quality (AHRQ) and is considered the largest all-payer inpatient care database in the United States. It contains data from approximately 8 million hospital stays each year, approximating a 20% stratified sample of community hospitals from participating states, including academic and specialty hospitals within the United States. The sampling frame for the NIS is a sample of hospitals that comprises approximately 90% of all hospital discharges in the United States. The NIS is the only national hospital database containing charge information on all patients, regardless of payer, including persons covered by Medicare, Medicaid, private insurance, and the uninsured.11

Relying on discharge records, and using the International Classification of Diseases, 9th revision (ICD-9), diagnostic codes (C67.0–C67.6, C67.8, C67.9), and procedure codes (57.71), we identified patients diagnosed with bladder cancer and treated with RC. Patients with unknown age and/or sex data at surgery were excluded from further analysis (n = 3). These selection criteria yielded 12,274 assessable patients, treated between the years 1998 and 2007.

Variables Definition

For all patients, the following covariates were abstracted: age (≤59 vs. 60–69 vs. 70–79 vs. ≥80 years), sex (male vs. female subjects), race (white vs. black vs. others vs. unknown), Charlson comorbidity index (CCI) that was quantified by using ICD-9 codes (0 vs. 1 vs. 2 vs. 3 vs. ≥4), urinary diversion type (ileal conduit vs. other external diversions vs. neobladder vs. diversion to intestine vs. bilateral nephrectomy vs. unknown), year of surgery (1998–1999, 2000–2001, 2002–2003, 2004–2005, 2006–2007), annual hospital caseload tertiles (1–2 vs. 3–6 vs. 7–122) defined using a previously reported methodology, location/teaching status of hospital (rural vs. urban nonteaching vs. urban teaching), hospital region (Northeast vs. Midwest vs. South vs. West), hospital bed size (small vs. medium vs. large).8,11,12 In our study, POM was defined as in-hospital mortality after RC.

Statistical Analyses

Descriptive statistics focused on frequencies and proportions for categorical variables. Means, medians and ranges were reported for continuously coded variables. Chi-square and t tests were used to compare the statistical significance of differences in respectively proportions and means.

The analyses consisted of two steps. In the first step, we stratified patients according to the available covariates, and quantified POM for each category. The 95% confidence interval was also calculated for each category by using a previously described methodology.13

In the second step, we randomly divided 12,274 patients into two similarly sized cohorts. The development cohort consisted of 6188 (50.4%) patients vs. 6086 (49.6%) patients in the validation cohort. The development cohort was used to fit univariable and multivariable logistic regression models predicting POM after RC. All available covariates were included in the initial multivariable model. Stepwise variable removal was then applied according to the Akaike’s information criterion, with the intent of developing the most accurate and parsimonious model.14,15 Subsequently, the multivariable logistic regression coefficients of predictors that remained in the most accurate and parsimonious model were used to generate a reference table predicting individual probabilities of POM after RC.

The external validation cohort of 6086 patients was used to assess the accuracy of the reference table in predicting POM. Model discrimination accuracy was quantified by using the receiver operating characteristics–derived area under the curve (AUC), where 0.5 implies random predictions and 1.0 indicates perfect agreement. The extent of over- or underestimation of the observed POM versus the reference table predicted POM was explored graphically by using a calibration plot.

All statistical analyses were performed by the R statistical package system (R Foundation for Statistical Computing, Vienna, Austria), with a two-sided significance level set at P < 0.05.

Results

Between 1998 and 2007, 12,274 patients underwent RC within the NIS database (Table 1). Average patient age was 68.5 years (median 70, range 40–95). Most patients were male (83.1%) and white (65.4%). Most patients had no comorbidity according to CCI (68.1%). Urinary diversion consisted of ileal conduit in the majority of cases (76.4%). Most patients were operated in hospitals of low caseload (38.9%). Finally, most patients were operated in the South region (35.7%), and within urban-teaching hospitals (62.9%). No statistically significant differences were observed between the development and validation cohorts (Table 1).
Table 1

Descriptive statistics for 12,274 patients with bladder cancer treated with RC between 1998 and 2007, within the Nationwide Inpatient Sample database

Variable

Overall (n = 12,274)

Development cohort (n = 6188)

External validation cohort (n = 6086)

P value

Age (year)

0.4

 Mean

68.5

68.6

68.4

 Median

70

70

70

 Range

40–95

40–94

40–95

Age category (year)

0.7

 ≤59

2382 (19.4)

1178 (19.0)

1204 (19.8)

 60–69

3620 (29.5)

1839 (29.7)

1781 (29.3)

 70–79

4667 (38)

2358 (38.1)

2309 (37.9)

 ≥80

1605 (13.1)

813 (13.1)

792 (13)

Sex

0.5

 Male

10203 (83.1)

5157 (83.3)

5046 (82.9)

 Female

2071 (16.9)

1031 (16.7)

1040 (17.1)

Race

0.4

 White

8028 (65.4)

4008 (64.8)

4020 (66.1)

 Black

407 (3.3)

214 (3.5)

193 (3.2)

 Other

642 (5.2)

329 (5.3)

313 (5.1)

 Unknown

3197 (26.0)

1637 (26.5)

1560 (25.6)

Charlson comorbidity index

0.8

 0

8355 (68.1)

4211 (68.1)

4144 (68.1)

 1

2815 (22.9)

1431 (23.1)

1384 (22.7)

 2

605 (4.9)

294 (4.8)

311 (5.1)

 3

365 (3.0)

183 (3.0)

182 (3.0)

 ≥4

134 (1.1)

69 (1.1)

65 (1.1)

Urinary diversion

0.1

 Ileal conduit

9382 (76.4)

4796 (77.5)

4586 (75.4)

 Other external diversions

121 (1.0)

61 (1.0)

60 (1.0)

 Neobladder

812 (6.6)

391 (6.3)

421 (6.9)

 Diversion to intestine

1272 (10.4)

606 (9.8)

666 (10.9)

 Bilateral nephrectomy

11 (0.1)

6 (0.1)

5 (0.1)

 Unknown

676 (5.5)

328 (5.3)

348 (5.7)

Year of surgery

0.5

 1998–1999

1861 (15.2)

952 (15.4)

909 (14.9)

 2000–2001

2438 (19.9)

1250 (20.2)

1188 (19.5)

 2002–2003

2612 (21.3)

1302 (21.0)

1310 (21.5)

 2004–2005

2517 (20.5)

1281 (20.7)

1236 (20.3)

 2006–2007

2846 (23.2)

1403 (22.7)

1443 (23.7)

Hospital volume

0.4

 Low (1–2)

4776 (38.9)

2405 (38.9)

2371 (39.0)

 Intermediate (3–6)

3666 (29.9)

1878 (30.3)

1788 (29.4)

 High (7–122)

3832 (31.2)

1905 (30.8)

1927 (31.7)

Location/teaching status of hospital

0.9

 Rural

828 (6.7)

414 (6.7)

414 (6.8)

 Urban nonteaching

3731 (30.4)

1890 (30.5)

1841 (30.2)

 Urban teaching

7715 (62.9)

3884 (62.8)

3831 (62.9)

Region of hospital

0.6

 Northeast

2350 (19.1)

1204 (19.5)

1146 (18.8)

 Midwest

2916 (23.8)

1448 (23.4)

1468 (24.1)

 South

4383 (35.7)

2202 (35.6)

2181 (35.8)

 West

2625 (21.4)

1334 (21.6)

1291 (21.2)

Bed size of hospital

0.5

 Small

1159 (9.4)

567 (9.2)

592 (9.7)

 Medium

2341 (19.1)

1193 (19.3)

1148 (18.9)

 Large

8774 (71.5)

4428 (71.6)

4346 (71.4)

In-hospital death

0.5

 No

11968 (97.5)

6028 (97.4)

5940 (97.6)

 Yes

291 (2.4)

152 (2.5)

139 (2.3)

Data were stratified according to the characteristics recorded within the development and external validation cohorts

In the entire cohort of 12,274 patients, the proportion of patients who succumbed to POM was 2.4% (Table 2). POM increased with age (≤59: 0.6% vs. 60–69: 1.6% vs. 70–79: 3.1% vs. ≥80: 4.6%, P < 0.001). Similarly, higher comorbidity was associated with higher POM (CCI 0: 1.7% vs. CCI 1: 3.0% vs. CCI 2: 4.2% vs. CCI 3: 4.3% vs. CCI ≥ 4: 12.1%, P < 0.001). Patients operated in hospitals with low annual caseload volume had higher POM than their counterparts operated in medium and high annual caseload hospital (3.0 vs. 2.2 vs. 1.6%, respectively; P < 0.001). Patients operated in rural hospital had higher POM than their counterpart operated in urban nonteaching and urban teaching hospitals (3.9 vs. 2.7 vs. 2.0%, respectively; P < 0.001).
Table 2

In-hospital mortality after RC for 12,274 patients treated with RC between 1998 and 2007, within the Nationwide Inpatient Sample database

Variable

All age categories

Age ≤ 59 year

Age 60–69 year

Age 70–79 year

Age ≥ 80 year

% (95% CI) (n:overall)

% (95% CI) (n:overall)

% (95% CI) (n:overall)

% (95% CI) (n:overall)

% (95% CI) (n:overall)

Overall

2.4 (2.1–2.7) (291:12274)

0.6 (0.4–1) (15:2382)

1.6 (1.2–2.1) (59:3620)

3.1 (2.6–3.6) (143:4667)

4.6 (3.6–5.8) (74:1605)

Sex

 Male

2.3 (2.0–2.6) (238:10203)

0.6 (0.3–1.1) (14:2059)

1.5 (1.1–2) (46:3063)

3.1 (2.5–3.6) (118:3833)

4.8 (3.6–6.1) (60:1248)

 Female

2.5 (1.9–3.3) (53:2071)

0.3 (0.0–1.7) (1:323)

2.33 (1.2–3.9) (13:557)

3.0 (1.9–4.3) (25:834)

3.9 (2.2–6.5) (14:357)

Race

 White

2.3 (2.0–2.7) (189:8028)

0.6 (0.2–1.1) (9:1434)

1.6 (1.2–2.2) (38:2305)

3.1 (2.5–3.7) (98:3192)

4.0 (2.9–5.3) (44:1097)

 Black

2.4 (1.1–4.4) (10:407)

2.4 (0.5–6.8) (3:125)

1.4 (0.1–5.2) (2:135)

2.78 (0.5–7.9) (3:108)

5.13 (0.6–17.3) (2:39)

 Other

2.4 (1.4–4.0) (16:642)

0.6 (0.02–3.6) (1:150)

2.5 (0.8–5.8) (5:197)

3.6 (1.6–7.1) (8:218)

2.6 (0.3–9.1) (2:77)

 Unknown

2.3 (1.8–2.9) (76:3197)

0.3 (0.0–1.1) (2:673)

1.4 (0.7–2.3) (14:983)

2.9 (2.1–4.1) (34:1149)

6.6 (4.3–9.5) (26:392)

Charlson comorbidity index

 0

1.8 (1.5–2.1) (148:8355)

0.5 (0.2–0.9) (9:1873)

1.2 (0.8–1.8) (30:2438)

2.4 (1.9–3) (71:2992)

3.6 (2.6–4.9) (38:1052)

 1

3.1 (2.5–3.8) (88:2815)

1.1 (0.3–2.9) (4:355)

2.3 (1.4–3.5) (20:875)

3.6 (2.6–4.8) (44:1218)

5.4 (3.4–8.3) (20:367)

 2

4 (2.6–5.8) (24:605)

1.2 (0–6.3) (1:86)

2.4 (0.7–6) (4:167)

4.9 (2.7–8.3) (13:263)

6.7 (2.5–14.1) (6:89)

 3

3.8 (2.1–6.4) (14:365)

1.9 (0–10.1) (1:53)

1.9 (0.2–6.8) (2:104)

5 (2–10) (7:140)

5.9 (1.6–14.4) (4:68)

 ≥4

12.7 (7.6–19.5) (17:134)

0 (0–21.8) (0:15)

8.3 (1.8–22.5) (3:36)

14.8 (6.6–27.1) (8:54)

20.7 (8–39.7) (6:29)

Urinary diversion

 Ileal conduit

2.5 (2.2–2.8) (237:9382)

0.6 (0.3–1.2) (10:1493)

1.8 (1.3–2.4) (49:2638)

2.8 (2.3–3.4) (111:3845)

4.7 (3.7–6.0) (67:1406)

 Other external diversions

1.6 (0.2–5.8) (2:121)

0.0 (0.0–15.4) (0:22)

3.0 (0.1–15.7) (1:33)

0.0 (0.0–6.8) (0:52)

7.1 (0.1–33.8) (1:14)

 Neobladder

0.3 (0.1–1.1) (3:812)

0.0 (0.0–1.1) (0:330)

0.3 (0.0–1.8) (1:307)

1.2 (0.1–4.3) (2:162)

0.0 (0.0–24.7) (0:13)

 Diversion to intestine

1.8 (1.1–2.7) (23:1272)

1.1 (0.32–2.96) (4:343)

0.9 (0.2–2.4) (4:417)

2.9 (1.5–5.1) (12:405)

2.8 (0.5–7.9) (3:107)

 Bilateral nephrectomy

9.1 (0.2–41.2) (1:11)

0.0 (0.0–70.7) (0:3)

0.0 (0.0–52.1) (0:5)

50.0 (1.2–98.7) (1:2)

0.0 (0.0–97.5) (0:1)

 Unknown

3.7 (2.4–5.4) (25:676)

0.52 (0.01–2.88) (1:191)

1.8 (0.5–4.5) (4:220)

8.4 (5–13.2) (17:201)

4.69 (0.9–13.0) (3:64)

Year of surgery

 1998–1999

2.9 (2.2–3.8) (55:1861)

0.8 (0.1–2.3) (3:370)

2.3 (1.2–3.9) (13:560)

3.7 (2.5–5.3) (28:746)

5.9 (3.0–10.3) (11:185)

 2000–2001

2.5 (1.9–3.2) (62:2438)

0.6 (0.1–1.8) (3:467)

1.8 (1–3.1) (13:698)

3.2 (2.2–4.5) (31:960)

4.7 (2.7–7.7) (15:313)

 2002–2003

2.6 (2.0–3.2) (68:2612)

0.4 (0.1–1.4) (2:491)

1.6 (0.8–2.7) (12:749)

3.9 (2.8–5.3) (40:1014)

3.9 (2.1–6.4) (14:358)

 2004–2005

1.9 (1.4–2.5) (49:2517)

0.6 (0.1–1.7) (3:487)

1.4 (0.7–2.5) (11:756)

2.1 (1.2–3.1) (19:926)

4.6 (2.6–7.3) (16:348)

 2006–2007

2.0 (1.5–2.5) (57:2846)

0.7 (0.2–1.8) (4:567)

1.1 (0.5–2.1) (10:857)

2.4 (1.5–3.5) (25:1021)

4.4 (2.6–7.0) (18:401)

Hospital volume (cases/year)

 Low (1–2)

3.0 (2.5–3.5) (145:4776)

0.9 (0.4–1.9) (8:820)

2.2 (1.5–3.1) (32:1456)

3.8 (2.9–4.7) (71:1866)

5.3 (3.7–7.4) (34:634)

 Intermediate (3–6)

2.2 (1.8–2.8) (83:3666)

0.5 (0.1–1.3) (4:733)

1.4 (0.8–2.4) (15:1022)

2.8 (2.1–3.8) (41:1423)

4.7 (3.0–6.9) (23:488)

 High (7–122)

1.6 (1.27–2.1) (63:3832)

0.3 (0.1–1.1) (3:829)

1.1 (0.5–1.8) (12:1142)

2.2 (1.5–3.2) (31:1378)

3.5 (2.1–5.5) (17:483)

Location/teaching status of hospital

 Rural

3.9 (2.7–5.5) (33:828)

0 (0–2.76) (0:132)

2.9 (1.18–5.92) (7:240)

4.8 (2.8–7.78) (16:329)

7.8 (3.8–14.0) (10:127)

 Urban nonteaching

2.7 (2.2–3.2) (101:3731)

1.0 (0.38–2.21) (6:587)

2.0 (1.2–3.0) (22:1096)

3.3 (2.51–4.4) (51:1516)

4.1 (2.6–6.1) (22:532)

 Urban teaching

2.0 (1.7–2.3) (157:7715)

0.5 (0.25–1.02) (9:1663)

1.3 (0.8–1.8) (30:2284)

2.6 (2.13–3.36) (76:2822)

4.4 (3.2–5.9) (42:946)

Region of hospital

 Northeast

2.8 (2.1–3.5) (66:2350)

1.3 (0.5–2.9) (6:438)

1.9 (1.1–3.3) (13:664)

3.3 (2.3–4.7) (31:922)

4.9 (2.8–7.8) (16:326)

 Midwest

1.7 (1.3–2.2) (51:2916)

0.0 (0.0–0.6) (0:558)

1.2 (0.6–2.2) (11:869)

2.2 (1.5–3.3) (26:1136)

3.9 (2.1–6.5) (14:353)

 South

2.6 (2.1–3.1) (116:4383)

0.6 (0.2–1.4) (6:877)

1.7 (1.1–2.5) (23:1330)

3.6 (2.7–4.6) (59:1621)

5.1 (3.3–7.2) (28:555)

 West

2.2 (1.6–2.8) (58:2625)

0.5 (0.1–1.7) (3:509)

1.5 (0.8–2.7) (12:757)

2.7 (1.8–3.9) (27:988)

4.3 (2.4–6.9) (16:371)

Bed size of hospital

 Small

2.5 (1.7–3.6) (30:1159)

1.4 (0.2–4.0) (3:214)

1.7 (0.6–3.7) (6:347)

3.4 (1.9–5.6) (15:436)

3.7 (1.3–7.8) (6:162)

 Medium

2.7 (2.1–3.4) (64:2341)

0.7 (0.1–2.1) (3:419)

1.8 (0.9–3.1) (12:664)

3.1 (2.1–4.4) (29:925)

6.0 (3.7–9.1) (20:333)

 Large

2.2 (1.9–2.5) (197:8774)

0.5 (0.2–0.9) (9:1749)

1.5 (1.1–2.1) (41:2609)

2.9 (2.4–3.6) (99:3306)

4.3 (3.2–5.6) (48:1110)

Differences in in-hospital mortality among all stratifications were statistically significant (P < 0.001)

In multivariable logistic regression analyses that was fitted in the development cohort (n = 6188), age category (P < 0.001), CCI (P < 0.001) and location/teaching status (P = 0.05) of the hospital were the only statistically significant predictors of POM (Table 3). After application of the Akaike’s information criterion methodology, only age category and CCI (both P < 0.001) remained in the multivariable model for prediction of POM (Table 3), and represented the two variables from which the reference table was devised (Table 4).14,15
Table 3

Univariable and multivariable logistic regression analyses predicting in-hospital death in 6188 patients treated with RC

Variable

Univariable model

Full multivariable model

Reduced multivariable model

Odds ratio (95% confidence interval)

P value

Odds ratio (95% confidence interval)

P value

Odds ratio (95% confidence interval)

P value

Age category (year)

 ≤59

1.00 (Ref.)

1.00 (Ref.)

1.00 (Ref.)

 60–69

0.01 (0.00–0.01)

<0.001

2.31 (1.05–5.09)

0.03

2.27 (1.04–4.99)

0.04

 70–79

2.42 (1.11–5.30)

0.02

4.57 (2.18–9.60)

<0.001

4.64 (2.23–9.66)

<0.001

 ≥80

4.94 (2.38–10.26)

<0.001

6.25 (2.85–13.67)

<0.001

6.40 (2.95–13.85)

<0.001

Sex

 Male

1.00 (Ref.)

1.00 (Ref.)

 Female

0.02 (0.02–0.03)

0.7

0.97 (0.63–1.49)

0.8

Race

 White

1.00 (Ref.)

1.00 (Ref.)

 Black

0.02 (0.02–0.03)

0.8

1.34 (0.57–3.16)

0.5

 Other

1.16 (0.50–2.68)

0.7

0.81 (0.35–1.88)

0.6

 Unknown

0.75 (0.33–1.72)

0.4

1.30 (0.88–1.94)

0.1

Charlson comorbidity index

 0

1.00 (Ref.)

1.00 (Ref.)

1.00 (Ref.)

 1

0.02 (0.02–0.02)

<0.001

1.46 (1.00–2.13)

0.05

1.46 (1.01–2.13)

0.04

 2

1.59 (1.10–2.31)

0.01

2.46 (1.31–4.63)

0.005

2.37 (1.27–4.43)

0.007

 3

2.43 (1.30–4.51)

0.005

2.86 (1.44–5.69)

0.003

2.73 (1.38–5.40)

0.004

 ≥4

3.03 (1.54–5.95)

0.001

5.09 (2.06–12.61)

<0.001

4.99 (2.07–12.07)

<0.001

Urinary diversion type

 Ileal conduit

1.00 (Ref.)

1.00 (Ref.)

 Other external diversions

0.03 (0.02–0.03)

0.1

0.67 (0.09–4.99)

0.7

 Neobladder

0.63 (0.09–4.56)

0.6

0.34 (0.08–1.42)

0.1

 Diversion to intestine

0.19 (0.05–0.79)

0.02

1.03 (0.57–1.86)

0.9

 Bilateral nephrectomy

0.83 (0.46–1.47)

0.5

0 (NA)

0.8

 Unknown

0.01 (NA)

0.8

1.70 (0.91–3.18)

0.09

Year of surgery

 1998–1999

1.00 (Ref.)

1.00 (Ref.)

 2000–2001

0.03 (0.02–0.04)

0.2

0.84 (0.50–1.41)

0.5

 2002–2003

0.92 (0.56–1.53)

0.7

0.95 (0.58–1.58)

0.8

 2004–2005

0.99 (0.61–1.63)

0.9

0.73 (0.42–1.25)

0.2

 2006–2007

0.71 (0.42–1.21)

0.2

0.60 (0.34–1.05)

0.7

Hospital volume

 Low (1–2)

1.00 (Ref.)

1.00 (Ref.)

 Intermediate (3–6)

0.03 (0.02–0.04)

<0.01

0.99 (0.67–1.45)

0.9

 High (7–122)

0.87 (0.60–1.25)

0.4

0.74 (0.45–1.20)

0.2

Location/teaching status of hospital

 Rural

1.00 (Ref.)

1.00 (Ref.)

 Urban nonteaching

0.05 (0.03–0.08)

0.003

0.53 (0.31–0.93)

0.02

 Urban teaching

0.53 (0.31–0.91)

0.02

0.52 (0.30–0.91)

0.02

Region of hospital

 Northeast

1.00 (Ref.)

1.00 (Ref.)

 Midwest

0.03 (0.02–0.04)

0.6

0.61 (0.35–1.05)

0.07

 South

0.72 (0.44–1.20)

0.2

0.93 (0.59–1.46)

0.7

 West

0.91 (0.59–1.40)

0.6

0.96 (0.58–1.58)

0.8

Bed size of hospital

 Small

1.00 (Ref.)

1.00 (Ref.)

 Medium

0.04 (0.02–0.06)

0.05

0.71 (0.40–1.25)

0.2

 Large

0.79 (0.45–1.36)

0.3

0.58 (0.35–0.96)

0.03

Table 4

Reference table for individual prediction of POM rate after RC using CCI and age categories

CCIa

Age category, % (95% CI)

≤59 year

60–69 year

70–70 year

≥80 year

0

0.7 (0.5–1.1)

1.3 (1.0–1.7)

2.4 (1.9–2.9)

4.2 (3.2–5.5)

1

1.0 (0.7–1.5)

1.8 (1.4–2.3)

3.2 (2.7–3.8)

5.6 (4.4–7.1)

2

1.4 (0.9–2.1)

2.4 (1.8–3.3)

4.3 (3.4–5.4)

7.5 (5.6–9.9)

3

1.9 (1.1–3.1)

3.3 (2.2–4.9)

5.8 (4.0–8.2)

9.9 (6.7–14.3)

≥4

2.5 (1.3–4.7)

4.4 (2.6–7.5)

7.7 (4.7–12.3)

13.0 (7.9–20.6)

The lookup table was developed within the cohort of 6188 patients. The discrimination accuracy of the lookup table was 70% accurate in the external validation cohort of 6086 patients

aCCI was calculated according to previously used methodology.12,13

The individual discrimination accuracy of age and CCI for prediction of POM was 65.2 and 60.6%, respectively. The reference table’s discriminatory accuracy (AUC) was 70.0% in the external validation cohort of 6086 patients. The calibration plot showed that predictions of the reference table were virtually perfect (Fig. 1).
Fig. 1

Calibration plot depicting the relationship between the predicted vs. the observed POM risk of the reference table. Dashed line ideal prediction, dotted line reference table prediction

Discussion

Our objective was to assess the proportion of POM in patients treated with RC according to several hospital and patient baseline characteristics. Moreover, we sought to develop and validate a tool that can predict the risk of POM after RC. We relied on 12,274 RC records within the NIS database.

The proportion of POM was 2.4%. This proportion significantly increased from 0.6% in patients aged ≤50 years to 4.6% in patients aged ≥80 years. Similarly, POM significantly increased from 1.7% in patients with CCI of 0 to 12.1% in patients with CCI of 4 or more. POM was virtually twice higher in patients operated in low volume hospitals in comparison to their counterparts operated in high volume hospitals (3.0 vs. 1.6%, respectively). Similarly, the proportion of POM was virtually twice higher in patients operated at rural hospitals in comparison to their counterparts operated at urban, teaching centers (3.9 vs. 2.0%, respectively).

In multivariable analyses, after removal of noninfluential variables, age and CCI represented the two most important independent predictors of POM. These two variables were then used to develop a reference table for prediction of individual age- and CCI-specific POM risk after RC (Table 4). Within that reference table, POM ranged from 0.7 to 13.0%. POM increased with increasing age and number of comorbidities. For example, patients aged ≤59 years with CCI of 0 have virtually negligible POM risk (0.7%). This risk is virtually three times higher in patients aged ≤59 years, who have a CCI of 4 or more. Similarly, this risk is six times higher in patients with a CCI of 0, aged ≥80 years. Taken together, our findings showed that when age and CCI are considered, an important and statistically significant dose-response effect on POM risk exists. The discrimination accuracy of the developed model was 70%. This implies that the predictions will be incorrect in 30% of cases. However, this accuracy estimate exceeded that of age (65%) and CCI (61%) alone. Our results also indicate that none of the other examined variables can improve discrimination accuracy in a meaningful fashion. Therefore, other examined variables may be omitted when predictions are made.

Our results corroborate previous reports, where the overall POM did not exceed 3%.3, 4, 5, 6,9,10 However, the data used by previous investigators originated form substantially smaller samples (n = 101–1166). The sample size of these previous cohorts limits their robustness and generalizability. While a an existing report that examined the data of patients treated with RC, between 1988 and 1999, within the NIS database, showed that the rate of POM was 2.9%, age- and CCI-specific POM rates were not reported.8 Our findings showed that relying solely on an overall POM rate might be highly misleading. For example, in our cohort, only patients aged ≤59 or aged 60–69 years with a CCI of 2 or more (n = 5862, 48%) had a POM ≤ 3%. Conversely, virtually all other age- and CCI-specific categories had a POM > 3%, except for patients aged 70–79 years with CCI of 0 (POM: 2.4%).

Despite its strengths, the use of population-based datasets has several limitations. First, the use of in-hospital ICD-9 codes without accounting for preadmission data, may limit the available information about comorbidity and represent a potential limitation. Nonetheless, we relied on a previously reported and validated methodology, which accounts for this limitation.12 Moreover, this methodology attempts to distinguish between comorbidity and postoperative complications by using ICD-9 codes. While this may not always be feasible, the CCI profile observed in our cohort was similar to a previous population-based report.16 Second, our reference table was intended as a preoperative model that can help in clinical decision-making. In consequence, we did not include postoperative covariate, such as complication rate, that may affect POM. Moreover, the pathological information of the tumor was not available. However, for the same aforementioned reason, this covariate cannot be included in a preoperative prediction model. Third, our study relies on retrospective data that were abstracted by trained personnel. In consequence, coding errors may have affected our findings. Nonetheless, the NIS represents an important source of data that was used in several previous reports.7,8,17, 18, 19 Third, POM was defined by using in-hospital mortality, because it is the only available information recorded by the NIS. More conventional follow-up periods include 30- and 90-day mortality. Nonetheless, in-hospital mortality rates still represent an important end point that was addressed several previous reports.8,9,20

Fourth, it should be noted that like several other studies, which examined POM after RC, the current study could be affected by surgical selection bias.3, 4, 5, 6, 7, 8, 9, 10 The latter represents an important screening method that restricts RC to those individuals that are considered fit according to their surgeon assessment. However, even in these selected patients, age and CCI still represent important predictors of POM as demonstrated by our findings. Finally, the combined model discrimination of age and CCI was 70%, as quantified by the AUC. This figure is comparable to several other predictive models that are used in clinical practice.21,22 However, it is not perfect. Other more informative variable may be needed to better stratify patients at risk of POM. It is also possible that a more informative system than the CCI is required to achieve this goal. These limitations are shared with virtually all of the previous studies that were also based on the NIS database, and that addressed a similar topic.7,8

In conclusion, age and CCI are the foremost determinants of POM after RC. Our findings represent the first systematic tool capable of predicting POM after RC, in individualized fashion. The discriminatory accuracy of the developed reference table is good (70%) and highly generalizable, due to the large population-based cohort from which it was derived.

Acknowledgment

P.I.K. is partially supported by the University of Montreal Health Center Fonds de la Recherche en Santé du Quebec, the University of Montreal Department of Surgery and the University of Montreal Health Center (CHUM) Foundation.

Copyright information

© Society of Surgical Oncology 2011

Authors and Affiliations

  • Firas Abdollah
    • 1
    • 2
  • Maxine Sun
    • 1
  • Jan Schmitges
    • 1
    • 3
  • Rodolphe Thuret
    • 1
  • Orchidee Djahangirian
    • 1
  • Claudio Jeldres
    • 1
  • Zhe Tian
    • 1
  • Shahrokh F. Shariat
    • 4
  • Paul Perrotte
    • 1
  • Francesco Montorsi
    • 2
  • Pierre I. Karakiewicz
    • 1
  1. 1.Cancer Prognostics and Health Outcomes UnitUniversity of Montreal Health CentreMontrealCanada
  2. 2.Department of UrologyVita Salute San Raffaele UniversityMilanItaly
  3. 3.Martini-ClinicProstate Cancer Center Hamburg-EppendorfHamburgGermany
  4. 4.Department of UrologyWeill Medical College of Cornell UniversityNew YorkUSA

Personalised recommendations