Original Article
Prognostic Index for Patients With Parotid Carcinoma International External Validation in a Belgian-German Database Vincent Vander Poorten, MD, PhD1,2,3, Augustinus Hart, MSc4, Tom Vauterin, MD1,2,3, Gert Jeunen, MD1,2, Joseph Schoenaers, MD, DDS2,5, Marc Hamoir, MD, PhD3,6, Alphonsus Balm, MD, PhD7, Eberhard Stennert, MD, PhD3,8, Orlando Guntinas-Lichius, MD, PhD3,8,9, and Pierre Delaere, MD, PhD1,2
BACKGROUND: Prognostic indices for recurrence-free interval in patients with parotid carcinoma were developed and validated in a nationwide database. International validation would increase generalizability. METHODS: In a Belgian-German database that contained 237 consecutive patients with parotid carcinoma, a pretreatment prognostic index (PS1) and a post-treatment prognostic index (PS2) were validated by calculating both indices for each patient, comparing coefficients, constructing survival curves, calculating the concordance measure C, and performing Wald tests for scale and weight optimization of included variables and for the possible inclusion of new variables. RESULTS: Sixty-nine percent of patients (standard error, 5%) were disease free at 5 years. The defined cutoff points for PS1 resulted in 5-year disease-free rates from 94% (PS1 ¼ 1) to 42% (PS1 ¼ 4), and the cutoff points for PS2 resulted in 5-year disease-free rates from 93% (PS2 ¼ 1) to 40% (PS2 ¼ 4). Concordance measure C was good with 0.74 for both PS1 and PS2. Neither index could be improved statistically using this international database. There was some evidence that additional inclusion of the variable ‘number of positive lymph nodes in the neck dissection specimen’ could enhance the prognostic power of PS2. CONCLUSIONS: The prognostic indices performed adequately in this validation sample. Prospective generalized use seems to be well supported. Cancer 2009;115:540C 2009 American Cancer Society. 50. V KEY WORDS: salivary gland neoplasm, parotid carcinoma, prognostic factors, prognostic index, validation.
In the search for a practical prognostic system for patients with parotid carcinoma, we previously constructed a prognostic index based on a Cox proportional-hazards analysis in a source population of
Corresponding author: Vincent Vander Poorten, MD, PhD, Department of Otorhinolaryngology, Head and Neck Surgery, University Hospitals Leuven, Kapucijnenvoer 33 3000 Leuven, Belgium; Fax: (011) 003216332335;
[email protected] 1 Department of Otorhinolaryngology, Head and Neck Surgery, University Hospitals, Leuven, Belgium; 2Leuven Cancer Institute, Leuven, Belgium; 3European Salivary Gland Society; 4Department of Radiotherapy, Netherlands Cancer Institute, Amsterdam, The Netherlands; 5Department of Oral and Maxillofacial Surgery, University Hospitals, Leuven, Belgium; 6Department of Otorhinolaryngology, Head and Neck Surgery, St. Luc University Clinics, Brussels, Belgium; 7 Department of Head and Neck Oncology, Netherlands Cancer Institute, Amsterdam, the Netherlands; 8Department of Otorhinolaryngology, Head and Neck Surgery, Cologne University Clinic, Cologne, Germany; 9Department of Otorhinolaryngology, Head and Neck Surgery, Jena University Clinic, Jena, Germany
Conducted under the auspices of the European Salivary Gland Society (available at: www.esgs.ch accessed on November 12, 2008) Tom Vauterin’s current address: Department of Otorhinolaryngology, Head and Neck Surgery, General Hospital St. Jan, Brugge, Belgium. Gert Jeunen’s current address: Department of Otorhinolaryngology, Head and Neck Surgery, General Hospital, Mariaziekenhuis Noord-Limburg, Overpelt, Belgium. Presented at the Third International Conference on Head and Neck Oncology, European Head and Neck Society, May 8-10, 2008, Zagreb, Croatia. Received: June 16, 2008; Revised: July 30, 2008; Accepted: August 1, 2008 C 2009 American Cancer Society Published online: January 9, 2009, V
DOI: 10.1002/cncr.24015, www.interscience.wiley.com
540
Cancer
February 1, 2009
Internationally Valid Prognostic Index/Vander Poorten et al
Table 1. The Pretherapeutic and Post-therapeutic Prognostic Indices
Pretherapeutic Prognostic Indexy
Post-therapeutic Prognostic Indexy
PS1 ¼ 0.024*Aþ0.62*Pþ0.44*T þ0.45*Nþ0.63Sþ0.91*F
PS2 ¼ 0.018*Aþ0.39*Tþ0.34*N þ0.70*Sþ0.56*Fþ0.78*PGþ0.65*PM
Variable
Numbers to Fill Out in the Above Calculation
Variable
Numbers to Fill Out in the Above Calculation
A: Age at diagnosis P: Pain on presentation
No. (y) 1 ¼ no pain, 2 ¼ pain or numbness
A: Age at diagnosis T: Clinical tumor classification
T: Clinical tumor classificationz N: clinical lymph classification S: Skin invasion F: Facial nerve dysfunction
T1 (6 cm) ¼ 3 N0 ¼ 0, N1 ¼ 1, N2a ¼ 2, N2b ¼ 3, N2c ¼ 4, N3 ¼ 5 1 ¼ no invasion, 2 ¼ invasion 1 ¼ intact function, 2 ¼ paresis-paralysis
N: Clinical lymph node classification S: Skin invasion
No. (y) T1 (6 cm) ¼ 3 N0 ¼ 0, N1 ¼ 1, N2a ¼ 2, N2b ¼ 3, N2c ¼ 4, N3 ¼ 5
PS1 Value
5.80
PS1 Level 1 2 3 4
1 ¼ No invasion, 2 ¼ invasion
F: Facial nerve dysfunction PG: Perineural growth in the resection specimen PM: Positive surgical margins
1 ¼ Intact function, 2 ¼ paresis-paralysis 1 ¼ No, 2 ¼ yes
PS2 Value
PS2 Level
5.67
1 ¼ No, 2 ¼ yes 1 2 3 4
PS1 indicates pretherapeutic prognostic index; PS2, post-therapeutic prognostic index. y PS1 and PS2 result from the sum of the different terms in the formula in the upper row of Table 1. For a patient aged 70 years with a painless, 3-cm (T2) nodule, N0 neck (N0), no skin invasion, and intact facial nerve, replacing the variables by their adequate values results in the formula PS1 ¼ 0.024*70 0.62*1þ0.44*1þ0.45*0þ0.63*1þ0.91*1 ¼ 4.28. This score translates to a PS1 level of 2 and, hence, predicts an 86% chance to be recurrence free after 5 years (see Table 4). z See International Union Against Cancer 1992.31
151 patients with parotid carcinoma from the Netherlands Cancer Institute.1 Table 1 shows the pretreatment prognostic index PS1, which combines information available before surgery, and the post-treatment prognostic index PS2, which incorporates information from the surgical specimen. For each patient, the index sums the properly weighted contributions of the important clinicopathologic characteristics into a number corresponding to an estimated possibility of tumor recurrence. These indices showed good discrimination in the source population1 and in an independent nationwide database of Dutch patients with parotid carcinoma.2,3 According to Justice et al,4 the next level of validation is to go on an international level (see Table 2, Level 4). For this purpose, an international database was constructed from patients who were treated in Leuven and Brussels (Belgium) and in Cologne (Germany), where the prognostic variables needed to calculate the indices were recorded, and predictions were compared with outcomes. In this way, we tried to achieve further clinical and statistical validation.3 Clinical validation implies verifying whether the Cancer
February 1, 2009
Table 2. Hierarchy of External Validity for Predictive Systems According to Justice 19994
Level of Validation 0. 1. 2. 3. 4. 5.
Internal validation Prospective validation Independent validation Multisite validation Multiple independent validations Multiple independent validations with life-table analyses
index predicts accurately by evaluating discrimination (the relative ranking of individual risk is in the correct order) and calibration (the predicted recurrence rate is neither too high nor too low).4 Clinical utility of the index is likely if the aforementioned accuracy proves generalizable (reproducible and transportable). Statistical validation evaluates whether the index can be made more robust statistically using new information arising in this international sample. 541
Original Article
MATERIALS AND METHODS The database contains 239 consecutive patients with primary parotid carcinoma who were treated with curative intent from the academic centers of Leuven (n ¼ 67 patients) and Brussels (n ¼ 61 patients) in Belgium and Cologne (n ¼ 101) in Germany during the period from 1983 to 2004. Two hundred thirty-seven patients remained after the exclusion of 2 patients who had metastatic disease (M1) at presentation. The mean follow-up for the 192 patients who were alive at the end of followup was 55 months. Histopathologic diagnoses were made by experienced head and neck pathologists using the 1972 and 1991 World Health Organization (WHO) classifications.5,6 For analysis in the Cox model, the levels of the WHO classification were reduced to the high-grade/lowgrade malignancy dichotomy explained in the footnotes to Table 3. Table 3 displays patient, tumor, and treatment characteristics. One hundred thirty-five of 237 patients (57%) were treated with surgery and radiotherapy. There was a difference in the treatment approach between the Belgian centers (69% of patients in Brussels 70% of patients in Leuven underwent surgery and received postoperative radiotherapy) and the German center (only 42% of patients underwent surgery and received postoperative radiotherapy). Both local and regional radiotherapy were given at a median dose of 50 grays (in median fraction doses of 200 centigrays) as elective neck treatment and for patients with metastatic lymph nodes. Ninety-eight patients underwent surgery, and 1 patient received radiotherapy as single-modality treatment. Three patients received combined chemotherapy and radiotherapy. We used the same methodology that was used in our former validations.2 First, the prognostic power of PS1 and PS2 was evaluated. To perform ‘clinical validation,’ we evaluated discrimination by calculating the score for each patient, subsequently constructing Kaplan-Meier recurrence-free interval curves7 by dividing the range of scores according to the cutoff points as defined in the source population.1 This enabled us to visualize discriminative capacity and correct the ordering of patient groups. The amount of fit of the indices to the international data was quantified by entering the calculated prognostic index into a Cox model and comparing the resulting coefficient with the value of 1 in the source population. Assessment of predictive power was done by comparing the concord542
ance of patient ranking according to the prognostic index with patient ranking according to the time to recurrence. This was done by computing concordance measure C,8 which estimates the proportion of pairs of patients in whom the order of their respective prognostic indices concurs with the order of their observed recurrence-free percentages. Calibration was evaluated by plotting the observed recurrence-free percentages for the levels of prognostic index in the international database population against the reported percentages in the source population and in the former validation study. Next, we performed a statistical validation in 4 ways. First, the correctness of the weights for the individual prognostic variables that resulted from the analysis of the source population was checked. For each variable, a Cox model with only the prognostic index and the evaluated variable as covariates was fitted, thus obtaining an estimate of univariate optimal change for 1 variable at a time. The Wald test was used to produce a P value indicating the significance of this resulting improvement. Next, a Cox model was fitted with all of the variables included in the prognostic index to define their individual optimal weight in the international validation sample and to compare these weights with those reported in the source population. In the second step, we reconsidered the scale definition of the variables with regard to categorization of the interval and ordinal variables (age, clinical tumor classification [cT], and clinical lymph node classification [cN]). Dummy variables that indicated the categories (eg, for age, 70 years) were entered into the Cox model in addition to PS, including the original interval or ordinal variable, to judge improvement from the Wald test. In the third step, we considered new clinicopathologic variables for inclusion in the prognostic model; improvement was judged again from the P value generated by the Wald test, always correcting coefficients for treatment differences. Finally, the proportional-hazards assumption was verified by plotting the weighted Schoenfeld residuals against time and drawing a smoothed spline through the points.9 On the basis of the spline appearance, a PS-time interaction term was tested for significance in a time-dependent proportional-hazards model. The linearity (ln) of the relation between the prognostic index and the ln(hazard) was tested by checking improvement in prognostic Cancer
February 1, 2009
Table 3. Patient, Tumor, and Treatment Characteristics (N¼237)
Variable
No.
Table 3. Patient, Tumor, and Treatment Characteristics (N¼237) (continued)
%*
Sex 114 123
48 52
37 142 58
16 60 24
67 61 109
28 26 46
132 105
56 44
1 50 104
0.5 21 44
80
33.5
1 1
0.5 0.5
6 91 140
3 38 59
98 135 1 3
41 57 0.5 1.5
Age, y 70
Center UZ Leuven UCL Brussels Uni Koln
Pain No pain Pain or numbness
Treatment characteristics Surgery Local excision Superficial parotidectomy Total parotidectomy with facial nerve preservation Total parotidectomy without facial nerve preservation Debulking No surgery
Spill Yes No Not recorded
Treatment combination Surgery alone Surgery and postoperative RT RT alone RT and chemotherapy
Skin 92 8
192 45
81 19
47 96 46 47 1
20 40 20 20
168 17 3 43 6
71 7 1 18 3
17 7 77 138
17 7 76
46 12 3 176
19 5 1 75
cT classification (UICC 199231) cT1 cT2 cT3 cT4 Missing
cN classification (UICC 199231) cN0 cN1 cN2a cN2b cN3
Cytology Not suspect Acinic cell carcinoma Other malignancy NA
Tumor localization Superficial lobe Deep lobe Superficial and deep Not recorded
%*
184 48 5
78 29 2
82 122 33
35 51 14
184 48 5
78 20 2
38 37 32 3 6 1 0 0 0 3 11 42 0 27 20 0 15 2
16 16 14 1 3 0.5 0 0 0 1 5 18 0 11 8 0 6 1
62 175
26 74
65 172
27 73
237
87
30
13
100 16 45 10 66
42 7 19 4 28
Vascular invasion Yes No Not recorded
Positive margins No Yes Not recorded
Histologic classification (Seifert & Sobin 19926) Acinic cell carcinoma Mucoepidermoid carcinoma (low grade; n¼24) Adenoid cystic carcinoma Polymorphous low-grade adenocarcinoma Epithelial-myoepithelial carcinoma Basal cell adenocarcinoma Sebaceous carcinoma Papillary cystadenocarcinoma Mucinous adenocarcinoma Oncocytic carcinoma Salivary duct carcinoma Adenocarcinoma Malignant myoepithelioma Carcinoma in pleomorphic adenoma Squamous cell carcinoma Small cell carcinoma Undifferentiated carcinoma Other carcinomas
Grade (WHO 1972)y Low grade High grade
Grade (WHO 1991)z
Extracapsular growth 217 20
Facial nerve dysfunction Intact function Paresis-paralysis
No Yes Not recorded
Low grade High grade
Tumor characteristics No invasion Invasion
No.
Perineural growth
Patient characteristics Men Women
Variable
(continued)
No/no lymph nodes in neck specimen/no neck specimen Yes
Metastatic lymph nodes 0 1 Positive lymph node 2-5 Positive lymph nodes 6 Positive lymph nodes NA
UZ indicates University Hospitals Leuven; UCL, University Hospital Brussels; RT, radiotherapy; cT classification, clinical tumor classification; UICC, International Union Against Cancer; cN classification, clinical lymph node classification; WHO, World Health Organization; NA, not available (because of no surgery). * Of the total number with known values. y Grade WHO 1972: dichotomy ‘high grade’ (high-grade mucoepidermoid carcinoma, adenocarcinoma, undifferentiated carcinoma, squamous cell carcinoma, malignant mixed tumor, adenoid cystic carcinoma) vs ‘lowgrade’ group (acinic cell carcinoma, low-grade mucoepidermoid carcinoma). z Grade WHO 1991: dichotomy: ‘high grade’ (high-grade mucoepidermoid carcinoma, adenoid cystic carcinoma, salivary duct carcinoma, adenocarcinoma not otherwise specified, carcinoma in pleomorphic adenoma, small cell carcinoma, squamous cell carcinoma, and undifferentiated carcinoma) vs ‘low grade’ (acinic cell carcinoma, polymorphous low-grade adenocarcinoma, epithelial-myoepithelial carcinoma, and basal cell adenocarcinoma; see Seifert 199032).
Original Article
FIGURE 1. Recurrence-free interval by pretherapeutic prognostic index (PS1) groups. Numbers above the time axis refer to the number of patients who were at risk at that particular time during follow-up.
power by adding dummy variables, which indicated the original categorization of PS, to the linear term. Then, this linearity was evaluated by inspecting the plot of the deviance residuals against PS.10 Multiple comparisons adjustment of P values was done only where stated. Survival-type calculations were done by using the Kaplan-Meier method.7 For standard error (SE) calculation the method of Tsiatis was used.11
RESULTS Treatment Results The overall survival rate at 5 years was 76% (SE, 3%). The 5-year disease free rate was 69% (SE, 5%), and the 10-year disease-free survival rate was 58% (SE, 8%). A recurrence was observed in 68 of 237 patients (local, n ¼ 12 patients; regional, n ¼ 2 patients; distant, n ¼ 16 patients; locoregional, n ¼ 21 patients; locoregional and distant, n ¼ 17 patients). This corresponds to the findings in the source population and in the national validation sample with 64% and 62% of patients recurrence free at 5 years, respectively1,2 Clinical Validation PS1 could be calculated in 235 patients. It could not be calculated for 2 patients with 1 recurrence each because of missing values in cT classification. PS1 scores ranged 544
FIGURE 2. Recurrence-free interval by post-treatment prognostic index (PS2) groups. Numbers above the time axis refer to the number of patients who were at risk at that particular time during follow-up.
from 2.79 to 8.32. According to the published cutoff points (5.80),1 the observed 5-year recurrence-free rates were 94% for the most favorable group, 86% for the second best group, 58% for the third group, and 42% for the least favorable group. These values from this international validation sample are listed next to the corresponding values in the source population and in the national validation in Table 4. PS1-based recurrence-free interval curves per group are displayed in Figure 1, in which correct ordering can be verified. PS2 could be calculated for 230 patients and ranged from 3.16 to 10.73. For 7 patients who had 3 recurrences each, including the 1 patient without surgery, PS2 could not be calculated because of missing values. No evidence of a different recurrence rate in patients with missing PS2 was observed (P ¼ .72). By using published cutoff points (5.67), the 5-year recurrence-free rates ranged from 93%, to 84% for the second group, to 61% for the third group, and 40% for the least favorable group. Figure 2 displays the Kaplan-Meier recurrence-free interval curves for the 4 groups. Calibration can be assessed in Tables 4 and 5, which show the recurrence-free rates according to PS1 and PS2 levels in the international sample next to the corresponding percentages in the source population and in the nationwide sample. This illustrates over-optimism in estimating Cancer
February 1, 2009
Internationally Valid Prognostic Index/Vander Poorten et al
Table 4. Recurrence-free Percentages at 5 Years for Source and Validation Samples According to the 4 Levels of the Pretherapeutic Prognostic Index
PS1 Level
PS1 Source Population*
1 2 3 4
PS1 National Validationy
PS1 International Validationz
% (SE)
No.
% (SE)
No.
% (SE)
No.
92 83 48 23
29 30 31 28
92 70 59 42
40 56 60 27
94 86 58 42
44 74 61 56
(5) (7) (11) (9)
(7) (9) (11) (32)
(5) (5) (7) (7)
PS1 indicates the pretreatment prognostic index; SE, standard error. * Original sample from the Netherlands’ Cancer Institute from which PS1 and PS2 were derived. y National validation sample from the Dutch Cooperative Group on Head and Neck Cancer. z International validation sample.
Table 5. Recurrence-free Percentages at 5 Years for Source and Validation Samples According to the 4 Levels of the Post-therapeutic Prognostic Index
PS2 Level
PS2 Source Population*
1 2 3 4
PS2 National Validationy
PS2 International Validationz
% (SE)
No.
% (SE)
No.
% (SE)
No.
95 83 56 42
26 25 26 26
90 87 70 40
21 56 50 44
93 84 61 40
58 64 44 64
(5) (8) (12) (5)
(10) (7) (10) (16)
(4) (5) (8) (7)
PS2 indicates the post-treatment prognostic index; SE, standard error. * Original sample from the Netherlands’ Cancer Institute from which PS1 and PS2 were derived. y National validation sample from the Dutch Cooperative Group on Head and Neck Cancer. z International validation sample.
Table 6. Optimal Weights in the Pretherapeutic Prognostic Index (PS1) for the International Belgian-German Dataset
Variable
Original Weights (SE)
Univariate Optimal Change (SE)
P*
Multivariate Weights (SE)
Age Pain cT classification cN classification Skin invasion Facial nerve dysfunction
0.024 0.62 0.44 0.45 0.63 0.91
0.0140 0.99 þ0.34 0.04 0.21 0.10
.12 .0003 .04 .74 .54 .77
0.0286 0.53 0.59 0.230 0.39 0.42
(0.0087) (0.30) (0.17) (0.11) (0.34) (0.33)
(0.009) (0.27) (0.16) (0.11) (0.34) (0.33)
(0.0089) (0.26) (0.14) (0.091) (0.34) (0.28)
SE indicates standard error; cT classification, clinical tumor classification; cN classification, clinical lymph node classification. * Resulting from the Wald test.
prognostic power when a prognostic index is applied to its source population: The 4 groups in both national and international samples had less diverging prognoses.12 However, the estimated recurrence-free rates were comparable and were within the same range in both validation samples. Cancer
February 1, 2009
The amount of fit of the index to the new data was quantified by entering the newly calculated the prognostic index into a Cox model, resulting in a coefficient for PS1 of 0.63 (SE, 0.10) and a coefficient for PS2 of 0.71 (SE, 0.10), both highly significant (P
