Society for Endocrinology
Multivariable model versus AJCC staging system: cancer-specific survival predictions in adrenocortical carcinoma
Letizia Maria Ippolita Jannello01,2,3, Simone Morra1,4, Lukas Scheipner1,5, Andrea Baudo1,3,6, Carolin Siech1,7, Mario de Angelis1,8, Nawar Touma1, Zhe Tian1, Jordan A Goyal1, Stefano Luzzago2,9, Francesco A Mistretta2,9, Mattia Luca Piccinelli2,9, Fred Saad1, Felix K H Chun7, Alberto Briganti8, Sascha Ahyai5, Luca Carmignani6,10, Nicola Longo4, Ottavio de Cobelli2,9, Gennaro Musi2,9 and Pierre I Karakiewicz1
1Cancer Prognostics and Health Outcomes Unit, Division of Urology, University of Montréal Health Center, Montréal, Québec, Canada
2Department of Urology IEO European Institute of Oncology, IRCCS, Via Ripamonti, Milan, Italy 3Università degli Studi di Milano, Milan, Italy
4Department of Neurosciences, Science of Reproduction and Odontostomatology, University of Naples Federico II, Naples, Italy 5Department of Urology, Medical University of Graz, Graz, Austria
6Department of Urology, IRCCS Policlinico San Donato, Milan, Italy
7Department of Urology, University Hospital Frankfurt, Goethe University Frankfurt am Main, Frankfurt am Main, Germany
8Division of Experimental Oncology/Unit of Urology, URI, Urological Research Institute, IRCCS San Raffaele Scientific Institute, Milan, Italy 9Department of Oncology and Haemato-Oncology, Università degli Studi di Milano, Milan, Italy
10Department of Urology, IRCCS Ospedale Galeazzi - Sant’Ambrogio, Milan, Italy
Correspondence should be addressed to L M I Jannello: letizia.jannello@unimi.it
Abstract
We developed a novel contemporary population-based model for predicting cancer-specific survival (CSS) in adrenocortical carcinoma (ACC) patients and compared it with the established 8th edition of the American Joint Committee on Cancer staging system (AJCC). Within the Surveillance, Epidemiology, and End Results database (2004-2020), we identified 1056 ACC patients. Univariable Cox regression model addressed CSS. Harrell’s concordance index (C-index) quantified accuracy after 2000 bootstrap resamples for internal validation. The multivariable Cox regression model included the most informative, statistically significant predictors. Calibration and decision curve analyses (DCAs) tested the multivariable model as well as AJCC in head-to-head comparisons. Age at diagnosis (>60 vs ≤60 years), surgery, T, N, and M stages were included in the multivariable model. Multivariable model C-index for 3-year CSS prediction was 0.795 vs 0.757 for AJCC. Multivariable model outperformed AJCC in DCAs for the majority of possible CSS-predicted values. Both models exhibited similar calibration properties. Finally, the range of the multivariable model CSS predicted probabilities raged 0.02-75.3% versus only four single AJCC values, specifically 73.2% for stage I, 69.7% for stage II, 46.6% for stage III, and 15.5% for stage IV. The greatest benefit of the multivariable model-generated CSS probabilities applied to AJCC stage I and II patients. The multivariable model was more accurate than AJCC staging when CSS predictions represented the endpoint. Additionally, the multivariable
model outperformed AJCC in DCAs. Finally, the AJCC appeared to lag behind the multivariable model when discrimination addressed AJCC stage I and II patients.
Keywords: adrenocortical carcinoma; multivariable model; cancer-specific mortality; prognostic model
Introduction
Adrenocortical carcinoma (ACC) is a rare malignancy characterized by its high aggressiveness and heterogeneity (Fassnacht & Allolio 2009). Its incidence is approximately 0.72 per million cases per year, leading to 0.2% of all cancer deaths in the USA (Bourdeau et al. 2013). Clinicians use staging systems to predict the prognosis of ACC patients due to its rarity. In North America, the American Joint Committee on Cancer (AJCC) staging system is commonly used for this purpose (Amin et al. 2017). However, the clinical outcome widely differs for any given tumor stage (Panunzio et al. 2023). For instance, within the 8th edition of the AJCC, stage I patients may exhibit very favorable survival, conversely, very unfavorable survival affects AJCC stage IV patients (Amin et al. 2017). Between these two extremes, a variety of survival trajectories may exist. Nonetheless, the application of the four AJCC stages may not reflect the survival trajectories of individual patients in the most accurate fashion. Based on this hypothesis, we postulated that a multivariable model that relies on clinical and pathological variables may better predict cancer-specific survival (CSS), according to standard validation metrics. These testing criteria for any predictive and prognostic model, including multivariable models and AJCC, consist of Harrell’s concordance index (C-index), calibration, and decision curve analyses (DCAs). To the best of our knowledge, there have been only few studies conducted on the clinical prognostic model of ACC (Zini et al. 2009, Kim et al. 2016, Li et al. 2018, Zhang et al. 2020). However, the accuracy of their prognostic outcomes has not been entirely satisfactory or precise enough. We relied on the Surveillance Epidemiology and End Results (SEER) database to address this objective.
Materials and methods
Data source and study population
Within the SEER database (2004-2020), we identified newly diagnosed and histologically confirmed ACC patients (International Classification of Disease for Oncology (ICD-O-3) site code C74.0/C74.9; histologic code: ‘8370/3: Adrenal cortical carcinoma’, ‘8010/3: Carcinoma, NOS’, and ‘8140/3: Adenocarcinoma of the adrenal, NOS’) aged ≥18 years (https://seer.cancer.gov/index.html, accessed September 2023). Moreover, patients with complete data regarding vital status, T stage, N stage, M stage, and surgical resection status (performed vs not performed) were included (Fig. 1). Since SEER is entirely
anonymous, study-specific ethics approval was waived by the institutional review board.
Variables and outcome of interest
Three years of CSS (death from ACC) represented the endpoints of interest. For each patient the following covariates were recorded: age at diagnosis (<60 vs ≥60 years), sex (male vs female), race/ethnicity (Caucasian vs others), tumor size (continuously coded), T stage (T1 vs T2 vs T3 vs T4), N stage (N0 vs N1 vs NX), M stage (M0 vs M1), surgical resection status (performed vs not performed), chemotherapy delivered (yes vs no), and radiotherapy delivered (yes vs no)
Statistical analyses
First, we tabulated baseline patient and tumor characteristics. Descriptive statistics included frequencies and proportions for categorical variables. For continuously coded variables (i.e. age and tumor size) minimum P-value approach was used to identify the most informative cutoff. Subsequently, Kaplan- Meier plots as well as univariable Cox regression models assessing CSS were fitted. Only predictor variables that achieved statistical significance (P < 0.05) in
ACC patients (SEER 2004-2020) (n=1,432)
Unknown T or N or M stage (n=300)
ACC patients with complete stage information (SEER 2004-2020) (n=1,132)
Unknown tumor size (n=48)
ACC patients with complete stage and tumor information (SEER 2004-2020) (n=1,084)
Unknown adrenalectomy status (n=28)
ACC patients with complete stage, tumor and treatment information (SEER 2004-2020) (n=1,056)
univariable Cox regression models, were considered for inclusion in multivariable Cox regression models. Here, backward stepwise selection was applied using Akaike’s Information Criterion (AIC) rules. The intent was to generate the most accurate and most parsimonious multivariable model predicting CSS (Harrell et al. 1996). The accuracy of CSS predictions for individual variables as well as the multivariable models was quantified using C-index (Heagerty & Zheng 2005). A C-index of 0.5 indicates random predictions, whereas a C-index of 1.0 indicates perfect predictions. The predictive accuracy of individual variables as well as multivariable models was internally validated using 2000 bootstrap resamples (Carpenter & Bithell 2000).
Finally, the most accurate and parsimonious multivariable model predicting CSS as well as AJCC was subjected to standard testing metrics for predictive and prognostic tools in a head-to-head fashion, namely, calibration and DCAs. Calibration plots quantified the difference between predicted CSS vs observed CSS. DCAs graphically plotted the performance of AJCC-derived CSS predictions relative to multivariable model-derived CSS predictions.
All statistical tests were two-sided, with the level of significance set at P < 0.05, and were performed with R Software Environment for Statistical Computing and Graphics (R version 4.1.3; https://www.r-project.org/).
Results
Descriptive characteristics
Of all 1056 ACC patients identified between 2004 and 2020, 644 (61%) were female, and 737 (70%) were Caucasian. The median age at diagnosis was 56 years (interquartile range (IQR) 44-65 years). The median tumor size was 11.0 cm (IQR 7.8-15.0 cm). The median follow-up was 18 months (IQR 6-52 months). Surgery was performed in 827 (78%) patients. Chemotherapy was used in 475 (45%) patients and radiotherapy was delivered in 157 (15%) patients (Table 1).
At 3 years of follow-up, CSS rates were 45.8% for all ACC patients across all four examined AJCC stages (Fig. 2A). After stratification, according to AJCC stages I to IV, 3-year CSS rates were 73.2% for stage I, 69.7% for stage II, 46.6% for stage III, and 15.5% for stage IV (Fig. 2B).
Univariable and multivariable CSS models
The accuracy of AJCC stages according to the C-index in CSS predictions at 3 years of follow-up was 0.757 after 2000 bootstrap resamples for internal validation. In univariate Cox regression models, the most accurate predictors of CSS at 3 years of follow-up and after 2000 bootstrap resamples were M stage (0.701), T stage (0.661), surgical resection status (0.652), N stage (0.577), chemotherapy status (0.563), age at diagnosis (0.553),
radiotherapy status (0.539), and tumor size (0.514). The most accurate and most parsimonious multivariable model predicting CSS relied on age at diagnosis, surgical resection status, T stage, N stage, and M stage. According to the C-index, its accuracy in CSS prediction at 3 years of follow-up was 0.795 after 2000 bootstrap resamples for internal validation (Table 2 and Fig. 3).
Calibration plots depicted observed vs predicted CSS: AJCC vs multivariable model
Calibration plots tested the agreement between predicted CSS vs observed CSS for AJCC stages as well as for the multivariable model. At 3 years of follow-up, AJCC predicted CSS was invariably higher than AJCC observed CSS. Specifically, predicted CSS for stage I was 76.0% vs 67.0 observed CSS. The latter resulted in a difference of 9.0%. For stage II, predicted CSS was 72.7 vs 67.5 % observed CSS, resulting in a difference of 5.2% For stage III, predicted CSS was 54.0 vs 47.2 % observed CSS, resulting in a difference of 6.9%. Finally, for stage IV predicted CSS was 16.4 vs 15.6% observed CSS, resulting in a difference of 0.8%.
The same methodology was applied to the multivariable model. Here, four equally sized groups were generated according to CSS probabilities (high, intermediate, low, and very low risk). At 3 years of follow-up, the differences between the multivariable model predicted CSS relative to the multivariable model observed CSS values were higher in the intermediate-risk group, followed by the low-risk
| n = 1056 | |
|---|---|
| Tumor size (cm), median (IQR) | 11.0 (7.8-15.0) |
| Age at diagnosis (≤60 years) | 669 (63%) |
| Sex (female) | 644 (61%) |
| Race/ethnicity (Caucasian) | 737 (70%) |
| T stage, n (%) | |
| T1 | 120 (12%) |
| T2 | 360 (34%) |
| T3 | 278 (26%) |
| T4 | 298 (28%) |
| N stage, n (%) | |
| N0 | 871 (82%) |
| N1 | 112 (11%) |
| NX | 73 (7%) |
| M stage, n (%) | |
| M0 | 701 (66%) |
| M1 | 355 (34%) |
| Surgery performed, n (%) | 827 (78%) |
| Chemotherapy delivered, n (%) | 475 (45%) |
| Radiotherapy delivered, n (%) | 157 (15%) |
IQR, interquartile range.
A
CSS + Overall population
1.00
+
+
Three-year CSS: 45.8%
+
0.75
Median CSS: 31 months
Survival Probability
0.50
0.25
0.00
0
6
12
18
24
30
36
42
48
54
60
Months
Number at risk
CSS
All
1056
804
642
531
451
388
339
306
277
262
241
0
6
12
18
24
30
36
42
48
54
60
Months
B
8th AJCC
1.00
Three-year CSS:
+
+
+
Stage I: 73.2%
+
+
Stage
Stage II: 69.7%
+
+
Stage Ih
Stage III: 46.6%
+
Stage IV: 15.5%
0.75
+
+
+
+
+
+
+
+
+
Survival Probability
+
+
+
Stage
+
+
+
+
+
0.50
+
+
Stage IV
0.25
p < 0.0001
+
+
+
+
0.00
0
6
12
18
24
30
36
42
48
54
60
Months
Number at risk
I
AJCC8th
84
76
67
58
48
38
30
29
27
24
19
II
268
249
230
218
197
184
172
162
150
144
141
III
322
267
208
158
128
104
90
77
69
66
56
IV
309
164
103
71
54
42
29
24
20
19
16
0
6
12
18
24
30
36
42
48
54
60
Months
groups, and the high-risk group, whereas it was inverted in the very low-risk group. Specifically, for the high-risk group the predicted CSS was 74.6% vs 74.5% observed CSS. The latter resulted in a difference of 0.1%. For the intermediate-risk group, the predicted CSS was 63.6% vs 55.2% observed CSS, resulting in a difference of 9.1%. For the low-risk group, the predicted CSS was 44.6% vs 35.5% observed CSS, resulting in a difference of 8.4%. Finally, for the very low-risk group predicted CSS was 9.1% vs 10.1% observed CSS, resulting in a difference of 1.0% (Fig. 4).
Decision curve analyses for CSS predictions: AJCC vs multivariable model
DCAs plotted the net benefit of AJCC stages as well as of the novel multivariable model across all possible CSS predicted values. Here, the multivariable model outperformed AJCC for the majority of possible CSS- predicted values. Namely, a greater net benefit resulted from CSS predicted values ranging from 0.30 to 0.34, and then ranging from 0.45 to 0.88 (Fig. 5).
| Variables tested | Univariable | Multivariable | ||||
|---|---|---|---|---|---|---|
| HR (95% CI) | P | 36-months C-index | HR (95% CI) | P | 36-month C-index | |
| Eighth edition AJCC (I) | Ref | 0.757 | – | – | 0.795 | |
| II | 1.17 (0.75-1.83) | 0.492 | – | – | ||
| III | 2.25 (1.45-3.48) | <0.001 | – | – | ||
| IV | 6.58 (4.28-10.12) | <0.001 | – | – | ||
| Age at diagnosis (>60 vs ≤60) | 1.42 (1.20-1.67) | <0.001 | 0.553 | 1.46 (1.23-1.73) | <0.001 | |
| Tumor size (cm) | 1.00 (1.00-1.00) | 0.784 | 0.514 | – | – | |
| Sex (male vs female) | 1.04 (0.88-1.23) | 0.653 | 0.479 | – | – | |
| T stage (T1) | Ref | 0.661 | Ref | |||
| T2 | 1.17 (0.83-1.64) | 0.365 | 1.25 (0.89-1.75) | 0.204 | ||
| T3 | 1.87 (1.33-2.64) | <0.001 | 1.97 (1.39-2.80) | <0.001 | ||
| T4 | 3.13 (2.23-4.38) | <0.001 | 1.75 (1.24-2.48) | 0.002 | ||
| N stage (N0) | Ref | 0.577 | Ref | |||
| N1 | 2.93 (2.33-3.69) | <0.001 | 1.55 (1.21-1.99) | <0.001 | ||
| NX | 1.54 (1.13-2.11) | 0.007 | 0.79 (0.57-1.10) | 0.160 | ||
| M stage (M1 vs M0) | 4.10 (3.47-4.85) | <0.001 | 0.701 | 2.43 (1.96-3.01) | <0.001 | |
| Surgery (yes vs no) | 0.20 (0.16-0.24) | <0.001 | 0.652 | 0.37 (0.29-0.47) | <0.001 | |
| Chemotherapy (yes vs no) | 1.37 (1.16-1.61) | <0.001 | 0.563 | - | - | |
| Radiotherapy (yes vs no) | 0.63 (0.48-0.81) | <0.001 | 0.539 | - | - | |
Values in bold indicate statistical significance.
Comparison of predicted probabilities: AJCC vs multivariable model
At 3 years AJCC offers four specific CSS predicted values, these were 73.2% for stage I, 69.7% for stage II, 46.6% for stage III, and 15.5% for stage IV. Conversely, the multivariable model demonstrated a wide range of predicted CSS values from 0.02 to 75.3%. Moreover, when AJCC stages were plotted within the multivariable model, the distributions of CSS values at 3 years recorded important overlap between stages I and II as evidenced in IQR of 66.1 to 75.3% for stage I, and 59.7-70.2% for
stage II. No overlap in IQR ranges was recorded for stages III and IV (Fig. 6).
Discussion
Currently, AJCC represents the standard of care for risk stratification in ACC patients (Amin et al. 2017). However, it does not provide a specific CSS estimate for each of its stages. Instead, it only stratifies ACC patients according to stage I vs stage II vs stage III vs stage IV. Based on this limitation regarding the consideration of
0
10
20
30
40
50
60
70
80
90
100
Points
Age at diagnosis
>60
≤60
Surgery
NO
YES
T stage
T2
T3
T1
T4
N stage
N0
NX
N1
M stage
M1
M0
Total points
0
50
100
150
200
250
300
350
400
36 months
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
A
2017 8th edition of the AJCC-staging system
1.0
Stage I: n=90 (9%)
0.8
Stage II: n=280 (27%)
Stage III: n=336 (32%)
Observed probability
Stage IV: n=333 (32%)
I
0.6
0.4
0.2
c-index: 0.757
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Predicted probability
B
Multivariable model
1.0
High risk group: n=264 (25%)
0.8
Intermediate risk group: n=264 (25%)
Low risk group: n=264 (25%)
Observed probability
Very low risk group: n=264 (25%)
High risk
0.6
Intermediate risk
0.4
Low risk
0.2
Very low risk
c-index:0.795
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Predicted probability
baseline risk factors in ACC patients (that are restricted to AJCC staging I vs II vs III vs IV) we postulated that a multivariable modeling approach of CSS relying on stages as well as other patient characteristics may yield better predictions than AJCC alone. We tested this hypothesis according to standard testing criteria (accuracy, calibration, and DCA) and made several noteworthy observations.
First, ACC is a rare malignancy. Its annual incidence in North America is 0.7 per million (Fassnacht & Allolio 2009). Staging systems are more important in rare malignancies than in more prevalent ones since clinicians tend to be less familiar with patients’ prognoses. Unfortunately, due to few available patients for clinical and epidemiological studies, it is difficult to improve staging and prognostic
models in rare malignancies. This situation is very different for more prevalent malignancies, where staging and prognostication are continuously updated and improved, such as in prostate (Hittelman et al. 2004, Ploussard et al. 2020), lung (Rimner et al. 2023), and colon cancers (Banias et al. 2022). To date, AJCC represents the staging standard of care in North American ACC patients (Amin et al. 2017). While AJCC may stratify patients according to stage-specific risk levels, it does not provide an estimate of survival for each of its four stages. This hypothesis validates the rationale for the current study. The latter postulates that more accurate and specific prognostic information may be obtained with the use of a prognostic model that relies on stages as well as surgical resection status and patient age at diagnosis.
0.5
- None
All
Multivariable_model
AJCC8th
0.4
Net benefit
0.3
0.2
0.1
0.0
0.0
0.2
0.4
0.6
0.8
Threshold probability
Second, our analyses relied on the most contemporary (2004-2020) SEER cohort of ACC patients. To date, databases used for testing multivariable models predicting CSS or other cancer controlled outcomes in ACC patients predominantly relied on historical patients. These patients were diagnosed and managed before the year 2000. Therefore, they may no longer be current. All available population-based studies addressing ACC prognosis relied on the SEER database. For instance, Zhang et al. studied 855 ACC patients (Zhang et al. 2020), Li et al. studied 751 ACC patients (Li et al. 2018), and Kong et al. studied 722 ACC patients (Kong et al. 2019). Interestingly, multi-institutional studies addressing ACC provided even smaller sample sizes. For example, the largest multi-institutional study (Kim et al. 2016) only provided 148 ACC patients, over 20 years (1994-2014) from within 13 major North American institutions (Kim et al. 2016). Given these considerations, the rare entity,
such as ACC, should ideally be investigated within the framework of even larger multi-institutional databases. Moreover, those observations also validate the hypothesis and the goal of the current study, which aims to develop a contemporary prognostic multivariable model.
Third, our study provided a head-to-head comparison of AJCC staging versus a multivariable approach for the prediction of CSS at 3 years of follow-up after ACC diagnosis. The first testing benchmark consisted of accuracy. Here, we relied on bootstrap resampling to simulate internal validation. Bootstrap resampling represents a widely used substitute for internal validation (Karakiewicz et al. 2007, Dell’Oglio et al. 2016, Nazzani et al. 2018). After internal validation using 2000 bootstrap resamples CSS predictions at 3 years resulted in a C-index of 0.757 for AJCC vs. 0.795 for the multivariable approach that relied on T, N, and M stages
Multivariable model prediction at 3 years
0.6
o
0.4
O
o
0
0.2
0
000
O
0.0
Stage I
Stage II
Stage III
Stage IV
8th edition of the AJCC staging system
as well as surgical resection status and patient age at diagnosis. Higher accuracy of the novel multivariable approach relative to AJCC indicates that the building blocks of AJCC staging, namely, T, N, and M stages, clearly should represent the backbones of any predictive or prognostic models. However, our observations also indicate that a substantial improvement in accuracy may be achieved with simple and universally available variables, such as patient age at diagnosis and surgical resection status. In consequence, additional studies, that replicate the methodological steps used in the current study, should ideally test the concept that additional variables, especially those portraying specific tumor characteristics, may further improve the accuracy of CSS prediction. Unfortunately, the lack of sufficient detail within the SEER database prevented us from including other potential predictors of CSS. Although the National Cancer Database (NCDB) provides an even larger cohort of ACC patients than the SEER database, lack of cancer- specific mortality data represents a critical limitation that prevents its use, when patients with non-metastatic disease are included since such individuals often die of other causes than ACC. However, the distinction between cancer-specific and other-cause mortality is not available within that database. In consequence, the use of a large multi-institutional database to circumvent that limitation should be encouraged and prioritized in ACC prognostic research.
Fourth, the calibration plot represented the second testing metric for both AJCC and the novel multivariable prognostic model. Here, predicted CSS vs. observed CSS values at 3 years were plotted for each approach. For all four AJCC stages, at 3 years of follow-up predicted CSS was invariably higher than observed CSS. Four equally sized groups based on CSS predictions were created for the multivariable prognostic model to allow the most objective and direct comparison between the multivariable model and AJCC-generated predictions. Here, plots displayed predicted CSS vs. observed CSS, showed a higher difference for the high, intermediate, and low-risk groups according to CSS probabilities and a lower difference for the very low-risk group. In consequence, according to the calibration metric, both models performed relatively equally well. Therefore, accuracy and DCA results indicate an advantage of the multivariable prognostic model. In consequence, the multivariable prognostic model should be used instead of AJCC when CSS predictions are required.
Fifth, in the final step of testing AJCC and the multivariable prognostic model were tested in DCAs. DCAs represent a valuable approach when comparing the net benefit of one prognostic approach versus an alternative. In the current study, the multivariable prognostic model provided a higher net benefit than AJCC, across the majority of possible CSS predicted values. Specifically, the multivariable approach outperformed AJCC for CSS predicted values ranging from 0.30-0.34 and 0.45- 0.88. It is important to notice that the vast majority
of multivariable model-predicted CSS values were situated in the 0.45-0.88 range. In consequence, the notable advantage of the multivariable model was predominantly applicable within a population where its net benefit was the greatest.
Taken together, the current study demonstrates that a multivariable prognostic model that relied on T, N, and M stages, surgical resection status, and patient age at diagnosis, provides more accurate predictions than AJCC. Moreover, the multivariable prognostic model also outperformed AJCC according to DCA criteria. Conversely, both models performed similarly well according to calibration. Based on these observations, it may be postulated that the multivariable prognostic model offers an advantage when CSS predictions are required. The multivariable prognostic model may be further improved regarding its accuracy, net benefit in DCAs, and possibly its calibration properties, with the inclusion of other variables that further define the host and the tumor. Testing of such variables should be undertaken based on multi-institutional databases. Unfortunately, more detailed testing using the SEER database is not possible due to a lack of additional variables, that would be considered for inclusion. Such variables could be the Weiss score (Sohail et al. 2021), biochemical and pathological values, such as the Ki-67 score (Elhassan et al. 2021), lymphovascular invasion (Luconi et al. 2023), and the presence of tumor necrosis (Yi et al. 2022). Moreover, the extent of surgical dissection, including the resection of metastases may also be of benefit. Similarly, further testing using the NCDB is also not possible, due to the lack of cancer-specific mortality information within that large database. In consequence, future prognostic research in ACC will invariably require multi-institutional collaboration.
Last but not least, the current multivariable prognostic model should be externally validated within an independent cohort. However, an independent cohort with a sufficient sample size from a different database may be difficult to identify. Additional limitations consist of the retrospective nature of the data that were used for the multivariable model development.
Conclusion
The multivariable model was more accurate than AJCC staging when CSS predictions represented the endpoint. Additionally, multivariable model CSS predictions outperform AJCC in DCAs. Finally, the AJCC appeared to lag behind the multivariable model when discrimination addressed AJCC stage I and II patients.
Declaration of interest
The authors declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of the study reported.
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
Author contribution statement
Conceptualization and Methodology: LMIJ, SM, LS; Acquisition of Data: LMIJ, AB (Andrea Baudo), CS, MdA; Formal analysis: LMIJ, ZT; Investigation and Data Curation: LMIJ, SL, FAM, MLP; Writing - Original Draft: LMIJ, SM, JAG; Visualization: FS, SFS, LC, SA, AB (Alberto Briganti), FKHC; Funding acquisition: N/A; Supervision: OdC, GM, PIK; Project Administration: PIK.
References
Amin MB, Edge SB, Greene FL, Byrd DR, Brookland RK, Washington MK, Gershenwald JE, Compton CC, Hess KR, Sullivan DC, et al. 2017 AJCC cancer staging manual. 8th ed, pp. 911-918. New York, NY, USA: Springer. Banias L, Jung I, Chiciudean R & Gurzu S 2022 From dukes-MAC staging system to molecular classification: evolving concepts in colorectal cancer. International Journal of Molecular Sciences 23 9455. (https://doi. org/10.3390/ijms23169455)
Bourdeau I, Mackenzie-Feder J & Lacroix A 2013 Recent advances in adrenocortical carcinoma in adults. Current Opinion in Endocrinology, Diabetes, and Obesity 20 192-197. (https://doi.org/10.1097/ MED.0b013e3283602274)
Carpenter J & Bithell J 2000 Bootstrap confidence intervals: when, which, what? A practical guide for medical statisticians. Statistics in Medicine 19 1141-1164. (https://doi.org/10.1002/(sici)1097- 0258(20000515)19:9<1141:aid-sim479>3.0.co;2-f)
Dell’Oglio P, Suardi N, Boorjian SA, Fossati N, Gandaglia G, Tian Z, Moschini M, Capitanio U, Karakiewicz PI, Montorsi F, et al. 2016 Predicting survival of men with recurrent prostate cancer after radical prostatectomy. European Journal of Cancer 54 27-34. (https://doi. org/10.1016/j.ejca.2015.11.004)
Elhassan YS, Altieri B, Berhane S, Cosentini D, Calabrese A, Haissaguerre M, Kastelan D, Fragoso MCBV, Bertherat J, Al Ghuzlan A, et al. 2021 S-GRAS score for prognostic classification of adrenocortical carcinoma: an international, multicenter ENSAT study. European Journal of Endocrinology 186 25-36. (https://doi.org/10.1530/EJE-21-0510)
Fassnacht M & Allolio B 2009 Clinical management of adrenocortical carcinoma. Best Practice and Research. Clinical Endocrinology and Metabolism 23 273-289. (https://doi.org/10.1016/j.beem.2008.10.008)
Harrell FEJ, Lee KL & Mark DB 1996 Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Statistics in Medicine 15 361-387. (https://doi.org/10.1002/(SICI)1097-0258(19960229)15:4<361 :: AID- SIM168>3.0.CO;2-4)
Heagerty PJ & Zheng Y 2005 Survival model predictive accuracy and ROC curves. Biometrics 61 92-105. (https://doi. org/10.1111/j.0006-341X.2005.030814.x)
Hittelman AB, Purohit RS & Kane CJ 2004 Update of staging and risk assessment for prostate cancer patients. Current Opinion in Urology 14 163-170. (https://doi.org/10.1097/00042307-200405000-00005)
Karakiewicz PI, Briganti A, Chun FKH, Trinh QD, Perrotte P, Ficarra V, Cindol L, De la Taille A, Tostain J, Mulders PFA, et al. 2007 Multi- institutional validation of a new renal cancer-specific survival nomogram. Journal of Clinical Oncology 25 1316-1322. (https://doi. org/10.1200/JCO.2006.06.1218)
Kim Y, Margonis GA, Prescott JD, Tran TB, Postlewait LM, Maithel SK, Wang TS, Evans DB, Hatzaras I, Shenoy R, et al. 2016 Nomograms to predict recurrence-free and overall survival after curative resection of adrenocortical carcinoma. JAMA Surgery 151 365-373. (https://doi. org/10.1001/jamasurg.2015.4516)
Kong J, Zheng J, Cai J, Wu S, Diao X, Xie W, Chen X, Liao C, Yu H, Fan X, et al. 2019 A nomogram for individualized estimation of survival among adult patients with adrenocortical carcinoma after surgery: a retrospective analysis and multicenter validation study. Cancer Communications 39 80. (https://doi.org/10.1186/s40880-019-0426-0)
Li Y, Bian X, Ouyang J, Wei S, He M & Luo Z 2018 Nomograms to predict overall survival and cancer-specific survival in patients with adrenocortical carcinoma. Cancer Management and Research 10 6949-6959. (https://doi.org/10.2147/CMAR.S187169)
Luconi M, Cantini G, van Leeuwaarde RS, Roebaar R, Fei L, Propato AP, Santi R, Ercolino T, Mannelli M, Canu L, et al. 2023 Prognostic value of microscopic tumor necrosis in adrenal cortical carcinoma. Endocrine Pathology 34 224-233. (https://doi.org/10.1007/s12022-023-09760-6)
Nazzani S, Bandini M, Marchioni M, Preisser F, Tian Z, Soulières D, Montanari E, Motta G, Acquati P, Briganti A, et al. 2018 A contemporary analysis of radiotherapy effect in surgically treated retroperitoneal sarcoma. Radiotherapy and Oncology 127 318-325. (https://doi. org/10.1016/j.radonc.2018.03.027)
Panunzio A, Barletta F, Tappero S, Cano Garcia C, Piccinelli M, Incesu RB, Law KW, Tian Z, Tafuri A, Tilki D, et al. 2023 Contemporary conditional cancer-specific survival rates in surgically treated adrenocortical carcinoma patients: a stage-specific analysis. Journal of Surgical Oncology 127 560-567. (https://doi.org/10.1002/jso.27161)
Ploussard G, Manceau C, Beauval JB, Lesourd M, Almeras C, Gautier JR, Loison G, Salin A, Soulié M, Tollon C, et al. 2020 Decreased accuracy of the prostate cancer EAU risk group classification in the era of imaging- guided diagnostic pathway: proposal for a new classification based on MRI-targeted biopsies and early oncologic outcomes after surgery. World Journal of Urology 38 2493-2500. (https://doi.org/10.1007/s00345- 019-03053-6)
Rimner A, Ruffini E, Cilento V, Goren E, Ahmad U, Appel S, Bille A, Boubia S, Brambilla C, Cangir AK, et al. 2023 The International Association for the Study of Lung Cancer thymic epithelial tumors staging project: an overview of the central database informing revision of the forthcoming (ninth) edition of the TNM classification of malignant tumors. Journal of Thoracic Oncology 18 1386-1398. (https://doi. org/10.1016/j.jtho.2023.07.008)
Sohail S, Azmat U, Khawaja S, Mir K, Azam M, Syed M, Mushtaq S, Hassan U, Siddiqui A, Raza SA, et al. 2021 Clinical and histopathological variables and prognostic factors of adrenocortical carcinoma. Cureus 13 e15721. (https://doi.org/10.7759/cureus.15721)
Yi X, Wan Y, Cao W, Peng K, Li X & Liao W 2022 Identification of four novel prognostic biomarkers and construction of two nomograms in adrenocortical carcinoma: a multi-omics data study via bioinformatics and machine learning methods. Frontiers in Molecular Biosciences 9 878073. (https://doi.org/10.3389/fmolb.2022.878073)
Zhang H, Naji Y, Yan M, Lian W, Xie M & Dai Y 2020 Development and validation of prognostic nomograms in patients with adrenocortical carcinoma: a population-based study. International Urology and Nephrology 52 1057-1071. (https://doi.org/10.1007/s11255-020-02413-1)
Zini L, Capitanio U, Jeldres C, Lughezzani G, Sun M, Shariat SF, Isbarn H, Arjane P, Widmer H, Perrotte P, et al. 2009 External validation of a nomogram predicting mortality in patients with adrenocortical carcinoma. BJU International 104 1661-1667. (https://doi. org/10.1111/j.1464-410X.2009.08660.x)