EJE
External validation of the S-GRAS score for predicting recurrence in patients with adrenocortical carcinoma: implications for adjuvant mitotane therapy
Paula Jimenez-Fonseca, 10D Cristina Alvarez-Escola,20D Inmaculada Ballester Navarro,3 Jorge Hernando Cubero,40D Laura González Fernández,5 Miguel Ángel Mangas Cruz,6 Clara Iglesias,10 Jesus Garcia-Donas,70D Maria Jose Picon,&D Miguel Paja,9[D
Lorena Gonzalez Batanero, 100D Lourdes Garcia, 11[D Javier Molina, 12[D Raquel Jimeno Mate, 13 [D Javier Aller, 14D Maria del Castillo Tous Romero, 15 Jersy Cardenas Salas, 16D Gala Gutierrez-Buey,170 Nerea Egaña Zunzunegui,18 Miguel Navarro,1 19 İD
Maria Jose Lecumberri,2º Nuria Valdes,21(D and Alberto Carmona-Bayonas22,* (D
1Medical Oncology Department, Hospital Universitario Central de Asturias, ISPA, Oviedo, 33011, Spain
2Endocrinology and Nutrition Department, Hospital Universitario La Paz, Madrid, 28046, Spain
3Medical Oncology Department, Hospital Universitario Morales Meseguer, Murcia, 33008, Spain
4Medical Oncology Department, Gastrointestinal and Endocrine Tumors Unit, Hospital Universitario Vall d’Hebrón, Vall d’Hebron Institute of Oncology (VHIO), Barcelona, 08035, Spain
5Endocrinology and Nutrition Department, Hospital General Universitario Gregorio Marañon, Madrid, 28007, Spain 6Endocrinology and Nutrition Department, Hospital Universitario Virgen del Rocío, Sevilla, 41013, Spain
7Medical Oncology Department, Centro Integral Oncológico HM Clara Campal, Madrid, 28050, Spain
8Endocrinology and Nutrition Department, Hospital Universitario Virgen de La Victoria, Biomedical Research Institute-IBIMA, Málaga, CIBER Pathophysiology of Obesity and Nutrition-CIBEROBN, Madrid, 29010, Spain
9Endocrinology Department, Hospital Universitario de Basurto, Bilbao, University of the Basque Country, University of Pais Vasco (UPV/EHU), 48013, Spain
10Medical Oncology Department, Hospital Universitario de Canarias, Tenerife, 38320, Spain
11Endocrinology and Nutrition Department, Hospital Universitario de Jerez, Jerez, 11407, Spain
12Medical Oncology Department, Hospital Universitario Ramón y Cajal, Madrid, 28034, Spain
13Medical Oncology Department, Hospital Universitario Marqués de Valdecilla, Instituto de Investigación Valdecilla (IDIVAL), Universidad de Cantabria (UNICAN), Santander, 39008, Spain
14Endocrinology and Nutrition Department, Hospital Universitario Puerta de Hierro, Madrid, 28222, Spain
15Endocrinology and Nutrition Department, Hospital Universitario Virgen de la Macarena, Sevilla, 41009, Spain
16Endocrinology and Nutrition Department, Fundación Jiménez Díaz, Madrid, 28040, Spain
17Endocrinology and Nutrition Department, Hospital Universitario de Cabueñes, Gijón, 33394, Spain
18Endocrinology and Nutrition Department, Hospital Donostia, San Sebastian, 20014, Spain
19Medical Oncology Department, Complejo Asistencial Universitario de Salamanca, Salamanca, 37007, Spain
20Medical Oncology Department, Hospital de Navarra, Pamplona, 31008, Spain
21 Endocrinology and Nutrition Department, Hospital Universitario de Cruces, Biobizkaia, University of Pais Vasco (UPV/EHU), CIBERDEM, CIBERER, Endo-ERN, Barakaldo, 48903, Spain
22Medical Oncology Department, Hospital Universitario Morales Meseguer, Instituto Murciano de Investigación Biosanitaria (IMIB), Universidad de Murcia (UMU), Calle Marqués de los Vélez s/n, Murcia 3008, Spain
*Corresponding author: Medical Oncology Department, Hospital Universitario Morales Meseguer, Instituto Murciano de Investigación Biosanitaria (IMIB), Universidad de Murcia (UMU), Calle Marqués de los Vélez s/n, Murcia 3007, Spain. Email: alberto.carmonabayonas@gmail.com
Abstract
Background: Adrenocortical carcinoma (ACC) is a rare, aggressive malignancy with variable outcomes post-adrenalectomy. The S-GRAS score integrates 5 clinical and pathological factors to predict prognosis but requires external validation in diverse settings.
Methods: We validated the S-GRAS score in 138 ACC patients from the Spanish ICARO-GETTHI/SEEN registry (1998-2023). Model performance was assessed using discrimination, calibration, and accuracy. Exploratory refinements included non-linear modeling of age and Ki-67% and an expanded model incorporating venous invasion and tumor size. Cox models examined the interaction between S-GRAS and adjuvant mitotane.
Results: A total of 76 recurrence events were recorded. The S-GRAS score demonstrated good discrimination for overall survival (C-index 0.706, 95% CI, 0.628-0.785) and recurrence-free survival (C-index 0.673, 95% CI, 0.601-0.745) with well-calibrated predictions. Five-year survival rates differed significantly across score groups: 100% for scores 0-1, 81.6% for 2-3, 55% for 4-5, and 33.8% for 6-7. Non-linear modeling of Ki-67%
improved performance (C-index 0.738 for RFS, 0.761 for OS), but adding clinical variables offered minimal benefit, leaving 75% of recurrence variability unexplained. Higher S-GRAS scores correlated with increased mitotane benefit (HR 0.57, 95% CI, 0.34-0.97 for score 4; HR 0.46, 95% CI, 0.23-0.94 for score 5), indicating a potential incremental benefit pattern.
Conclusions: Our findings validate the S-GRAS score in a multicenter cohort and support its use in identifying candidates for adjuvant mitotane. Non-linear modeling of Ki-67% enhances predictive precision without increasing complexity, but the performance plateau of clinical variables suggests that integrating molecular biomarkers may be necessary to improve prognostic accuracy.
Keywords: adrenocortical carcinoma, S-GRAS score, Ki-67% index, prognosis, survival, non-linear modelling, mitotane
Significance
Adrenocortical carcinoma (ACC) prognosis after surgery is highly variable. In a large Iberian cohort, we externally validated the ENSAT-derived S-GRAS score, confirming its robust performance for predicting survival and recurrence. S-GRAS also suggested patterns in adjuvant mitotane outcomes that may help guide risk-adapted care. Using Ki-67 as a continuous vari- able via splines enhanced precision, while adding extra clinical predictors yielded limited gains, suggesting a ceiling for pure- ly clinical models. Our findings support S-GRAS as a practical tool for guiding follow-up intensity, patient counseling, and adjuvant therapy decisions in ACC, and motivate integration of molecular biomarkers to refine risk and treatment strategies in endocrine malignancies.
Introduction
Adrenocortical carcinoma (ACC) is a rare and aggressive ma- lignancy, traditionally associated with high recurrence rates following adrenalectomy.1,2 However, this characterization oversimplifies the clinical heterogeneity of ACC, which spans from slow-progressing to highly aggressive forms. For in- stance, in the ADIUVO trial, low-risk patients exhibited recur- rence rates below 25% at 8 years, whereas those with high-risk features-such as nodal involvement, positive surgi- cal margins, or high-grade histology-experienced recurrence rates exceeding 75%.2,3 To address this variability, recent ef- forts have focused on designing reliable prognostic tools to en- hance patient stratification and guide risk-adapted adjuvant treatment strategies.2,4,5 Over time, this work culminated in the development of the S-GRAS score by the European Network for the Study of Adrenal Tumors (ENSAT), which incorporates 5 clinical and pathological parameters: stage (ENSAT classification), grade (Ki-67% index), resection sta- tus (surgical margins), age at diagnosis, and symptoms of tu- mor or hormone hypersecretion.6,
Although initial studies have attempted to validate the S-GRAS score, 8,9 its predictive accuracy and reliability require further confirmation, particularly given the rarity of ACC and the variability in healthcare settings.10 Such validation could establish S-GRAS as a robust tool for guiding clinical decisions in high-risk patients, especially those who may benefit from in- tensive monitoring or adjuvant therapies such as mitotane or chemotherapy.11,12
Given this background, the primary objective of this study was to validate the prognostic performance of the S-GRAS score in an independent cohort of patients with ACC who underwent adrenalectomy. We also explored potential refine- ments, such as incorporating novel covariates or extracting additional information from continuous variables via non- linear modeling. Finally, we evaluated the effect of adjuvant mitotane at different S-GRAS score levels.
Methods
Study design and patients
This multicenter, retrospective observational study included patients from the ICARO-GETTHI/SEEN registry, an
initiative of the Spanish Group for Transversal Oncology and Rare and Orphan Tumors (GETTHI) and the Spanish Society of Endocrinology and Nutrition (SEEN). Recruitment was conducted consecutively to minimize selec- tion bias. The patient cohort was recruited from a network of 38 Spanish hospitals.
This study was conducted and reported in accordance with the Transparent Reporting of a multivariable prediction mod- el for Individual Prognosis or Diagnosis statement13 (Supplementary Material).
Eligibility criteria included a histological diagnosis of ACC, age over 18 years, and having undergone adrenalectomy. Patients with insufficient follow-up (<3 months) were ex- cluded unless they experienced early death within this period. Additionally, patients lacking data for any of the predictors re- quired to calculate the S-GRAS score were excluded.
Informed consent was obtained from all living patients at the time of registry inclusion. The study was conducted in ac- cordance with the Declaration of Helsinki and applicable laws and was approved by the ethics committees of all participating institutions.
Variables
The S-GRAS score comprises 5 variables: stage, grade, resec- tion status, age, and symptoms (Annex Table S1). The original S-GRAS model was applied exactly as proposed, using the as- signed point-based scoring system without modifications. All predictor data were retrospectively extracted from patient medical records and pathology reports by the treating physi- cians at each center. Stage was determined using the ENSAT staging system, an established ACC staging method that cate- gorizes tumors from stage I to stage IV based on tumor size, local invasion, extension to regional lymph nodes, and metas- tasis.14 Scoring was assigned as 0 points for stage I-II, 1 point for stage III, and 2 points for stage IV. Grade was evaluated using the Ki-67% proliferation index, measured as a continu- ous variable and further categorized into thresholds (<10%, 10%-19%, and ≥20%) in accordance with prior specifica- tions,7 and scored as 0, 1, and 2 points, respectively. Surgical margin status following resection was classified as R0 (complete resection with negative margins), Rx (indeter- minate margins), R1 (microscopically positive margins), or
R2 (macroscopically positive margins), corresponding to scores of 0, 1, 2, and 3 points, respectively. Age at diagnosis was recorded as a continuous variable and scored as 0 points for patients <50 years and 1 point for those ≥50 years. Hormone hypersecretion and tumor-associated symptoms at diagnosis were combined into a single variable indicating the presence (1) or absence (0) of either type of symptom. Following previous recommendations, the S-GRAS score was evaluated either as a cumulative score (0-9) or, in a simpli- fied manner, by grouping the score into 4 prognostic risk cat- egories: 0-1, 2-3, 4-5, and 6-9 points.7 To attempt to expand the model, additional inclusion of primary tumor size (cm) and microscopic evidence of venous invasion-features pre- sent in other multiparametric diagnostic models-was consid- ered.15 The registry also includes data on the use of adjuvant mitotane in these patients.
The endpoints were overall survival (OS) and recurrence- free survival (RFS), both defined from the time of surgery to death from any cause and to recurrence, respectively, with in- dividuals censored at the time of last follow-up if no event oc- curred. Recurrence was primarily detected by routine computed tomography scans, performed every 3-6 months ac- cording to each center’s local practice. While there was no for- mal blinding for outcome assessment, the physicians and radiologists evaluating the clinical and imaging data were not aware of the S-GRAS score at the time of their evaluation.
Statistical methods
The external validation of the S-GRAS prognostic score fol- lowed the usual recommendations for this type of validation studies.16 Calibration and discrimination were evaluated as key metrics of model performance. Calibration was assessed through plots comparing predicted probabilities to observed outcomes, using LOWESS curves to identify miscalibration patterns. Optimism correction was applied via resampling (1000 replications) to reduce the risk of overfitting and ensure robust estimates. Discrimination, assessed using Harrell’s c-index, evaluated the model’s ability to distinguish between individuals with different survival outcomes. Kaplan-Meier survival curves stratified by the S-GRAS score were plotted. The precision of the model was further examined using the Royston-Sauerbrei’s Discrimination R2D statistic, where higher values indicate better performance, and the Brier score at 2, 3, and 5 years, where lower values suggest better calibra- tion and discrimination, with scores ≤0.25 generally consid- ered acceptable.17 Confidence intervals for the Brier scores and the R2D statistic were generated using a percentile boot- strap method based on 500 replicates.
Initial validation was performed using the prognostic index, derived from the originally assigned S-GRAS points.7
Meanwhile, exploratory analyses involved model refine- ments, such as non-linear transformations and additional co- variates, which were clearly described as separate from the primary validation.
The refitted model was obtained by re-estimating the coeffi- cients using the same predictor variables as the original S-GRAS model within a Cox proportional hazards framework.
To assess whether the S-GRAS model could be improved, we evaluated the impact of non-linear transformations of con- tinuous variables, particularly Ki-67%, using restricted cubic splines (piecewise cubic polynomials that join smoothly at pre- defined knots) to avoid arbitrary cutoffs.1
This refined version was referred to as the S-GRAS spline-adjusted model. Furthermore, to enhance precision, a third model (hereafter referred to as the Expanded model) was developed by incorporating venous invasion and tumor size as additional covariates. All 3 models were fitted using Cox proportional hazards models.
To demonstrate the practical utility of spline-based models, we developed an interactive Shiny web application showcas- ing the S-GRAS spline-adjusted model, allowing users to dy- namically explore the impact of covariates on patient outcomes (https://albertocarm.shinyapps.io/SGRAS/).
Model comparisons were conducted using the likelihood ra- tio (LR) test for nested models and the Akaike Information Criterion (AIC), which evaluates model fit while penalizing for complexity to reduce overfitting.
A Cox proportional hazards model was fitted, incorporat- ing the interaction between S-GRAS and adjuvant mitotane treatment to estimate the hazard ratio (HR) associated with mitotane use at each S-GRAS score level.
The sample size was limited by the number of patients avail- able in the registry. Consequently, the robustness of the con- clusions should be interpreted based on the width of the confidence intervals.
Missing data were handled using complete-case analysis, as Ki-67% was missing in 45% of cases. Since its absence was mainly linked to the year of diagnosis rather than tumor biol- ogy, and given the high proportion of missing values, multiple imputation was not considered reliable. Therefore, all in- cluded cases had complete data for all mandatory variables. Statistical analyses were performed using R version 4.3.2, with the rms and survival packages employed for model fitting and validation.19 The underlying R code for model refitting and validation is available upon request to facilitate replication.
Results
Patients
The study cohort included 138 patients with ACC treated with adrenalectomy between 1998 and 2023. Baseline characteris- tics are detailed in Table 1. The median age was 51 years (range: 21-88), and 93 patients (67.3%) were female. Most patients (n = 123, 89.0%) had an ECOG performance status (ECOG-PS) of 0-1. Before surgery, 97 patients (70.2%) were symptomatic, including 47 (34.0%) with hormone-related symptoms, 36 (26.0%) with tumor-related symptoms, and 14 (10.1%) with both. According to ENSAT staging, 6 pa- tients (4.3%) were classified as stage I, 70 (50.7%) as stage II, 61 (44.2%) as stage III, and 1 (0.7%) as stage IV.
Regarding surgical outcomes, 105 patients (76.1%) achieved R0 resection, while 29 (21.0%) had R1 margins, 1 (0.7%) had R2 margins, and margin status was unknown in 3 cases (2.2%). The median Ki-67% index was 19% (inter- quartile range [IQR] 28). The primary tumors had a median size of 10 cm (IQR 7.5).
With regard to other perioperative treatments, 97 patients (70.2%) received adjuvant mitotane, 22 (15.9%) received ad- juvant radiotherapy, and 13 (9.4%) underwent perioperative chemotherapy, either before or after adrenalectomy. The me- dian initial mitotane dose was 2500 mg/day, with a median treatment duration of 23 months (95% CI, 18.8-24.3 months). Therapeutic drug monitoring was performed in 76 of 97 patients (78%). Analysis of peak mitotane concentrations
| Variable | Entire cohort, N = 138 (100%) |
|---|---|
| Age, median (range) | 51.35 (21-88) |
| <50 years | 61 (44.2) |
| ≥50 years | 77 (55.8) |
| Sex, female | 93 (67.3) |
| ECOG-PS | |
| 0 | 71 (51.4) |
| 1 | 52 (37.6) |
| 2 | 5 (3.6) |
| 3 | 3 (2.1) |
| 4 | 1 (0.7) |
| No available | 6 (4.3) |
| Multidisciplinary committee | 89 (64.4) |
| Symptoms | |
| Asymptomatic | 41 (29.7) |
| Symptomatic | 97 (70.2) |
| Hormone-related | 47 (34.0) |
| Tumor-related | 36 (26.0) |
| Both | 14 (10.1) |
| Clinical presentation | |
| Incidental diagnosis | 40 (28.9) |
| Cushing syndrome | 35 (25.3) |
| New hypertension | 22 (15.9) |
| Hypokalemia | 11 (7.9) |
| New diabetes | 9 (6.5) |
| Hirsutism | 38 (27.5) |
| Virilization | 22 (15.9) |
| Amenorrhea | 17 (12.3) |
| Excess mineralocorticoids | 15 (10.8) |
| Paraneoplastic symptoms | 4 (2.9) |
| Erectile dysfunction | 4 (2.9) |
| Feminization | 1 (0.72) |
| Ki-67%, median (IQR) | 19 (28.0) |
| <10% | 28 (20.2) |
| 10%-19% | 41 (29.7) |
| ≥20% | 69 (50.0) |
| Mitotic rate, median (IQR) | 12 (15) |
| Weiss index | |
| 3 | 4 (10.8) |
| 4 | 6 (16.2) |
| 5 | 4 (10.8) |
| 6 | 5 (13.5) |
| 7 | 6 (16.2) |
| 8 | 8 (21.6) |
| 9 | 4 (10.8) |
| NA | 101 |
| ENSAT stage | |
| I | 6 (4.3) |
| II | 70 (50.7) |
| III | 61 (44.2) |
| IV | 1 (0.72) |
| Primary tumor size (cm), median (IQR) | 10 (7.2) |
| Surgical margin | |
| R0 | 105 (76.1) |
| Rx | 3 (2.2) |
| R1 | 29 (21.0) |
| R2 | 1 (0.7) |
| Adjuvant treatment | |
| Adjuvant mitotane | 97 (70.2) |
| Adjuvant chemotherapy | 13 (9.4) |
| Adjuvant radiotherapy | 22 (15.9) |
Abbreviations: ACC, adrenocortical carcinoma; ECOG-PS, Eastern Cooperative Oncology Group Performance Status; ENSAT, European Network for the Study of Adrenal Tumors; IQR, interquartile range.
showed that 33% of patients had maximum levels ≤14 µg/mL, indicating that one-third did not reach the target therapeutic range. Conversely, 24% achieved peak levels within the thera- peutic range (14-20 µg/mL), while 43% exceeded this range at some point, reaching levels >20 µg/mL. Additionally, a longitu- dinal assessment revealed that 45% of samples (231/508) were below the therapeutic range (<10 mg/L), although these were predominantly early treatment samples. These data are pre- sented in Figure S1.
The median follow-up for patients without an event was 60.9 months (95% CI, 51.4-89). The database includes 77 death events, with a median OS of 69 months (95% CI, 60.9-99.4), and 76 recurrence events, with a median RFS of 29.4 months (95% CI, 23.4-86).
Validation of the S-GRAS score
The distribution of S-GRAS scores is shown in the Figure S2. At 5 years, survival-based endpoints, including OS and RFS, differed significantly across S-GRAS score categories. Patients with scores of 0-1 had a 5-year survival probability of 100%, while those with scores of 2-3 had a survival prob- ability of 81.6% (95% confidence interval [CI], 68.1%-97.8%). For patients scoring 4-5, the survival prob- ability dropped to 55.0% (95% CI, 42.7%-70.9%), and for those scoring 6-7, it was 33.8% (95% CI, 19.3%-59.2%). Similar differences were observed for RFS (see Figure 1). The C-index was 0.706 (95% CI, 0.628-0.785) for OS and 0.673 (95% CI, 0.601-0.745) for RFS. In a sensitivity analysis restricted to patients without perioperative chemotherapy or adjuvant radiotherapy, consistent results were observed, with a C-index for OS of 0.723 (95% CI, 0.633-0.813) and for DFS of 0.695 (95% CI, 0.613-0.778). Calibration curves for the OS and RFS endpoint demonstrated good model per- formance at 2, 3, and 5 years, with predictions closely aligning with observed outcomes (Figure 2). The Brier score for RFS was 0.16 (95% CI, 0.11-0.21) at 2 years, 0.24 (95% CI, 0.17-0.27) at 3 years, and 0.28 (95% CI, 0.21-0.30) at 5 years, indicating acceptable predictive accuracy at these time points. For OS, the Brier scores at 2, 3, and 5 years were 0.11 (95% CI, 0.07-0.14), 0.18 (95% CI, 0.14-0.21), and 0.20 (95% CI, 0.17-0.25), respectively, demonstrating good predictive accuracy. The R2D value for RFS was 0.28 (95% CI, 0.04-0.31), and for OS was 0.20 (95% CI, 0.08-0.37).
Effect of mitotane according to S-GRAS scores
We fitted Cox models for RFS and OS that included an inter- action term between adjuvant mitotane use and S-GRAS scores. In both cases, the models suggested a greater effect of mitotane in patients with higher S-GRAS scores (Figure 3). For example, for patients with an S-GRAS score of 4, the HR for RFS was 0.57 (95% CI, 0.34-0.97, P =. 039). Similarly, for those with a score of 5, the HR was 0.46 (95% CI, 0.23-0.94, P =. 032), 0.37 (95% CI, 0.14, 0.97), and for those with a score of 6, the HR was 0.37 (95% CI, 0.14-0.97, P =. 044), suggesting a possible stronger benefit of mitotane in these higher-risk groups. Figure S3 presents Kaplan-Meier curves for OS and RFS stratified by adjuvant mitotane use in adrenalectomy patients.
Recurrence-free survival, % >
1.00
0.75
Grouped
0.50
S-GRAS
score
0-1 points
0.25
-2-3 points
4-5 points
P = 0.00035 (log-rank test)
0.00
→5 points
0
6
12
18
24
30
36
42
48
54
60
Months
Number at risk
16
16
16
16
14
13
12
8
8
7
7
58
56
53
49
44
41
38
33
28
22
21
43
39
32
29
27
20
17
17
16
12
11
20
18
14
12
10
8
7
7
6
5
4
B
1.00
Overall survival, %
0.75
Grouped
0.50
S-GRAS
score
0-1 points
0.25
- 2-3 points
- 4-5 points
P < 0.0001
0.00
>5 points
(log-rank test)
0
6
12
18
24
30
36
42
48
3
54
60
Months
Number at risk
16
16
16
16
16
13
12
11
10
8
8
59
59
56
51
48
44
39
35
25
23
43
41
32
29
28
21
18
18
16
15
12
-20 20 15 12 10 8 7 7 6 6 4
Refitting and updating the model
Subsequently, we refitted the original model, incorporating improvements as specified in the Methods section. The coeffi- cients of the re-fitted S-GRAS model, derived from the new da- taset, were consistent with the originally proposed scoring system, reflecting a similar weighting of prognostic factors. Nonetheless, the spline-adjusted model captured more nuanced effects and demonstrated a better trade-off between predictive accuracy and complexity. Thus, this spline-based specification significantly outperformed the refitted S-GRAS model for RFS (LR test, x = 18.66; P <. 001), achieving super- ior discrimination with a bootstrap bias-corrected C-index of 0.738 (95% CI, 0.675-0.800) for RFS and 0.760 (95% CI, 0.690-0.820) for OS. The Supplementary Material pro- vides additional details on this comparison. Incorporating
non-linear terms enhanced the model’s performance, with strong evidence of significant non-linearity for Ki-67% (P =. 0012) and the combined non-linear terms (x2=14.52, df =2, P =. 0007) (Figure 4). The spline-adjusted model is available at https://albertocarm.shinyapps.io/SGRAS/, provided solely to exemplify its straightforward use. The improved perform- ance of the spline-adjusted model likely reflects its ability to capture non-linear prognostic effects, particularly for Ki-67% and tumor size, while avoiding unnecessary model complexity. In contrast, the expanded model with additional covariates showed a slightly higher R2D for RFS (0.38; 95% CI, 0.32-0.64 vs. 0.36; 95% CI, 0.27-0.58). However, when compared with the spline-adjusted model using a LR test, the difference in model performance was not statistically sig- nificant (P =. 233). Similarly, for OS, its marginally higher R2D (0.46; 95% CI, 0.34-0.63 vs. 0.43; 95% CI, 0.40-0.73) and minimal AIC differences (475.2 vs. 404.6) indicate that the added complexity of the expanded model is not justified by clinically relevant improvements in performance compared with the spline-adjusted model without additional covariates. These new models maintain good calibration, comparable to the original model. Further details are provided in the Supplementary Material.
Discussion
This study confirms the validity of the S-GRAS score in a con- temporary Iberian cohort, supporting its clinical applicability for stratifying ACC patients. Our findings reinforce its utility in guiding adjuvant mitotane therapy, particularly for intermediate-risk patients, and highlight the need for addition- al prognostic markers to refine treatment decisions.
The S-GRAS model evolved through progressive refine- ments of prognostic models developed by the ENSAT network over the past decade.6,7 While the ENSAT stage became the standard classification system and possibly the most critical prognostic factor after its development,5 substantial hetero- geneity within each stage soon became apparent, indicating the need for more complex multivariable models that incorpo- rated additional variables, such as grade, resection status, age, tumor functionality, or treatment-related factors, tumor size, and venous invasion.20 In line with this approach, Libé et al.6 showed that, beyond ENSAT stage, 4 additional prog- nostic parameters-grouped under the GRAS acronym (grade, resection status, age, symptoms)-provided incremen- tal prognostic value.6 Some S-GRAS parameters, such as Ki-67% and age, were dichotomized using predefined cutoffs, generally based on the medians in the derivation cohorts. The S-GRAS model, validated in 2021 by Elhassan et al.7 in a mul- ticenter cohort of 942 patients treated with adrenalectomy, demonstrated strong prognostic performance, with a Harrell’s C-index of 0.77 and a Royston and Sauerbrei R2D statistic of 0.46 for disease-specific survival.7 One year later, Lin et al.8 conducted a study in China evaluating the SEER-based nomogram against the S-GRAS model and the ENSAT stage.8 In this assessment, while the SEER model dem- onstrated good performance, the ENSAT stage achieved an area under the receiver operating characteristic curve (AUC) of 0.640 (95% CI, 0.543-0.737) for 5-year survival. Notably, the S-GRAS model outperformed the ENSAT stage, with an AUC of 0.683 (95% CI, 0.602-0.764). Since then, sev- eral groups have validated the S-GRAS model in Europe and Asia.
8,9,21-23 Consistent with previous validations, our
A Calibration at 2 years
B Calibration at 3 years
C Calibration at 5 years
Fraction surviving
2 years
0.0 0.2 0.4 0.6 0.8 1.0
Fraction surviving
3 years
0.0 0.2 0.4 0.6 0.8 1.0
Fraction surviving
0.0 0.2 0.4 0.6 0.8 1.0
5 years
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
Predicted 2-year survival
Predicted 3-year survival
Predicted 5-year survival
D Calibration at 2 years
E Calibration at 3 years
F Calibration at 5 years
2-year recurrence-free
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
fraction
3-year recurrence-free
fraction
5-year recurrence-free
fraction
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
Predicted 2-year RFS rate
Predicted 3-year RFS rate
Predicted 5-year RFS rate
A
Effect of mitotane on RFS
7
S-GRAS score
6
5
4
3
2
1
0
0
1
2
3
4
Hazard Ratio (HR)
HR (95% CI, p-value)
0.30 (0.08, 1.05), 0.059
0.37 (0.14, 0.97), 0.044
0.46 (0.23, 0.94), 0.032
0.57 (0.34, 0.97), 0.039
0.71 (0.43, 1.20), 0.204
0.89 (0.45, 1.77), 0.741
1.11 (0.43, 2.85), 0.828
1.38 (0.40, 4.74), 0.605
B Effect of mitotane on OS
7
S-GRAS score
6
5
4
3
2
1
0
0
1
2
3
4
5
Hazard Ratio (HR)
0.27 (0.07, 1.02), 0.054
0.33 (0.12, 0.92), 0.033
0.41 (0.20, 0.86), 0.018
0.51 (0.29, 0.89), 0.019
0.63 (0.34, 1.14), 0.128
0.77 (0.34, 1.75), 0.540
0.96 (0.31, 2.91), 0.938
1.18 (0.28, 5.00), 0.820
Hazard ratio for varying levels of Ki-67% (Spline-adjusted model for OS)
3
Chi-Square: 14.52 (p = 0.0001)
Hazard Ratio (HR)
Nonlinear: 10.57 (p = 0.0012)
2
1
0
20
40
60
Ki-67 (%)
B Hazard ratio across different age levels (Spline-adjusted model for OS)
Hazard Ratio (HR)
6
Chi-Square: 20.52 (p <0.0001)
Nonlinear: 2.31 (p = 0.128)
4
2
0
30
40
50
60
70
Age (years)
findings support that the S-GRAS model maintains strong dis- criminative ability for OS in the Iberian ÍCARO registry. However, its discriminatory capacity was somewhat lower for RFS. Furthermore, our findings support Elhassan et al.7 proposal that higher S-GRAS scores could help identify opti- mal candidates for adjuvant mitotane therapy, although only scores of 4-5 reached statistical significance in their series,7 while our data suggest a more gradual effect modification, with smoother transitions across score ranges. The statistical significance observed across different strata, both in our data- set and that of Elhassan et al.7 likely reflects sample size varia- tions within each stratum rather than true differences in therapeutic benefit. Notably, in our cohort, the S-GRAS scores displayed an approximately bell-shaped distribution with mo- dest negative skewness, peaking in the 3-5 range and including fewer patients at the extremes. Consequently, the limited num- ber of cases with S-GRAS scores >5 leads to wider confidence intervals crossing HR 1 for the therapeutic effect of adjuvant mitotane in these strata, reflecting a lack of statistical evidence rather than an absence of effect. This pattern is consistent with findings from both our dataset and Elhassan’s series.7
The most concerning finding in our analysis was that despite good calibration and discrimination capacity, the original S-GRAS score applied to our cohort appeared somewhat less accurate than in Elhassan’s analysis, with R2D values of 0.281 and 0.202 for RFS and OS, respectively. Potential
sources of this discrepancy could include non-linear effects of continuous variables or heterogeneity due to unmeasured prognostic factors, which we decided to explore more formal- ly through 2 refined models. To address the first issue, we at- tempted to model age and other continuous variables in a non-linear fashion. For the second issue, we incorporated tu- mor size and venous invasion. These were selected based on theoretical grounds, while other common variables like ECOG-PS were excluded as they might be relevant in ad- vanced disease but were of questionable value in the post- adrenalectomy setting.
In more detail, this is what we have learned. First, this ana- lysis serves as yet another example supporting the longstand- ing warnings from multiple authors regarding the loss of information and limitations introduced by the dichotomization of continuous variables, as well as the benefits of non-linear modeling. Interested readers can find relevant discussions in the literature on this topic.24-26 While simplicity is often cited as an argument for dichotomization, our work demonstrates that developing online applications to implement these more precise models can now be accomplished very easily and at no cost.
Second, beyond these improvements, our models suggest that progressive refinements based solely on clinical covariates are gradually approaching a performance plateau. While fur- ther optimization remains possible, substantial enhancements in predictive accuracy appear unlikely without incorporating additional types of data. This remains important, as the low R2 D values observed-even with non-linear spline-based mod- els-indicate that a substantial proportion of variance remains unexplained.27 As our findings suggest, further improvements through clinical methods alone are unlikely.
Therefore, the field should prioritize the validation of mo- lecular biomarkers, either to substantially enhance prognostic accuracy or to predict the benefit of adjuvant therapy. For ex- ample, Lippert et al.28 demonstrated that optimal prediction of outcomes was achieved by integrating molecular data (in- cluding Wnt/B-catenin and p53 pathway alterations and methylation patterns) with clinical and histopathological pa- rameters.28 Given that adjuvant mitotane therapy is the most commonly used adjuvant treatment, the search for mo- lecular predictive biomarkers appears to be a priority. Recent work by Zhang et al.29 has demonstrated a connection between mitotane response and lipid metabolism pathways, with CYP27A1 and ABCA1 expression emerging as potential predictors.29 Such efforts could complement current strategies based solely on prognostic stratification (eg, treating high-risk patients) by enabling the identification of those at high-risk that are also likely to benefit from therapy.
This study has several limitations. First, its retrospective de- sign introduces potential biases, including the impact of changes in clinical practice over the more than 25-year period covered by the cohort. While this long-time frame allows for extended follow-up, it may not fully reflect contemporary treatment approaches. However, unfortunately, treatment al- gorithms have barely been updated in recent years. The poten- tial for confounding bias in the mitotane-survival association warrants careful consideration, though the inclusion of S-GRAS score as a covariate provides some protection against this bias by capturing key determinants of treatment selection.
Second, the need for complete-case analysis to validate a prognostic model, coupled with the rarity of the disease (1-2 cases per million/year) (4), represents a key limitation that
constrains our statistical power. Although the low number of events limits the feasibility of expanding the model with great- er complexity, we believe it is unlikely that we have over- looked variables with a strong effect in our dataset.
In conclusion, our validation of the S-GRAS score using the ACC registry ICARO-GETTHI/SEEN demonstrates that non- linear modeling of continuous variables-most notably the Ki-67% index-offers an immediate and parsimonious path to improved precision without increasing model complexity. However, as clinical and histopathological variables appear to be reaching their performance ceiling, substantial unex- plained variance is likely to persist despite these refinements, suggesting the influence of unmeasured factors. Therefore, meaningful enhancements in predictive accuracy could poten- tially be achieved by integrating molecular biomarkers. Additionally, our results support the utility of S-GRAS scoring for identifying candidates for adjuvant mitotane therapy, with evidence suggesting an incremental benefit pattern at higher scores.
Acknowledgments
The authors express their gratitude to the ICARO registry re- searchers for their contributions to this study, and to GETTHI and SEEN for their support in its promotion. Special thanks to Natalia Cateriano, Miguel Vaquero, and the IRICOM SL team for designing, overseeing, and continuously supporting the registry’s web platform.
Supplementary material
Supplementary material is available at European Journal of Endocrinology online.
Funding
The database was funded by Esteve-HRA Pharma Rare Diseases. The funders had no role in the study design, data registration, statistical analysis, manuscript writing, or the de- cision to submit for publication.
Authors’ contributions
Paula Jimenez-Fonseca (Conceptualization [equal], Funding acquisition [equal], Investigation [equal], Methodology [equal], Project administration [equal], Resources [equal], Supervision [equal], Validation [equal], Writing-review & editing [equal]), Cristina Alvarez-Escola (Funding acquisition [equal], Investigation [equal], Methodology [equal], Project administration [equal], Resources [equal], Writing-review & editing [equal]), Inmaculada Ballester Navarro (Project ad- ministration [equal], Resources [equal]), Jorge Hernando Cubero (Resources [equal], Supervision [equal]), Laura González (Resources [equal]), Miguel Ángel Mangas (Resources [equal], Writing-review & editing [equal]), Clara Iglesias (Resources [equal], Writing-review & editing [equal]), Jesús García-Donas (Conceptualization [equal], Methodology [equal], Project administration [equal], Writing-review & editing [equal]), María Picón César (Supervision [equal], Writing-review & editing [equal]), Miguel Paja Fano (Writing-review & editing [equal]), Lorena González Batanero (Writing-review & editing [equal]), Lourdes Garcia (Writing-review & editing [equal]), Javier Molina (Writing-review & editing [equal]), Raquel
Jimeno Maté (Writing-review & editing [equal]), Javier Aller (Conceptualization [equal], Investigation [equal], Supervision [equal], Writing-review & editing [equal]), M. Tous Romero (Writing-review & editing [equal]), Jersy Cardenas-Salas (Writing-review & editing [equal]), Gala Gutierrez-Buey (Writing-review & editing [equal]), Nerea Egaña (Writing-review & editing [equal]), Miguel Navarro (Writing-review & editing [equal]), Maria José Lecumberri (Writing-review & editing [equal]), Nuria Valdés (Conceptualization [equal], Funding acquisition [equal], Methodology [equal], Project administration [equal], Resources [equal], Writing-original draft [equal], Writing -review & editing [equal]), Alberto Carmona-Bayonas (Conceptualization [equal], Data curation [equal], Formal analysis [equal], Investigation [equal], Methodology [equal], Project administration [equal], Resources [equal], Software [equal], Supervision [equal], Validation [equal], Visualization [equal], Writing-original draft [equal]), and Writing-review & editing [equal])
A.C .- B., P.J .- F., C.A .- E., N.V., and I.B. conceived and de- signed the project. A.C .- B. and P.J .- F. processed the data and prepared the manuscript. The other authors supplied clin- ical information from their enrolled patients and contributed to refining the manuscript. All authors have read and ap- proved the final version.
Conflict of interest: P.J .- F. has received honoraria for speaking engagements and advisory board participation from Astellas, AstraZeneca, Bristol-Myers Squibb (BMS), HRA Pharma-Esteve, Merck Sharp & Dohme (MSD), Novartis, Nutricia, Pfizer, Rovi, Takeda, and Viatris. C.A .- E. and N.V. has received honoraria as speaker fees, served on the ad- visory board, and received sponsorship for travel and accom- modation at scientific meetings from HRA-Pharma-Esteve. A.C .- B. reports receiving lecture grants from Esteve, Lilly, and Astellas; travel support from Amgen; and research fund- ing from HRA Pharma-Esteve. J.H.C. reports funding from Novartis, Eisai, Ipsen, Bayer, and HRA Pharma-Esteve. All other authors declare no competing interests related to the scope of this work.
Data availability
The data and analyses performed are available from the corre- sponding author upon reasonable request.
Ethics approval and consent to participate
This study was conducted in compliance with the principles of Good Clinical Practice and all applicable local regulations and laws. The study protocol was initially approved on December 29th, 2017, by the Ethics Committee (CEIM) of Hospital Universitario Central de Asturias (Ref: 218/17), and subse- quently by the CEIMs of all participating centers. In cases where patients were deceased at the time of registration, the re- quirement for written informed consent was waived.
References
1. Ayala-Ramirez M, Jasim S, Feng L, et al. Adrenocortical carcinoma: clinical outcomes and prognosis of 330 patients at a tertiary care center. Eur J Endocrinol. 2013;169(6):891-899. https://doi.org/ 10.1530/EJE-13-0519
2. Bilimoria KY, Shen WT, Elaraj D, et al. Adrenocortical carcinoma in the United States: treatment utilization and prognostic factors.
Cancer. 2008;113(11):3130-3136. https://doi.org/10.1002/cncr. 23886
3. Terzolo M, Fassnacht M, Perotti P, et al. Adjuvant mitotane versus surveillance in low-grade, localised adrenocortical carcinoma (ADIUVO): an international, multicentre, open-label, randomised, phase 3 trial and observational study. Lancet Diabetes Endocrinol. 2023;11(10):720-730. https://doi.org/10.1016/S2213-8587(23)00 193-6
4. He S, Huang X, Zhao P, Zhang P. The prognosis difference between elderly and younger patients with adrenocortical carcinoma. Front Genet. 2023;13:1029155. https://doi.org/10.3389/fgene.2022. 1029155
5. Kong J, Zheng J, Cai J, et al. A nomogram for individualized estima- tion of survival among adult patients with adrenocortical carcin- oma after surgery: a retrospective analysis and multicenter validation study. Cancer Commun. 2019;39(1):80. https://doi.org/ 10.1186/s40880-019-0426-0
6. Libé R, Borget I, Ronchi CL, et al. Prognostic factors in stage III-IV adrenocortical carcinomas (ACC): an European Network for the Study of Adrenal Tumor (ENSAT) study. Ann Oncol. 2015;26(10):2119-2125. https://doi.org/10.1093/annonc/mdv329
7. Elhassan YS, Altieri B, Berhane S, et al. S-GRAS score for prognostic classification of adrenocortical carcinoma: an international, multi- center ENSAT study. Eur J Endocrinol. 2022;186(1):25-36. https://doi.org/10.1530/EJE-21-0510
8. Lin W, Dai J, Xie J, et al. S-GRAS score performs better than a model from SEER for patients with adrenocortical carcinoma. Endocr Connect. 2022;11(6):e220114. https://doi.org/10.1530/ EC-22-0114
9. Mihai R. Stage, Grade, Resection status, Age, Symptoms (S-GRAS) score in predicting survival after resection of adrenocortical carcin- omas. BJS open. 2024;8(1):zrae001. https://doi.org/10.1093/ bjsopen/zrae001
10. Van Calster B, Steyerberg EW, Wynants L, Van Smeden M. There is no such thing as a validated prediction model. BMC Med. 2023;21(1):70. https://doi.org/10.1186/s12916-023-02779-w
11. Calabrese A, Basile V, Puglisi S, et al. Adjuvant mitotane therapy is beneficial in non-metastatic adrenocortical carcinoma at high risk of recurrence. Eur J Endocrinol. 2019;180(6):387-396. https:// doi.org/10.1530/EJE-18-0923
12. Tőke J, Uhlyarik A, Lohinszky J, et al. Prognostic factors and mito- tane treatment of adrenocortical cancer. Two decades of experience from an institutional case series. Front Endocrinol (Lausanne). 2022;13:952418. https://doi.org/10.3389/fendo.2022.952418
13. Collins GS, Reitsma JB, Altman DG, Moons K. Transparent report- ing of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMC Med. 2015;13(1):1. https://doi.org/10.1186/s12916-014-0241-z
14. Lughezzani G, Sun M, Perrotte P, et al. The European Network for the Study of Adrenal Tumors staging system is prognostically super- ior to the international union against cancer-staging system: a North American validation. Eur J Cancer. 2010;46(4):713-719. https://doi.org/10.1016/j.ejca.2009.12.007
15. Mete O, Erickson LA, Juhlin CC, et al. Overview of the 2022 WHO classification of adrenal cortical tumors. Endocr Pathol. 2022;33(1):155-196. https://doi.org/10.1007/s12022-022-09710-8
16. Moons KGM, Kengne AP, Grobbee DE, et al. Risk prediction mod- els: II. External validation, model updating, and impact assessment. Heart. 2012;98(9):691-698. https://doi.org/10.1136/heartjnl-2011- 301247
17. Biancari F, Vasques F, Mikkola R, Martin M, Lahtinen J, Heikkinen J. Validation of EuroSCORE II in patients undergoing coronary artery bypass surgery. Ann Thorac Surg. 2012;93(6): 1930-1935. https://doi.org/10.1016/j.athoracsur.2012.02.064
18. Rutherford MJ, Crowther MJ, Lambert PC. The use of restricted cubic splines to approximate complex hazard functions in the analysis of time-to-event data: a simulation study. J Stat Comput Simul. 2015;85(4):777-793. https://doi.org/10.1080/00949655. 2013.845890
19. Harrell FE Jr. Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis. 2nd ed. Cham: Springer International Publishing; 2015.
20. Zhang H, Naji Y, Yan M, Lian W, Xie M, Dai Y. Development and validation of prognostic nomograms in patients with ad- renocortical carcinoma: a population-based study. Int Urol Nephrol. 2020;52(6):1057-1071. https://doi.org/10.1007/s11255- 020-02413-1
21. Baek SK, Lee SH, Park SS, et al. The modified S-GRAS scoring sys- tem for prognosis in Korean with adrenocortical carcinoma. Endocrinol Metab. 2024;39(5):803-812. https://doi.org/10.3803/ EnM.2024.2086
22. Benyei E, Tőke J, Huszty G, et al. Clinical utility of S-GRAS prog- nostic system in adrenocortical carcinomas: confirmatory results from a single-centre institutional registry. Endocr Abstr. 2024;99: EP254. https://doi.org/10.1530/endoabs.99.EP254
23. Liang J, Liu Z, Zhou L, et al. The clinical utility of ‘GRAS’ param- eters in stage I-III adrenocortical carcinomas: long-term data from a high-volume institution. Endocrine. 2020;67(2):449-456. https:/ doi.org/10.1007/s12020-019-02141-2
24. Harrell FE. Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis. Cham: Springer International Publishing; 2015.
25. Royston P, Altman DG, Sauerbrei W. Dichotomizing continuous predictors in multiple regression: a bad idea. Stat Med. 2006;25(1):127-141. https://doi.org/10.1002/sim.2331
26. Carmona-Bayonas A, Jiménez-Fonseca P, Lamarca Á, et al. Prediction of progression-free survival in patients with advanced, well-differentiated, neuroendocrine tumors being treated with a somatostatin analog: the GETNE-TRASGU study. J Clin Oncol. 2019;37(28):2571-2580. https://doi.org/10.1200/JCO.19.00980
27. Gupta A, Stead TS, Ganti L. Determining a meaningful R-squared value in clinical medicine. Acad Med Surg. 2024. 10.62186/001c. 125154 [Epub ahead of print].
28. Lippert J, Appenzeller S, Liang R, et al. Targeted molecular analysis in adrenocortical carcinomas: a strategy toward improved personal- ized prognostication. J Clin Endocrinol Metab. 2018;103(12): 4511-4523. https://doi.org/10.1210/jc.2018-01348
29. Zhang J, Wu L, Su T, et al. Pharmacogenomic analysis in adrenocort- ical carcinoma reveals genetic features associated with mitotane sen- sitivity and potential therapeutics. Front Endocrinol (Lausanne). 2024;15:1365321. https://doi.org/10.3389/fendo.2024.1365321