ENDOCRINE SOCIETY
OXFORD
Serum Steroid Profiling in the Diagnosis of Adrenocortical Carcinoma: A Prospective Cohort Study
Kai Yu,1,20D Shobana Athimulam,3[D Jasmine Saini,20D Ravinder Jeet Kaur,2 Qingping Xue,4 Travis J. Mckenzie,5 Ravinder J. Singh,6 Stefan Grebe,6 and Irina Bancos2,6 [D
1Division of Endocrinology and Metabolism, West China Hospital, Sichuan University, Chengdu, Sichuan 610041, China 2Division of Endocrinology, Diabetes, Metabolism and Nutrition, Mayo Clinic, Rochester, MN 55905, USA
3Division of Endocrinology, Diabetes, Bone and Mineral Disorder, Henry Ford Health, Detroit, MI 48202, USA
4Department of Epidemiology and Biostatistics, School of Public Health, Chengdu Medical College, Chengdu, Sichuan 610041, China 5Department of Surgery, Mayo Clinic, Rochester, MN 55905, USA
6Department of Laboratory medicine and Pathology, Mayo Clinic, Rochester, MN 555905, USA
Correspondence: Irina Bancos, MD, MSc, Division of Endocrinology, Diabetes, Metabolism and Nutrition, Mayo Clinic, 200 First Street SW, Rochester, MN 55905, USA. Email: bancos.irina@mayo.edu.
Abstract
Context: Guidelines suggest performing urine steroid profiling in patients with indeterminate adrenal tumors to make a noninvasive diagnosis of adrenocortical carcinoma (ACC). However, urine steroid profiling is not widely available.
Objective: To determine the accuracy of clinically available serum 11-deoxycortisol, 17OH-progesterone, and 17OH-pregnenolone in diagnosing ACC.
Methods: We conducted a prospective single-center cohort study of patients with adrenal masses evaluated between 2015 and 2023. Serum was analyzed by liquid chromatography-mass spectrometry for 17OH-pregnenolone, 17OH-progesterone, and 11-deoxycortisol. Reference standard for adrenal mass included histopathology, imaging characteristics, imaging follow up of 2 years, or clinical follow up of 5 years. Localized Generalized Matrix Learning Vector Quantization analysis was used to develop serum steroid score and assessed with area under receiver operating curve.
Results: Of 263 patients with adrenal masses, 44 (16.7%) were diagnosed with ACC, 161 (61%) with adrenocortical adenomas (ACAs), 27 (10%) with other adrenal malignancies, and 31 (12%) with other. Hounsfield unit ≥ 20 was demonstrated in all ACCs, in all but 1 other adrenal malignancy, and in 58 (31%) ACAs. All 3 steroids were higher in patients with ACCs vs non-ACCs, including when comparing ACCs with functioning ACAs, and with ACAs with Hounsfield unit ≥ 20 (P <. 0001 for all). Localized Generalized Matrix Learning Vector Quantization analysis yielded a serum steroid score that discriminated between ACC and non-ACC groups with a mean threshold fixed area under receiver operating curve of 0.823.
Conclusion: We showed that measurements of 11-deoxycortisol, 17OH-progesterone, and 17OH-pregnenolone could be valuable in diagnosing ACC. After appropriate validation, serum steroid score could be integrated in clinical practice.
Key Words: steroid, adrenocortical carcinoma, adrenal incidentaloma, diagnosis
Abbreviations: ACA, adrenocortical adenoma; ACC, adrenocortical carcinoma; AUC, area under the receiver operating characteristic; HU, Hounsfield unit; LC-MS/MS, liquid chromatography-tandem mass spectrometry; LGMLVQ, Localized Generalized Matrix Learning Vector Quantization; ROC, receiver operating characteristic; SSM, serum steroid metabolite.
Adrenal tumors are common, reported in up to 7% of patients undergoing cross-sectional imaging (1-4). Malignancy was reported in 8.6% of patients with adrenal tumors, most being adrenal metastases, with <0.5% being adrenocortical carcin- omas (ACCs) in a population setting (1). In the endocrine set- ting, ACCs are more common, representing approximately 5% of all adrenal tumors (5). The overall prognosis of patients with ACC is poor but much better when diagnosed at an early stage (6-8).
Distinguishing ACC from lipid poor adenoma is sometimes challenging, especially in smaller ACCs that do not present with a combined adrenal hormone excess (9). In addition, ACCs share overlapping imaging characteristics with other
adrenal malignancies, with most being larger than 4 cm and demonstrating unenhanced Hounsfield unit (HU) on computed tomography of >20 HU (5, 10-12). Biopsy is contraindicated in a localized ACC because of a potential for iatrogenic tumor dissemination, whereas adrenalectomy may not be the manage- ment of choice in a patient with adrenal metastasis (2). Therefore, an accurate noninvasive diagnosis is also needed to distinguish ACC from other malignancies.
Disorganized steroidogenesis in ACC was noted more than 7 decades ago with studies reporting increased concentrations of urinary 17-ketosteroids and tetrahydrocortisol, as well as 11-deoxycortisol, 17OH-progesterone, progesterone, and andro- gens in the venous blood (13-15). A larger study of 147 patients
2015-2019: Consecutive patients with adrenal masses who consented to provide fasting serum sample, n=184
2019-2023: Consecutive patients with indeterminate adrenal masses with available clinical measurements of serum steroids, n=85
Excluded:
· Congenital adrenal hyperplasia
n=4
· Exogenous glucocorticoid use
· Absence of unenhanced computed tomography
n=2
· Lack of reference standard
· Lack of available serum sample
Total analyzed n=263
that was first to use machine learning analysis demonstrated that 3 urinary steroid metabolites of 17OH-pregnenolone, pregnenolone, and 11-deoxycortisol were most informative in discriminating ACCs from benign adrenal tumors (16).
Single serum steroid measurements for 11-deoxycortisol, 17OH-progesterone, dehydroepiandrosterone sulfate, and 17-beta-estradiol were suggested by the ACC guidelines from 2018 (17). However, as illustrated by a recent large multicenter study (18), only 17% of patients with ACC had 11-deoxycortisol measurements, and only 24% had 17OH-progesterone measurements, with abnormal results in 30% to 60%. Urine-based multisteroid profiling was shown to distinguish ACC from other tumors with high accuracy in a prospective validation study (5). Based on urine steroid pro- filing, urine metabolite of 17OH-pregnenolone (in addition to metabolite of 11-deoxycortisol and 17OH-progesterone) demonstrated a high diagnostic value. Subsequently, adrenal incidentaloma guidelines from 2023 recommended consider- ing urine steroid profiling in patients with indeterminate ad- renal masses (2). In addition, plasma multisteroid profiling has been reported to distinguish between ACCs and adenomas (19-21) but have not been available outside research centers or several laboratories.
To address the clinical implementation gap, we designed a prospective cohort study with an objective to determine the diagnostic performance of clinically available serum 17OH- progesterone, 17OH-pregnenolone, and 11-deoxycortisol over- all, and among patients with indeterminate adrenal masses.
Methods
Participants
This research was in accordance with the institutional review board of Mayo Clinic (Rochester, MN, United States; IRB 13-005838). All participants provided written informed con- sent. Between 2015 and 2023, we prospectively enrolled pa- tients with adrenal masses who consented to participate in the adrenal mass registry and biobank study. Only patients who agreed to provide a fasting serum sample were included
between 2015 and 2019. Between 2019 and 2023, only pa- tients with clinically available serum steroid measurements were included (Fig. 1). Exclusion criteria were congenital ad- renal hyperplasia, exogenous glucocorticoid use within 3 months before enrollment, absence of unenhanced HU measurement, and lack of reference standard.
We prospectively collected clinical data (demographics, mode of discovery, results of routine hormonal workup, im- aging, histopathology, clinical follow up). Reference standard included 1 or more of the following: histopathology, at least 2-year imaging follow up, and at least 5-year clinical follow up.
Adrenal Hormone Excess
Based on routine clinical hormonal work-up, patients were defined as demonstrating no adrenal hormone excess (nonfunc- tioning), cortisol, mineralocorticoid, androgen, catecholamine, or combined hormone excess (more than 2 types of hormonal excess). Specifically, cortisol excess was defined as postdex- amethasone morning cortisol (1-mg dexamethasone) > 1.8 g/dL in patients without features of overt Cushing syn- drome, and abnormal 1-mg dexamethasone, 24-hour urine cortisol, and/or late-night salivary cortisol in patients with features of overt Cushing syndrome. Mineralocorticoid ex- cess was defined based on the most recent guidelines (22). Catecholamine excess was defined based on elevated plasma metanephrines or 24-hour urine catecholamines/metanephrines. Androgen excess was defined when dehydroepiandrosterone sulfate, androstenedione, or testosterone were > upper nor- mal range.
Serum Steroid Measurements
Serum 11-deoxycortisol (reference range: 10-79 ng/dl) was measured using liquid chromatography-tandem mass spec- trometry (LC-MS/MS) using an Applied Biosystems Tandem Mass Spectrometer with Atmospheric Pressure Chemical Ionization Source. Serum 17OH-pregnenolone (reference range: 55-455 ng/dl in men, 31-455 ng/dl in women) and 17OH-progesterone (reference range <220 ng/dl in men,
| Variables | All (n = 263) | HU < 10 (n = 86) | HU 10-20 (n = 44) | HU ≥ 20 (n = 133) | P value |
|---|---|---|---|---|---|
| Age, y, median (IQR) | 62.00 | 62.50 | 62.00 | 61.00 | .156ª |
| (49.50, 70.50) | (54.25, 71.75) | (44.00, 67.75) | (47.00, 70.00) | .9126 | |
| Women, n (%) | 160 (60.84%) | 54 (62.79%) | 30 (68.18%) | 76 (57.14%) | .490ª |
| .2646 | |||||
| Mode of discovery, n (%) | .017ª | ||||
| Incidental | 175 (66.54%) | 67 (77.91%) | 29 (65.91%) | 79 (59.40%) | .775b |
| Cancer staging imaging | 15 (5.70%) | 4 (4.65%) | 2 (4.55%) | 9 (6.77%) | |
| Other | 73 (27.76%) | 15 (17.44%) | 13 (29.55%) | 45 (33.83%) | |
| Location, n (%) | |||||
| Bilateral | 42 (15.97%) | 17 (19.77%) | 13 (29.55%) | 12 (9.02%) | .006ª |
| Left | 123 (46.77%) | 46 (53.49%) | 17 (38.64%) | 60 (45.11%) | .003b |
| Right | 98 (37.26%) | 23 (26.74%) | 14 (31.82%) | 61 (45.86%) | |
| Size, cm, median (IQR) | 3.10 | 2.10 | 2.10 | 4.80 | <. 001ª |
| (1.80, 5.55) | (1.43, 3.70) | (1.65, 2.93) | (2.60, 8.60) | <. 001b | |
| Hounsfield units, median (IQR) | 20.00 | 1.00 | 13.50 | 35.00 | — |
| (6.25, 35.00) | (-6.12, 5.00) | (10.00, 15.00) | (27.00, 40.00) | ||
| Hormone excess, n (%) | <. 001ª | ||||
| Cortisol excess | 54 (20.53%) | 12 (13.95%) | 10 (22.73%) | 32 (24.06%) | .036b |
| Mineralocorticoid excess | 16 (6.08%) | 4 (4.65%) | 5 (11.36%) | 7 (5.26%) | |
| Hyperandrogenism | 4 (1.52%) | 0 (0.00%) | 0 (0.00%) | 4 (3.01%) | |
| Catecholamine excess | 4 (1.52%) | 0 (0.00%) | 1 (2.27%) | 3 (2.26%) | |
| Combined hormone excess | 19 (7.22%) | 1 (1.16%) | 0 (0.00%) | 18 (13.53%) | |
| None | 166 (63.12%) | 69 (80.23%) | 28 (63.64%) | 69 (51.88%) | |
| 17-hydroxypregnenolone, ng/dL, median (IQR) | 49.00 | 38.00 | 48.50 | 64.00 | <. 001ª |
| (25.00, 104.00) | (21.25, 63.75) | (24.00, 70.75) | (34.00, 205.00) | .0106 | |
| 11-deoxycortisol, ng/dL, median (IQR) | 33.00 | 26.50 | 29.50 | 50.00 | <. 001ª |
| (20.00, 82.50) | (16.00, 45.25) | (22.75, 38.25) | (24.00, 140.00) | .001b | |
| 17-hydroxyprogesterone, ng/dL, median (IQR) | 47.00 | 40.00 | 40.00 | 66.00 | <. 001ª |
| (40.00, 93.50) | (40.00, 63.25) | (40.00, 84.25) | (40.00, 122.00) | .015b | |
| Diagnosis of adrenal mass, n (%) | <. 001ª | ||||
| Adrenocortical carcinoma | 44 (16.73%) | 0 (0.00%) | 0 (0.00%) | 44 (33.08%) | <. 001b |
| Other malignant | 27 (10.27%) | 0 (0.00%) | 1 (2.27%) | 26 (19.55%) | |
| Pheochromocytoma | 6 (2.28%) | 0 (0.00%) | 1 (2.27%) | 5 (3.76%) | |
| Adrenocortical adenoma | 161 (61.22%) | 81 (94.19%) | 37 (84.09%) | 43 (32.33%) | |
| Other benign | 25 (9.51%) | 5 (5.81%) | 5 (11.36%) | 15 (11.28%) |
P values were calculated via Mann-Whitney U test or chi-squared test (category).
Abbreviations: HU, Hounsfield unit; IQR, interquartile range.
“HU < 10 vs ≥20.
HU 10-20 vs ≥20.
<80 ng/dL in women during follicular phase, <285 ng/dL dur- ing luteal phase and <51 ng/dl postmenopausal) were meas- ured using LC-MS/MS on the SCIEX 5000 mass spectrometer.
Statistics
Continuous variables were summarized as median and inter- quartile range unless otherwise stated. Category variables were reported as numbers and percentages. Comparisons were based on Mann-Whitney U test (for continuous varia- bles) and the chi-square test (for categorical variables).
Machine Learning-based Classification
Localized generalized matrix learning vector quantization (LGMLVQ) was applied to create a classifier for the
discrimination of ACC and non-ACC by incorporating serum steroid data measured by LC/MS-MS. LGMLVQ, an extension of learning vector quantization, is prototype-supervised and easily interpretable (23, 24). LGMLVQ learns from labeled dataset to determine the optimal prototype and adaptive matrix of relevant factors per class following nearest prototype and minimum classification error scheme. Mathematically, all samples and prototypes are vectors of nx 1 (number of features) matrices, here, 3x1 steroid matrices. Initially, the prototypes are set as the means of random subsets of training samples selected from the corresponding class. The initial matrices are set as identity matrices to avoid selection bias. Prototypes and matrices are then updated by classification error, Ei-1f[(d(xi, a) - d(xi, ak)/(d(xi, a) + d(xi, ak)], to draw correct nearest prototype close and push wrong
| Variables | ACC (n = 44) | Non-ACC (n = 89) | P value |
|---|---|---|---|
| Age, y, median (IQR) | 57.00 | 63.00 | .051 |
| (38.00, 68.25) | (49.00, 71.00) | ||
| Women, n (%) | 28 (63.64%) | 48 (53.93%) | .380 |
| Mode of discovery, n (%) | |||
| Incidental | 24 (54.55%) | 55 (61.80%) | |
| Cancer staging imaging | 1 (2.27%) | 8 (8.99%) | .171 |
| Other | 19 (43.18%) | 26 (29.21%) | |
| Location, n (%) | |||
| Bilateral | 0 (0.00%) | 12 (13.48%) | |
| Left | 14 (31.82%) | 46 (51.69%) | <. 001 |
| Right | 30 (68.18%) | 31 (34.83%) | |
| Size, cm, median (IQR) | 9.20 | 3.70 | <. 001 |
| (5.85, 15.00) | (2.40, 5.50) | ||
| HU, median (IQR) | 36.00 | 34.00 | .016 |
| (32.50, 42.50) | (25.00, 40.00) | ||
| Hormone excess, n (%) | <. 001 | ||
| Cortisol excess | 15 (34.09%) | 17 (19.10%) | |
| Mineralocorticoid excess | 0 (0.00%) | 7 (7.87%) | |
| Hyperandrogenism | 3 (6.82%) | 1 (1.12%) | |
| Catecholamine excess | 0 (0.00%) | 3 (3.37%) | |
| Combined hormone excess | 17 (38.64%) | 1 (1.12%) | |
| None | 9 (20.45%) | 60 (67.42%) | |
| 17-hydroxypregnenolone, ng/dL, median (IQR) | 317.50 | 44.00 | <. 001 |
| (133.75, 790.75) | (24.00, 83.00) | ||
| 11-deoxycortisol, ng/dL, median (IQR) | 316.50 | 31.00 | <. 001 |
| (109.25, 1662.50) | (19.00, 67.00) | ||
| 17-hydroxyprogesterone, ng/dL, median (IQR) | 125.00 | 42.00 | <. 001 |
| (69.00, 214.25) | (40.00, 76.00) | ||
| Diagnosis of adrenal mass, n (%) | — | ||
| ACC | 44 (100.00%) | — | |
| Other malignant | — | 26 (29.21%) | |
| Pheochromocytoma | — | 5 (5.62%) | |
| Benign cortical | — | 43 (48.31%) | |
| Other benign | — | 15 (16.85%) |
P values were calculated via Mann-Whitney U-test or Chi-squared test (category). Abbreviations: ACC, adrenocortical carcinoma; HU, Hounsfield unit; IQR, interquartile range.
nearest prototype away. [(d(xi, @)- d(xi, @k))/(d(xi, @) + d(xi, @k))] is in the range of [-1, 1] with a negative value indi- cating correct classification. In this study, d(xi, @ACC) and d(xi, @non-ACC) denoted the adaptive Euclidean distance, d(xi, @ACC/non-ACC)=(xi, @ACC/non-ACC)TA(xi, @ACC/non-ACC), between sample xi and prototype @ACC/non-ACC. A is a nxn (number of features, 3 x 3 in this study) matrix with 2; on the di- agonal with _; 2; = 1 and 0 in off-diagonal.
For LGMLVQ analysis, all numeric steroids were log2- -transformed and normalized by the respective mean values divided by corresponding SDs. All samples were subsequently split into training dataset and testing dataset (8:2). To avoid the selection bias, we randomly selected 90% of samples from the training dataset. We employed a single prototype for each class. The initial prototypes were set as the class con- ditional mean of the randomly selected training dataset. The initial matrices were set as identity (ie, all steroids were equally
important to identify or exclude ACC/non-ACC). We employed the sklvq package, an open-source Python implementation of LGMLVQ algorithm, with default param- eters and repeated the training process 1000 times (25). The final @ACC, @non-ACC, AACC, and Anon-ACC were the mean value of 1000 run. We employed sigmoid algorithm, 1/(1+ exp(d(x, @ACC) - d(x, @“on-ACC)/4)), to obtain serum steroid metabolites (SSMs) score for ACC vs non-ACC discrim- ination. In addition, a separate LGMLVQ model that included serum steroids, tumor size, and HU was trained for the discrim- ination of ACC and non-ACC. Patients with adrenal masses were further divided into 5 groups, ACC, adrenocortical aden- oma (ACA), other benign mass, other malignancy, and pheo- chromocytoma and paraganglioma. Similarly, we employed LGMLVQ to determine the prototype and SSMs for each class.
Then, to evaluate the discrimination of LGMLVQ scores and to determine the best threshold to distinguish ACC and
| Variables | ACC (n = 44) | Other malignant (n = 27) | P value |
|---|---|---|---|
| Age, y, median (IQR) | 57.00 | 63.00 | .069 |
| (38.00, 68.25) | (51.00, 71.00) | ||
| Women, n (%) | 28 (63.64%) | 12 (44.44%) | .181 |
| Mode of discovery, n (%) | .028 | ||
| Incidental | 24 (54.55%) | 11 (40.74%) | |
| Cancer staging imaging | 1 (2.27%) | 6 (22.22%) | |
| Other | 19 (43.18%) | 10 (37.04%) | |
| Location, n (%) | |||
| Bilateral | 19 | 2 (7.41%) | |
| Left | 14 (31.82%) | 15 (55.56%) | .013 |
| Right | 30 (68.18%) | 10 (37.04%) | |
| Size, cm, median (IQR) | 9.20 | 6.30 | .026 |
| (5.85, 15.00) | (3.40, 9.50) | ||
| HU, median (IQR) | 36.00 | 38.00 | .275 |
| (32.50, 42.50) | (35.50, 43.50) | ||
| Hormone excess, n (%) | <. 001 | ||
| Cortisol excess | 15 (34.09%) | 0 (0.00%) | |
| Hyperandrogenism | 3 (6.82%) | 0 (0.00%) | |
| Combined hormone excess | 17 (38.64%) | 0 (0.00%) | |
| None | 9 (20.45%) | 27 (100.00%) | |
| 17-hydroxypregnenolone, ng/dL, median (IQR) | 317.50 | 35.00 | <. 001 |
| (133.75, 790.75) | (22.50, 81.50) | ||
| 11-deoxycortisol, ng/dL, median (IQR) | 316.50 | 31.00 | <. 001 |
| (109.25, 1662.50) | (21.00, 50.50) | ||
| 17-hydroxyprogesterone, ng/dL, median (IQR) | 125.00 | 40.00 | <. 001 |
| (69.00, 214.25) | (40.00, 71.00) |
P values were calculated via Mann-Whitney U test or chi-squared test (category). Abbreviations: ACC, adrenocortical carcinoma; HU, Hounsfield unit; IQR, interquartile range.
non-ACC, we randomly selected 90% samples from testing da- taset to determine receiver operating characteristics (ROC) for 100 times. Vertical mean ROC with fixed false-positive rate and threshold mean ROC with fixed threshold were deter- mined. The best cutoff for each run was determined by Youden’s Index. Based on the mean best cutoff, true-positive rate, false-positive rate, positive predictive value, and negative predictive value over 100 testing were summarized as mean and SD. Aiming to improve false-negative rate or false-positive rate, the discriminative performance of higher or lower cutoff values were also assessed in a similar way. Both the discrimin- ation of different models and the diagnostic performance of the established models between women and men were assessed by 1000 sample permutations during each run of test.
All analyses were conducted by Python for macOS (version 3.9.6) and R for macOS (version 4.0.2). P <. 05 was consid- ered as statistically significant.
Results
Patients
Of 263 patients included in this study (median age of 62 years, 60.8% women), ACA was diagnosed in 161 (61.2%), ACC in 44 (16.7%), other malignant adrenal masses in 27 (10.3%), oth- er benign adrenal masses in 25 (9.5%), and pheochromocytomas in 6 (2.3%) (Table 1). In most patients, adrenal mass was
discovered incidentally (175, 66.5%). Median tumor size was 3.1 cm (interquartile range 1.8-5.6), median unenhanced HU of 20 (6.3-35.0), and 42 (16.0%) patients’ adrenal mass was bi- lateral. Based on routine clinical hormonal workup, the majority of adrenal masses were classified as nonfunctioning (166, 63.1%) (Table 1). Cortisol excess was demonstrated in 54 (20.5%) patients, mineralocorticoid excess in 16 (6.1%), com- bined adrenal hormone excess in 19 (7.2%), androgen excess in 4 (1.5%), and catecholamine excess in 4 (1.5%). Notably, all ACCs demonstrated unenhanced HU ≥20, and most other malignant adrenal masses demonstrated HU ≥ 20, with only 1 adrenal mass measuring between 10 and 20 HU. Adrenal aden- omas and other benign adrenal masses presented a spectrum of unenhanced HU measurements, with 58 (31.2%) in the ≥ 20 HU category (Table 1).
In the 133 patients with adrenal mass with HU ≥ 20, ACC was diagnosed in 44 (Table 2). No differences in age, sex, or mode of discovery were noted between the 2 groups. None of ACCs was bilateral (as opposed to 12, 13.5%) of non-ACCs (Table 2). ACCs were significantly larger (median size of 9.2 cm vs 3.7 cm of non-ACCs). Notably, concentra- tions of 17OH-pregnenolone, 17OH-progesterone, and 11-deoxycortisol were 3- to 10-fold higher in ACCs vs non-ACCs (Table 2). Similar trends were seen when limiting the analysis only to patients with adrenal mass ≥4 cm with HU ≥ 20 (Supplementary Table S1) (21).
| Variables | Functioning ACC (n = 35) | Functioning lipid poor adenomas (n = 19) | P value |
|---|---|---|---|
| Age, y, median (IQR) | 57.00 | 61.00 | .574 |
| (34.00, 69.00) | (43.00, 68.50) | ||
| Women, n (%) | 22 (62.86%) | 13 (68.42%) | .912 |
| Mode of discovery, n (%) | .612 | ||
| Incidental | 17 (48.57%) | 12 (63.16%) | |
| Cancer staging imaging | 1 (2.86%) | 0 (0.00%) | |
| Other | 17 (48.57%) | 7 (36.84%) | |
| Location, n (%) | .004 | ||
| Bilateral | 0 (0.00%) | 2 (10.53%) | |
| Left | 11 (31.43%) | 12 (63.16%) | |
| Right | 24 (68.57%) | 5 (26.32%) | |
| Size, cm, median (IQR) | 8.70 | 4.20 | <. 001 |
| (5.70, 14.35) | (3.05, 5.10) | ||
| HU, median (IQR) | 36.00 | 30.00 | .003 |
| (31.00, 43.00) | (23.50, 34.50) | ||
| Hormone excess, n (%) | |||
| Cortisol excess | 15 (42.86%) | 16 (84.21%) | |
| Mineralocorticoid excess | 0 (0.00%) | 2 (10.53%) | <. 001 |
| Hyperandrogenism | 3 (8.57%) | 1 (5.26%) | |
| Combined hormone excess | 17 (48.57%) | 0 (0.00%) | |
| 17-hydroxypregnenolone, ng/dL, median (IQR) | 355.00 | 42.00 | <. 001 |
| (137.50, 808.50) | (20.00, 123.00) | ||
| 11-deoxycortisol, ng/dL, median (IQR) | 636.00 | 78.00 | <. 001 |
| (149.50, 1985.00) | (39.00, 108.00) | ||
| 17-hydroxyprogesterone, ng/dL, median (IQR) | 157.00 | 60.00 | <. 001 |
| (86.50, 242.00) | (40.00, 81.50) |
P values were calculated via Mann-Whitney U test or chi-squared test (category). Abbreviations: ACC, adrenocortical carcinoma; HU, Hounsfield unit; IQR, interquartile range.
When comparing patients with ACC with patients diag- nosed with other adrenal malignancies (Table 3), we found that ACCs were larger (median size of 9.2 cm vs 6.3 cm, P =. 026), but demonstrated similar unenhanced HU (median of 36 vs 38). Concentrations of 17OH-pregnenolone, 17OH-progesterone, and 11-deoxycortisol were significantly higher in patients diagnosed with ACCs vs other malignant adrenal masses (Table 3). Similarly, we found that patients with nonfunctioning ACC still had higher level of 17OH- pregnenolone and 11-deoxycortisol (Supplementary Table S2) (21). However, there were no significant differences in 17OH-progesterone and tumor size between groups (nonfunc- tioning ACC vs other malignancy) (Supplementary Table S2) (21).
We further examined a group of patients with adrenal mass ≥2 cm and HU ≥ 20 HU who demonstrated adrenal hormone excess (Table 4). In this category, no differences in age, sex, or mode of adrenal mass discovery were noted between groups. However, ACCs were larger (median tumor size of 8.7 vs 4.2 cm), had slightly higher unenhanced HU (36 vs 30), and were more likely to have combined hormone excess (49% vs 0%). Even in this group, concentrations of 17OH- pregnenolone, 17OH-progesterone, and 11-deoxycortisol were much higher in ACCs vs functioning lipid poor aden- omas (Table 4).
Computational Analysis of Serum Steroid Data LGMLVQ successfully classified ACC and non-ACC based on the adaptive Euclidean distances that derived from the val- ues of 17OH-pregnenolone, 11-deoxycortisol, and 17OH- progesterone (Fig. 2A). However, LGMLVQ was not able to separate adrenal adenomas, other benign, other malignant ad- renal masses, or pheochromocytoma and paraganglioma (Supplementary Fig. S1) (21). Based on 1000 runs, and the mean relevance matrix specific to ACC, we found that 17OH-pregnenolone and 11-deoxycortisol were the most informative steroid precursors in identifying ACC and non-ACC (Fig. 2B). 17OH-pregnenolone was slightly more important for identifying ACC (Fig. 2B, top), whereas 11-deoxycortisol was more relevant in diagnosing non-ACC (Fig. 2B, bottom).
ROC was applied for each run of test with 90% randomly se- lected samples from testing dataset (Fig. 2C, dotted line). Our results showed that SSM score derived from LGMLVQ had a mean vertical area under the ROC (AUC) of 0.961 (Fig. 2C, sol- id black line), mean threshold AUC of 0.823 and SD of 0.008 (Fig. 2C, solid navy line). According to our results, the best threshold of SSM score for the discrimination of ACC and non-ACC was 0.40. Our results demonstrated that LGMLVQ had a mean true-positive rate of 89.5% (SD: 4.6%), mean false- positive rate of 8.7% (1.5%), mean false-negative rate of 10.5%
A
ACC
non-ACC
B
1.00
Training
w.r.t. AACC
ACC
6
4
Weight
0.50
2
0.51 (0.11)
0.48 (0.11)
0
0.01 (0.02)
-2
0.00
-4
1.00
w.r.t. A
non-ACC
non-ACC
6
0.27 (0.20)
4
Weight
0.50
2
0.68 (0.19)
0.05
0
(0.03)
-2
0.00
17OH-preg
11DCORT
17OH-prog
4
-7.5 -5.0 -2.5 0.0 2.5
C
Testing
w.r.t. AACC
1.00
6
True Positive Rate
0.80
4
0.60
2
0.40
0
0.20
ROC,
ROCy (AUC 0.961)
-2
ROC- (AUC 0.823)
0.00
0.00
0.25 0.50 0.75 1.00
-4
w.r.t. A
non-ACC
False Positive Rate
D
6
4
2
0
-2
4
-7.5 -5.0 -2.5 0.0 2.5
| ACC (mean±SD) | |
|---|---|
| TPR (%) | 89.5 ± 4.6 |
| FPR (%) | 8.7 ± 1.5 |
| FNR (%) | 10.5 ±4.6 |
| PPV (%) | 67.4 ±4.3 |
| NPV (%) | 97.8 ± 1.0 |
| Accuracy | 91.0 ± 1.5 |
Figure 2. Results of LGMLVQ analysis. (A) Projection of individual samples (small circles) and prototypes @ACC and @non-ACC (large circles with solid border) with reference to different relevance matrices AACC and A non-ACC (top, training dataset [n = 210]; bottom, testing dataset [n = 53]). Ninety percent of samples were randomly selected from the training dataset every single run. The prototypes (both top and bottom) were the average prototypes derived from 1000 randomized training process. (B) The relevance matrices specific for ACC and non-ACC, average matrices derived from 1000 randomized training process. The relative weight of individual steroid was shown as mean ± SD (error bar). (C) ROC curve for the discrimination of ACC and non-ACC based on 100 randomized testing process (dotted line). The ROC curve was averaged based on fixed threshold (ROCT) and fixed false-positive rate (ROCv), separately. The shadow showed the SD of the ROC curve over 100 randomly testing. (D) The diagnostic performance of serum steroid metabolite score at a cutoff of 0.4 during 100 randomly testing, showed as mean ± SD.
Abbreviations, ACC, adrenocortical carcinoma; FNR, false-negative rate; FPR, false-positive rate; NPV, negative predictive value; PPV, positive predictive value; ROC, receiver op- erating characteristics; TPR, true-positive rate.
A
200
☒ TP = ACC
Number of patients with positive results
FP = non-ACC
150
100
50
0
B
ACC
non-ACC
C
ACC
OM
ACA
OB
PPGL
0.7-
0.7-
0.6
0.6
2
Steroids
00.5
0
0.4
0.4
0.3
0.3
5
5
10
S
15
20
40
10
0
Size
15
40
0
0
20
20
-60
-40
-20
HU
20
-20
-60
-40
HU
| Model | Serum steroids | Serum steroids+HU | Serum steroids+HU+size | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Score | 0.29 | 0.40 | 0.61 | 0.26 | 0.41 | 0.66 | 0.23 | 0.38 | 0.61 |
| FNR (%) | 0.0 | 15.9 | 56.8 | 0.0 | 18.2 | 56.8 | 0.0 | 15.9 | 47.7 |
Figure 3. Diagnostic accuracy of serum steroids for the discrimination of ACC. (A) Diagnostic performance of serum steroids alone or in combination with imaging characteristics. (B) Scatter plot of imaging characteristics and steroid profiling features of ACC and non-ACC. (C) Scatter plot of imaging characteristics and steroid profiling features of multiclass adrenal masses.
Abbreviations, ACA, adrenocortical adenoma; ACC, adrenocortical carcinoma; FNR, false-negative rate; FP, false positive; HU, Hounsfield unit; OB, other benign; OM, other ma- lignant; PPGL, pheochromocytoma and paraganglioma; TP, true positive.
(4.6%), mean positive predictive value of 67.4% (4.3%), mean negative predictive value of 97.8% (1.0%), and mean accuracy of 91.0% (1.5%) based on 100 analyses using testing dataset (Fig. 2D). Based on the entire study population, we found the diagnostic performances of SSM score was similar in women and men (Supplementary Table S3) (21). We found that based on an SSM score of 0.29, no ACC was missed (Fig. 3, Supplementary Tables S3 and 4) (21).
Computational Data of Steroid and Imaging Data
Tumor size and HU alone, as well as in combination yielded a poor accuracy to diagnose ACC (Supplementary Table S5) (21). In addition to serum steroids, we also integrated serum steroids, tumor size, and HU measurements in developing a ma- chine learning model to diagnose ACC (Fig. 3). We have found a minor improvement in accuracy in diagnosing ACC when us- ing serum steroids combined with tumor size and HU in com- parison with using serum steroids alone (Supplementary Fig. S2, Supplementary Table S6) (21).
Discussion
In this study, we show that 11-deoxycortisol, 17OH-pregnenolone, and 17OH-progesterone concentrations are approximately 3- to 8-fold higher in patients with ACC vs patients with other adrenal masses. We further developed a novel machine learning model (LGMLVQ) using the 3 serum steroids and demonstrate a good discrimination to distinguish ACC from other adrenal masses with a mean vertical AUC of 0.961. We also integrated imaging characteristics together with serum steroids demon- strating no significant advantage of combined vs serum steroid only approach.
When examining studies with at least 15 patients with ACC, 2 previous studies employed plasma multisteroid profiling to diagnose ACC (20, 21). In 1 study of 66 patients with ACA and 42 patients with ACC, authors measured 15 steroids and showed higher concentrations of several steroids, including 11-deoxycortisol and 17OH-progesterone. Authors applied lo- gistic regression modeling that included 6 steroids yielding an AUC of 0.95 in men and 0.94 in women (20). In another study
that included 577 patients with an adrenal mass, but only 19 patients with ACC, plasma steroid profiling demonstrated simi- lar results, with 8 steroids yielding a sensitivity of 74% and spe- cificity of 98% (21). 17OH-pregnenolone was not one of the steroids analyzed in either study (20, 21). Although 1 study pro- vided an additional analysis considering tumor size (21), none looked at stratified analysis based on unenhanced HU (20, 21). Notably, plasma multisteroid profiling is not widely avail- able, and, as such, has not been integrated in clinical practice.
In our study, we showed that 11-deoxycortisol, 17OH- pregnenolone, and 17OH-progesterone concentrations were high in ACCs. These 3 steroids discriminated ACCs from other malignancies, as well as from ACAs. Even when we limited our analysis to functioning adrenal masses or indeterminate masses (HU>20), we found similar results. Previous studies demonstrated similar results regarding 11-deoxycortisol and 17OH-progesterone (19-21). Similarly, finding elevated 17OH-pregnenolone concentrations in patients with ACC was not surprising, considering that 24-hour urine-based multi- steroid profiling identified increased urine metabolites of 17OH-pregnenolone in patients with ACC (19). We further showed that SSM of 0.40 based on our machine learning model distinguished ACC from non-ACC with a mean threshold fixed AUC of 0.823. We found that SSM of 0.29 had a mean false- negative rate of 0% and SSM of 0.61 had a mean false-positive rate of 0% in diagnosing ACC. Applying machine learning ana- lysis, 3 steroids demonstrated a mean false-positive rate fixed AUC of 0961.
Notably, all ACCs in our study demonstrated HU>20. This finding is consistent with previous reports in which 99% to 100% of ACCs has HU >20 or were heterogeneous (5, 10). However, although HU <20 makes ACC unlikely, most ad- renal masses with HU >20 are not ACCs, and as such HU alone cannot accurately diagnose ACC (1, 5, 9, 12). In a pre- vious study of urine steroid profiling, adding imaging charac- teristics to the diagnostic algorithm improved the diagnostic accuracy of urine steroid profiling (5). When we combined im- aging data with serum steroid data, we also found a minor im- provement in the diagnostic accuracy of ACC.
Comparing diagnostic accuracy of serum steroids to urine steroid profiling is difficult, as no studies to date measured both in patients with adrenal masses. Considering that 24-hour urine steroid metabolome is a more comprehensive as- sessment of one’s steroid secretion (26), whereas serum steroids have a disadvantage of lower concentrations and diurnal vari- ation, it is likely that serum steroid profiling may not be as accur- ate as urine steroid assessment. On the other hand, serum steroid measurements may serve as an easier, more widely available ini- tial diagnostic test that may be followed by a more comprehen- sive analysis in selected patients. Guidelines already suggest measurement of 11-deoxycortisol, 17OH-progesterone, and de- hydroepiandrosterone sulfate (17), though serum steroid precur- sor measurement in clinical practice is performed in less than a quarter of patients with ACC (18). Our data suggest that adding 17-OH-pregnenolone to the initial diagnostic workup is valu- able and using the proposed 3-steroid machine learning model further improves the diagnostic accuracy of indeterminate ad- renal tumors.
The strengths of this study include analysis of serum steroid measurements that are widely available clinically and thus more readily available to integrate in most practices, consideration of imaging characteristics, a stringent reference standard, and use of machine learning analysis. Limitations include referral bias, a
relatively small sample size, especially for patients with malig- nant adrenal masses, and inability to include participants without any or sufficient biobanked biomaterial. External mul- ticenter prospective validation is needed before widespread in- tegration in clinical practice.
In conclusion, we showed that measurement of 17OH- progesterone, 17OH-pregnenolone, and 11-deoxycortisol is valuable in distinguishing ACC from non-ACC overall and in a subgroup analysis of adrenal masses with HU >20.
Funding
This research was partially supported by the Melinda Nelson ACC Fund, by a small grant from the general surgery depart- ment, Mayo Clinic, Rochester, and by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) and National Institute of Aging (NIA) of the National Institutes of Health (NIH) USA under award K23DK121888, R03DK132121, and R03AG71934 (to I.B.). The views expressed are those of the author and not ne- cessarily those of the NIH.
Disclosures
I.B. reports consulting, advisory board, or data safety moni- toring board participation fees (to institution) from Diurnal, Neurocrine, Spruce, Adrenas, Recordati, Corcept, Sparrow, Xeris, AstraZeneca, NovoNordisk, Crinetics, Camurus, and HRA Pharma, outside this work. I.B. received research fund- ing (to institution) from Recordati and HRA Pharma for an investigator-initiated award, outside this work. All other au- thors declared no conflict of interest.
Data Availability
Some data, including prototypes and relevant matrices, for this study were not publicly available but available from cor- responding author upon reasonable request.
References
1. Ebbehoj A, Li D, Kaur RJ, et al. Epidemiology of adrenal tumours in Olmsted County, Minnesota, USA: a population-based cohort study. Lancet Diabetes Endocrinol. 2020;8(11):894-902.
2. Fassnacht M, Tsagarakis S, Terzolo M, et al. European Society of Endocrinology clinical practice guidelines on the management of adrenal incidentalomas, in collaboration with the European net- work for the study of adrenal tumors. Eur J Endocrinol. 2023;189(1):G1-G42.
3. Bancos I. Adrenal incidentalomas: insights into prevalence. Ann Intern Med. 2022;175(10):1481-1482.
4. Jing Y, Hu J, Luo R, et al. Prevalence and characteristics of adrenal tumors in an unselected screening population: a cross-sectional study. Ann Intern Med. 2022;175(10):1383-1391.
5. Bancos I, Taylor AE, Chortis V, et al. Urine steroid metabolomics for the differential diagnosis of adrenal incidentalomas in the EURINE-ACT study: a prospective test validation study. Lancet Diabetes Endocrinol. 2020;8(9):773-781.
6. James BC, Aschebrook-Kilfoy B, Cipriani N, Kaplan EL, Angelos P, Grogan RH. The incidence and survival of rare cancers of the thy- roid, parathyroid, adrenal, and pancreas. Ann Surg Oncol. 2016;23(2):424-433.
7. Chandrasekar T, Goldberg H, Klaassen Z, et al. The who, when, and why of primary adrenal malignancies: insights into the epi- demiology of a rare clinical entity. Cancer. 2019;125(7): 1050-1059.
8. Zhang H, Naji Y, Yan M, Lian W, Xie M, Dai Y. Development and validation of prognostic nomograms in patients with adrenocorti- cal carcinoma: a population-based study. Int Urol Nephrol. 2020;52(6):1057-1071.
9. Bancos I, Prete A. Approach to the patient with adrenal incidenta- loma. J Clin Endocrinol Metab. 2021;106(11):3331-3353.
10. Iniguez-Ariza NM, Kohlenberg JD, Delivanis DA, et al. Clinical, biochemical, and radiological characteristics of a single-center retrospective cohort of 705 large adrenal tumors. Mayo Clin Proc Innov Qual Outcomes. 2018;2(1):30-39.
11. Yalon T, Yalon M, Assaf D, et al. Differentiating between adreno- cortical carcinoma and lipid-poor cortical adenoma: a novel cross- sectional imaging-based score. Surgery. 2023;173(1):35-42.
12. Mao JJ, Dages KN, Suresh M, Bancos I. Presentation, disease progression and outcomes of adrenal gland metastases. Clin Endocrinol (Oxf). 2020;93(5):546-554.
13. Levine AC, Mitty HA, Gabrilove JL. Steroid content of the periph- eral and adrenal vein in Cushing’s syndrome due to adrenocortical adenoma and carcinoma. J Urol. 1988;140(1):11-15.
14. Bertagna C, Orth DN. Clinical and laboratory findings and results of therapy in 58 patients with adrenocortical tumors admitted to a single medical center (1951 to 1978). Am J Med. 1981;71(5): 855-875.
15. Bulaschenko H, Dohan FC, Richardson EM, Touchstone JC. The excretion of pregnane-3alpha, 17alpha, 21-triol-20-one (tetrahy- dro S) in normal and pathologic urine. J Clin Endocrinol Metab. 1957;17(2):250-255.
16. Arlt W, Biehl M, Taylor AE, et al. Urine steroid metabolomics as a biomarker tool for detecting malignancy in adrenal tumors. J Clin Endocrinol Metab. 2011;96(12):3775-3784.
17. Fassnacht M, Dekkers OM, Else T, et al. European Society of Endocrinology clinical practice guidelines on the management of adrenocortical carcinoma in adults, in collaboration with the European network for the study of adrenal tumors. Eur J Endocrinol. 2018;179(4):G1-G46.
18. Sada A, Foster TR, Al-Ward R, et al. The effect of hormonal secre- tion on survival in adrenocortical carcinoma: a multi-center study. Surgery. 2024;175(1):80-89.
19. Taylor DR, Ghataore L, Couchman L, et al. A 13-steroid serum panel based on LC-MS/MS: use in detection of adrenocortical car- cinoma. Clin Chem. 2017;63(12):1836-1846.
20. Schweitzer S, Kunz M, Kurlbaum M, et al. Plasma steroid metabo- lome profiling for the diagnosis of adrenocortical carcinoma. Eur J Endocrinol. 2019;180(2):117-125.
21. Berke K, Constantinescu G, Masjkur J, et al. Plasma steroid profil- ing in patients with adrenal incidentaloma. J Clin Endocrinol Metab. 2022;107(3):e1181-e1192.
22. Funder JW, Carey RM, Mantero F, et al. The management of pri- mary aldosteronism: case detection, diagnosis, and treatment: an endocrine society clinical practice guideline. J Clin Endocrinol Metab. 2016;101(5):1889-1916.
23. Schneider P, Biehl M, Hammer B. Adaptive relevance matrices in learning vector quantization. Neural Comput. 2009;21(12): 3532-3561.
24. Schneider P, Biehl M, Hammer B. Distance learning in discrimina- tive vector quantization. Neural Comput. 2009;21(10):2942-2969.
25. Veen RV, Biehl M, de Vries G-J. Sklvq: Scikit learning vector quant- ization. J Mach Learn Res. 2021;22(231):1-6.
26. Storbeck KH, Schiffer L, Baranowski ES, et al. Steroid metabolome analysis in disorders of adrenal steroid biosynthesis and metabol- ism. Endocr Rev. 2019;40(6):1605-1625.
Downloaded from https://academic.oup.com/jcem/article/110/4/1177/7749469 by WT Cox Information Services user on 04 April 2026