Simplified urinary steroid profiling by LC-MS as diagnostic tool for malignancy in adrenocortical tumors

Preoperative identification of malignant adrenal tumors is challenging. 24-h urinary steroid profiling by LC-MS/MS and machine learning has demonstrated high diagnostic power, but the unavailability of bioinformatic models for public use has limited its routine application. We here aimed to increase usability with a novel classification model for the differentiation of adrenocortical adenoma (ACA) and adrenocortical carcinoma (ACC).

Eleven steroids (5-pregnenetriol, dehydroepiandrosterone, cortisone, cortisol, α-cortolone, tetrahydro-11-deoxycortisol, etiocholanolone, pregnenolone, pregnanetriol, pregnanediol, and 5-pregnenediol) were quantified by LC-MS/MS in 24-h urine samples from 352 patients with adrenal tumor (281 ACA, 71 ACC). Random forest modelling and decision tree algorithms were applied in training (n = 188) and test sets (n = 80) and independently validated in 84 patients with paired 24-h and spot urine.

After examining different models, a decision tree using excretions of only 5-pregnenetriol and tetrahydro-11-deoxycortisol classified three groups with low, intermediate, and high risk for malignancy. 148/217 ACA were classified as being at low, 67 intermediate, and 2 high risk of malignancy. Conversely, none of the ACC demonstrated a low-risk profile leading to a negative predictive value of 100% for malignancy. In the independent validation cohort, the negative predictive value was again 100% in both 24-h urine and spot urine with a positive predictive value of 87.5% and 86.7%, respectively.

This simplified LC-MS/MS-based classification model using 24-h-urine provided excellent results for exclusion of ACC and can help to avoid unnecessary surgeries. Analysis of spot urine led to similarly satisfactory results suggesting that cumbersome 24-h urine collection might be dispensable after future validation.