Predicting treatment outcome in congenital adrenal hyperplasia using urine steroidomics and machine learning.

Objective: Treatment monitoring of individuals with congenital adrenal hyperplasia (CAH) remains unsatisfactory. Comprehensive 24 h urine steroid profiling provides detailed insight into adrenal steroid pathways. We investigated whether 24 h urine steroid profiling can predict treatment control in children and adolescents with CAH using machine learning (ML).

Design: Prospective observational cohort study.

Methods: This study included children with 21-hydroxylase deficiency. On 24 h urines of 2 consecutive visits 40 steroids were measured by gas chromatography-mass spectrometry. Treatment outcome was clinically classified as undertreated, optimally treated or overtreated. We used sparse partial least squares discriminant analysis (sPLS-DA) to investigate prediction of treatment outcome. We computed area under the ROC-curve (AUC) of 2 sPLS-DA models: (1) using only 24 h urine metabolites and (2) adding clinical variables.

Results: We included 112 visits (68 optimal, 44 undertreatment) from 59 patients: 27 (46%) girls, 46 (78%) classic CAH, and 19 (32%) prepubertal. Mean age at first visit was 11.9 ± 4.0 years and mean BMI SDS 0.6 ± 1.1. SPLS-DA using 24 h urine metabolites showed clear clustering of optimally treated patients on 2 components, while undertreated patients were more heterogeneous (AUC 0.88). The model selected pregnanetriol and 17α-hydroxypregnanolone contributing to excluding optimal treatment and 5 metabolites contributing to excluding undertreatment: 17β-estradiol, cortisone, tetrahydroaldosterone, androstenetriol, and etiocholanolone. Addition of clinical variables marginally improved classification (AUC 0.90).

Conclusions: Using ML on 24 h urine steroid profiling predicted treatment outcome in children with CAH, even in the absence of clinical data, suggesting that routine comprehensive 24 h urine steroid profiling could improve treatment monitoring in CAH.