ECE2017 Oral Communications Adrenal-Basic & Clinical (5 abstracts)
1Institute of Metabolism and Systems Research, University of Birmingham, Birmingham, UK; 2Centre for Endocrinology, Diabetes and Metabolism, Birmingham Health Partners, Birmingham, UK; 3Johann Bernoulli Institute for Mathematics and Computer Science, University of Groningen, Groningen, The Netherlands; 4School of Computer Science, University of Birmingham, Birmingham, UK; 5Department of Pediatric Endocrinology and Diabetes, Marmara University, Istanbul, Turkey.
Background: Urinary steroid metabolite profiling is an accurate reflection of adrenal and gonadal steroid output and metabolism in peripheral target cells of steroid action. Measurement of steroid metabolite excretion by gas chromatography-mass spectrometry (GCMS) is considered reference standard for biochemical diagnosis of steroidogenic disorders. However, performance of GCMS analysis and interpretation of the resulting data requires significant expertise and age- and sex-specific reference ranges. Here we developed novel computational approaches for rapid interpretation of GCMS data for diagnosis of inborn steroidogenic disorders
Methods: We analysed the urinary steroid metabolome by GCMS in 829 healthy controls (302 neonates and infants, 167 children and 360 adults) and 118 untreated patients with genetically confirmed inborn disorders (21-hydroxylase deficiency, 17-hydroxylase deficiency, POR deficiency, 11β-hydroxylase deficiency, 3β-HSD2 deficiency, 17β-HSD3 deficiency, 5α-reductase type 2 deficiency, cytochrome b5 deficiency). We calculated age-related normative values for established metabolite ratios representing distinct enzymatic functions. We developed a novel interpretable machine learning technique, Angle Learning Vector Quantisation (ALVQ), which looks at all possible metabolite ratios, computationally reduces these to the most relevant for discrimination, and differentiates disease states by comparison to a representative prototype. The method runs independent of sex and age information, units of measurement and method of urine collection.
Results: Conventional biochemical ratios had 100% sensitivity but only very poor specificity. By contrast, ALVQ predicted affected urine vs healthy urine with 100% sensitivity and 97% specificity. For our three most prevalent conditions (PORD, SRD5A2 and CYP21A2), the specific condition was identified correctly in 96% of cases.
Conclusion: We developed a novel Steroid Metabolomics approach to automatically diagnose inborn steroidogenic disorders with very high sensitivity and specificity, superior to current methods, and with high potential for implementation in routine clinical care.