The development of liquid chromatography tandem mass spectrometry (LC-MS/MS) allows clinical laboratory to propose targeted and quantitative steroidomics analysis for diagnostic and follow-up explorations. Quantitative profiles of circulating steroids are now available in one single run. However, the screening and diagnostic thresholds for LC-MS/MS are yet to be completely defined in both basal and post synacthen states. These thresholds are likely to differ from those established with immunoassays due to the enhanced specificity and sensitivity of LC-MS/MS, a method less prone to cross-reactivity and interferences (Fiet et al., 2017). The concept of profiling or multiplexing the circulating steroidome is a paradigm shift as it appears to be way more accurate than single specie immunoassay quantitation to stratify patients. In subnormal or atypical cases, we showed that an exhaustive profile in a basal state (i.e. without ACTH stimulation test) could differentiate 21-hydroxylase deficiency from other enzyme defects and establish a more accurate diagnosis (Fiet et al., 2017). In parallel to these analytical progresses, machine learning (ML) approaches are emerging as essential tools in diagnosis thank to its classification strength compare to classical univariate statistics. Our main hypothesis is that these statistical models can be all the more powerful as, when combined with the metabolic profiling of the steroid biosynthesis pathway, it will be possible to avoid, or at least, limit the use of ACTH stimulation tests. In our experience, ML is indeed, particularly adapted to convert these metabolomic signatures, potentially concatenated with other bio-clinical parameters, into straightforward diagnostic outputs for clinicians.

液相色谱串联质谱（LC-MS / MS）的发展使临床实验室可以提出针对性和定量的类固醇组学分析，以进行诊断和后续探索。现在可以一次运行获得循环类固醇的定量曲线。但是，LC-MS / MS的筛查和诊断阈值在基础状态和后合成状态均尚未完全定义。由于LC-MS / MS的特异性和敏感性增强，这些阈值可能不同于免疫测定所建立的阈值，该方法较不易发生交叉反应和干扰（Fiet et al。，2017）。对循环类固醇组进行分析或复用的概念是一种范式转变，因为它似乎比单种免疫测定定量法更准确地对患者进行分层。在亚正常或非典型病例中，我们显示了基础状态下的详尽图谱（即没有ACTH刺激试验）可以将21-羟化酶缺乏症与其他酶缺陷区分开，并建立更准确的诊断（Fiet等人，2017）。与这些分析进展并行的是，由于机器学习（ML）方法的分类强度与经典单变量统计数据相比，因此正在成为诊断中的重要工具。我们的主要假设是，这些统计模型可能更强大，因为当与类固醇生物合成途径的代谢谱分析相结合时，将有可能避免或至少限制ACTH刺激试验的使用。根据我们的经验，ML确实特别适合转换这些代谢组学特征，