MSACL 2019 EU Abstract

Multidimensional Diagnostics with Machine Learning for Primary Aldosteronism

Graeme Eisenhofer,1 Claudio Durán,1 Carlo Vittorio Cannistaci,1 Tracy Ann Williams,2,3 Aleksander Prejbisz,4 Mirko Peitzsch,1 Andrzej Januszewicz,4 Felix Beuschlein,5 Paolo Mulatero,2 Jacques W.M. Lenders1 and Martin Reincke3
(1) Technische Universität Dresden, Dresden, Germany; (2) University of Torino, Torino, Italy; (3) Universität München, Munich, Germany; (4) Institute of Cardiology, Warsaw, Poland; (5) Universitätsspital Zürich, Switzerland.(

Warning: Undefined variable $headshot in /var/www/html/view_abstract/view_abstract_in_program.php on line 704

Graeme Eisenhofer (Presenter) Technische Universität Dresden

Presenter Bio: Graeme Eisenhofer received his PhD in 1983 from the University of Otago, New Zealand, with clinical research on autonomic and neuroendocrine systems. He moved to the NIH in 1985, followed by the Baker Research Institute (Melbourne) in 1988, before returning to the NIH in 1991. At the NIH he developed measurements of plasma free metanephrines as a biochemical test for diagnosis of pheochromocytoma and was responsible the first ever synthesis of 18F-fluorodopamine as a PET imaging agent for localizing catecholamine-producing tumors. In 2007 he took up a Professorship at the University Hospital Dresden, where he is the coordinator of a Clinical Research Unit focusing on adrenal hypertension and disorders of adrenal function.

Relevant Financial Disclosures (within past 24 months) Committee/Board/Advisory Board VHL Family Alliance, MSCAL Endocrine committee Royalty / IP / Other Income books etc Discussing Products? NO, presenter will NOT present educational content related to products/services of the ineligible company(ies) listed above. Conflict mitigation NOT REQUIRED.

Background: Mass spectrometric steroid panels have potential for efficient single-sample methods for diagnosis of multiple disorders of steroidogenesis. Moreover within a specific disorder, such methods can facilitate subtyping for more rapid therapeutic intervention than currently allowed. Combined with advances in computational mathematics, such as machine learning (ML), it is now possible to move from traditional uni-dimensional approaches for interpreting diagnostic data to methods that can interpret patterns in multidimensional data.

Aim: Establish an LC-MS/MS based profile of adrenal steroids that with ML can be applied to patients with primary hypertension to first diagnose those with primary aldosteronism (PA) and then further identify patients with adenomas due to KCNJ5 mutations who may specially benefit from surgical intervention.

Methods: LC-MS/MS-based plasma measurements of 15 adrenal steroids were applied to 359 primary hypertensive and 273 patients in who PA was confirmed by conventional diagnostics. Among the latter, adrenal venous sampling enabled identification of 134 patients with bilateral disease and 139 with unilateral disease confirmed by biochemical cure on follow-up after adrenalectomy. Among the latter, 58 had adenomas with KCNJ5 mutations; the other 81 were without KCNJ5 mutations. A series of nine ML algorithms and 12 different batch corrections, data normalizations and feature selection rules was applied in 10-fold cross-validation steps to establish optimal models for multiclass classification of patients with and without PA and in the former the presence of unilateral disease due to KCNJ5 mutations.

Results: After adrenalectomy, patients with adenomas due to KCNJ5 mutations exhibited both higher (P<0.01) rates of biochemical cure and reductions of blood pressure than those without mutations. According to logistic regression a selection of steroids performed similarly to routine aldosterone:renin ratios for diagnosis of patients with PA, but showed superior (P<0.005) discrimination of patients with unilateral disease. With a random forest model, initial screening for PA could be achieved at over 80% sensitivity and close to 100% specificity. On the other hand, models involving support vector machine with non-linear kernel and recursive feature elimination provided optimal identification of patients with adenomas harboring KCNJ5 mutations (>92% sensitivity, >95% specificity) and reasonable screening of patients with PA (>75% sensitivity, >95% specificity).

Conclusion: Mass spectrometry-based steroidomics provides an ideal platform for advancing multi-dimensional approaches for disease diagnosis, in this particular study directed to PA. With such approaches it should be possible to more efficiently and effectively both diagnose patients with disorders of steroidogenesis and stratify them for effective therapeutic intervention.