Objectives: Use of machine learning (ML) in diagnostics offers promise to optimise interpretation of laboratory data and guide clinical decision-making. For this, ML-based outputs should provide robustly reproducible results at least as good as the underlying laboratory data. The objective of this study was to assess robustness of ML-based steroid-probability-scores for diagnosis of primary aldosteronism (PA). Methods: Reproducibility of ML-based steroid-probability-scores was assessed from coefficients of variation (CVs) for pools of quality control plasma from selected groups of patients with and without PA. Intra-patient measurement variability was assessed from CVs of three consecutive plasma specimens obtained on different days from 77 patients. Inter-laboratory reproducibility was assessed from 47 duplicate plasma specimens analysed in two different laboratories. Results: Support vector machine-derived steroid-probability-scores for diagnosis of PA for seven sets of quality control plasma pools yielded an averaged CV (2.5 % CI 0.4–4.4 %) that was lower (p=0.0078) than the averaged CV for seven steroids employed in that model (12.0 % CI 7.4–16.6). Using three sets of plasma samples from 77 patients, CVs for intra-patient measurement variability of steroid-probability-scores were 7 % (CI 5–9 %) and lower (p<0.0001) than CVs for measurements of aldosterone (38 % CI 32–42 %), 18-oxocortisol (36 % CI 29–43 %), 18-hydroxycortisol (25 % CI 21–28 %) and the aldosterone:renin ratio (46 % CI 38–55 %). ML-derived probability scores for 47 duplicate plasma samples analysed at two separate laboratories displayed excellent agreement and negligible bias. Conclusions: ML-based steroid-probability-scores for diagnosis of PA display remarkably high robustness according to reproducibility of measurements within and between laboratories as well as within patients.