Machine learning for classification of hypertension subtypes using multi-omics: A multi-centre, retrospective, data-driven study

Abstract

Background: Arterial hypertension is a major cardiovascular risk factor. Identification of secondary hypertension in its various forms is key to preventing and targeting treatment of cardiovascular complications. Simplified diagnostic tests are urgently required to distinguish primary and secondary hypertension to address the current underdiagnosis of the latter.

Methods: This study uses Machine Learning (ML) to classify subtypes of endocrine hypertension (EHT) in a large cohort of hypertensive patients using multidimensional omics analysis of plasma and urine samples. We measured 409 multi-omics (MOmics) features including plasma miRNAs (PmiRNA: 173), plasma catechol O-methylated metabolites (PMetas: 4), plasma steroids (PSteroids: 16), urinary steroid metabolites (USteroids: 27), and plasma small metabolites (PSmallMB: 189) in primary hypertension (PHT) patients, EHT patients with either primary aldosteronism (PA), pheochromocytoma/functional paraganglioma (PPGL) or Cushing syndrome (CS) and normotensive volunteers (NV). Biomarker discovery involved selection of disease combination, outlier handling, feature reduction, 8 ML classifiers, class balancing and consideration of different age- and sex-based scenarios. Classifications were evaluated using balanced accuracy, sensitivity, specificity, AUC, F1, and Kappa score.

Findings: Complete clinical and biological datasets were generated from 307 subjects (PA=113, PPGL=88, CS=41 and PHT=112). The random forest classifier provided ∼92% balanced accuracy (∼11% improvement on the best mono-omics classifier), with 96% specificity and 0.95 AUC to distinguish one of the four conditions in multi-class ALL-ALL comparisons (PPGL vs PA vs CS vs PHT) on an unseen test set, using 57 MOmics features. For discrimination of EHT (PA + PPGL + CS) vs PHT, the simple logistic classifier achieved 0.96 AUC with 90% sensitivity, and ∼86% specificity, using 37 MOmics features. One PmiRNA (hsa-miR-15a-5p) and two PSmallMB (C9 and PC ae C38:1) features were found to be most discriminating for all disease combinations. Overall, the MOmics-based classifiers were able to provide better classification performance in comparison to mono-omics classifiers.

Interpretation: We have developed a ML pipeline to distinguish different EHT subtypes from PHT using multi-omics data. This innovative approach to stratification is an advancement towards the development of a diagnostic tool for EHT patients, significantly increasing testing throughput and accelerating administration of appropriate treatment.

Funding: European Union's Horizon 2020 Research and Innovation Programme under Grant Agreement No. 633983, Clinical Research Priority Program of the University of Zurich for the CRPP HYRENE (to Z.E. and F.B.), and Deutsche Forschungsgemeinschaft (CRC/Transregio 205/1).

Keywords: Biomarkers; Cushing syndrome; Hypertension; Machine learning; Multi-omics; Pheochromocytoma/paraganglioma; Primary aldosteronism.

Figure 1

Schematic showing the details of the 3 stages ML-based pipeline.

Figure 2

(a) Classification metrics of top-performing classifiers on the test set of 5 disease combinations trained using multi-omics and 5 mono-omics. (b) The prediction performance of top-performing classifiers (on test set) for ALL-ALL and EHT-PHT (top row), PA-PHT and PPGL-PHT (middle row) and CS-PHT (bottom row) combinations. Each symbol represents a test sample. A diamond and crosshair symbol represent a correct and incorrect prediction respectively. The y-axis represents the decision value (probability) of a trained classifier. The value of 0.5 and 0.25 was considered as an outcome of a random classifier for binary (e.g. PA-PHT) and multi-class (e.g. ALL-ALL) data. (c) ROC curves for EHT-PHT, PA-PHT, PPGL-PHT and CS-PHT (left to right) showing the top-performing classifiers and their respective AUC values. The grey line represents the performance of a random classifier.

Figure 3

Count and percentage contribution of (a) different omics in the whole multi-omics dataset. (b) Count and percentage contribution of selected features for multi-omics classification within each of 5 disease combinations. Common features amongst disease combinations shown as (c) Venn diagram and (d) Circular heatmap showing the top features selected for the classification of the 5 disease combinations using multi-omics and 5 individual omics.