Machine learning integration of multimodal data identifies key features of blood pressure regulation

Abstract

Background: Association studies have identified several biomarkers for blood pressure and hypertension, but a thorough understanding of their mutual dependencies is lacking. By integrating two different high-throughput datasets, biochemical and dietary data, we aim to understand the multifactorial contributors of blood pressure (BP).

Methods: We included 4,863 participants from TwinsUK with concurrent BP, metabolomics, genomics, biochemical measures, and dietary data. We used 5-fold cross-validation with the machine learning XGBoost algorithm to identify features of importance in context of one another in TwinsUK (80% training, 20% test). The features tested in TwinsUK were then probed using the same algorithm in an independent dataset of 2,807 individuals from the Qatari Biobank (QBB).

Findings: Our model explained 39·2% [4·5%, MAE:11·32 mmHg (95%CI, +/- 0·65)] of the variance in systolic BP (SBP) in TwinsUK. Of the top 50 features, the most influential non-demographic variables were dihomo-linolenate, cis-4-decenoyl carnitine, lactate, chloride, urate, and creatinine along with dietary intakes of total, trans and saturated fat. We also highlight the incremental value of each included dimension. Furthermore, we replicated our model in the QBB [SBP variance explained = 45·2% (13·39%)] cohort and 30 of the top 50 features overlapped between cohorts.

Interpretation: We show that an integrated analysis of omics, biochemical and dietary data improves our understanding of their in-between relationships and expands the range of potential biomarkers for blood pressure. Our results point to potentially key biological pathways to be prioritised for mechanistic studies.

Funding: Chronic Disease Research Foundation, Medical Research Council, Wellcome Trust, Qatar Foundation.

Keywords: Blood pressure; Diet; Genomics; Machine learning; Metabolomics.

Figure 1

Consort diagram of data quality control, machine learning model building and model evaluation. Data included traditional risk factors (age, sex and BMI), biochemical measures, 206 known metabolites, a BP polygenic risk score, and energy intake and dietary intake for 45 nutrients (including salt intake).

Figure 2

Importance of top 50 features in SBP. Bars represent SHAP values indicating the average relative importance of each feature, coloured according to the type of data. Base layer labels indicate metabolite super pathways, where; A=Amino Acids, Ch=Carbohydrates, L=Lipids, N=Nucleotides, P=Peptides, and CoV=Cofactors and vitamins.

Figure 3

SHAP plot of top 50 features influencing our model's prediction of SBP. Features are ranked in descending order based on their influence on our model and the x-axis denotes SHAP values. Each dot represents an individual subject and coloured according to the magnitude of the feature (red depicts a higher feature value, and blue depicts a lower value). The horizontal location of a dot (x-axis) depicts whether it corresponds with a higher or lower prediction.

Figure 4

Scatter plot of SBP within our sample and predicted SBP of the XGBoost algorithm. Actual SBP of each subject within our sample plot along the y-axis and predicted SBP from our model across the x-axis (in mmHg). The colour gradient of each point denotes the density of participants within a particular region of the plot.

Figure 5

Bi plot depicting the principal component analysis of features most influential to our SBP model. Lines depict how strongly each feature influences a principal component, and the angle between each feature represents the correlations between those features.