Machine learning analysis of metabolomic biomarkers for diagnosis of heart failure

Abstract

Background: There is emerging evidence that the pathophysiological mechanisms of heart failure are associated with alterations in serum metabolites. Such metabolomic signatures may be useful for heart failure diagnosis, stratification and prognosis. Purpose: To evaluate the utility of including metabolomic biomarkers in addition to traditional cardiac biomarkers in a machine learning prediction model of heart failure diagnosis in the well-characterised Canagliflozin Cardiovascular Assessment Study (CANVAS) cohort. Methods: A subgroup of the CANVAS/CANVAS-R study cohort was analysed. 101 metabolites in plasma were measured by HPLC (HILIC)-mass spectrometry. A 10-times 5-fold cross-validated support vector machine model with radial basis kernel function was constructed to predict heart failure diagnosis using traditional biomarkers alone and using the combination of traditional biomarkers and metabolomic biomarkers. Model performance and variable importance were both evaluated by area under the curve (AUC) of the receiver operating characteristics (ROC) curve. Results are shown as mean ± standard deviation. Results: 967 patients (of which 402 patients had heart failure) were included in the analysis with 341 females, mean age 63±8 years and mean body mass index (BMI) 33±5 kg/m2. All patients had diabetes mellitus with mean HbA1c 8.2±0.9%. The prediction model based on only traditional biomarkers had mean AUC 72±3% and the prediction model based on both traditional biomarkers and metabolomic biomarkers had mean AUC 80±3%. The top metabolomic biomarkers for predicting heart failure were threonine, L-homoserine, creatine and deoxyadenosine. Conclusion: Metabolomic biomarkers improved diagnostic performance of a heart failure prediction model and captured variation not encompassed by traditional cardiac biomarkers.