Combining lipidomics and machine learning to measure clinical lipids in dried blood spots

Abstract

Abstract: Introduction: Blood-based sample collection is a challenge, and dried blood spots (DBS) represent an attractive alternative. However, for DBSs to be an alternative to venous blood it is important that these samples are able to deliver comparable associations with clinical outcomes. To explore this we looked to see if lipid profile data could be used to predict the concentration of triglyceride, HDL, LDL and total cholesterol in DBSs using markers identified in plasma. Objectives: To determine if DBSs can be used as an alternative to venous blood in both research and clinical settings, and to determine if machine learning could predict ‘clinical lipid’ concentration from lipid profile data. Methods: Lipid profiles were generated from plasma (n = 777) and DBS (n = 835) samples. Random forest was applied to identify and validate panels of lipid markers in plasma, which were translated into the DBS cohort to provide robust measures of the four ‘clinical lipids’. Results: In plasma samples panels of lipid markers were identified that could predict the concentration of the ‘clinical lipids’ with correlations between estimated and measured triglyceride, HDL, LDL and total cholesterol of 0.920, 0.743, 0.580 and 0.424 respectively. When translated into DBS samples, correlations of 0.836, 0.591, 0.561 and 0.569 were achieved for triglyceride, HDL, LDL and total cholesterol. Conclusion: DBSs represent an alternative to venous blood, however further work is required to improve the combined lipidomics and machine learning approach to develop it for use in health monitoring.