Abstract Aims Dynamic retinal vessel analysis (DVA) provides a non-invasive way to assess microvascular function in patients and potentially to improve predictions of individual cardiovascular (CV) risk. The aim of our study was to use untargeted machine learning on DVA in order to improve CV mortality prediction and identify corresponding response alterations. Methods and results We adopted a workflow consisting of noise reduction and extraction of independent components within DVA signals. Predictor performance was assessed in survival random forest models. Applying our technique to the prediction of all-cause mortality in a cohort of 214 haemodialysis patients resulted in the selection of a component which was highly correlated to maximal venous dilation following flicker stimulation (vMax), a previously identified predictor, confirming the validity of our approach. When fitting for CV mortality as the outcome of interest, a combination of three components derived from the arterial signal resulted in a marked improvement in predictive performance. Clustering analysis suggested that these independent components identified groups of patients with substantially higher CV mortality. Conclusion Our results provide a machine learning workflow to improve the predictive performance of DVA and identify groups of haemodialysis patients at high risk of CV mortality. Our approach may also prove to be promising for DVA signal analysis in other CV disease states.