OFSET_mine:an integrated framework for cardiovascular diseases risk prediction based on retinal vascular function

Abstract

As cardiovascular disease (CVD) represents a spectrum of disorders that often manifestfor the first time through an acute life-threatening event, early identification of seemingly healthy subjects with various degrees of risk is a priority.More recently, traditional scores used for early identification of CVD risk are slowly being replaced by more sensitive biomarkers that assess individual, rather than population risks for CVD. Among these, retinal vascular function, as assessed by the retinal vessel analysis method (RVA), has been proven as an accurate reflection of subclinical CVD in groups of participants without overt disease but with certain inherited or acquired risk factors. Furthermore, in order to correctly detect individual risk at an early stage, specialized machine learning methods and featureselection techniques that can cope with the characteristics of the data need to bedevised.The main contribution of this thesis is an integrated framework, OFSET_mine, that combinesnovel machine learning methods to produce a bespoke solution for Cardiovascular Risk Prediction based on RVA data that is also applicable to other medical datasets with similar characteristics. The three identified essential characteristics are 1) imbalanced dataset,2) high dimensionality and 3) overlapping feature ranges with the possibility of acquiring new samples. The thesis proposes FiltADASYN as an oversampling method that deals with imbalance, DD_Rank as a feature selection method that handles high dimensionality, and GCO_mine as a method for individual-based classification, all three integrated within the OFSET_mine framework.The new oversampling method FiltADASYN extends Adaptive Synthetic Oversampling(ADASYN) with an additional step to filter the generated samples and improve the reliability of the resultant sample set. The feature selection method DD_Rank is based on Restricted Boltzmann Machine (RBM) and ranks features according to their stability and discrimination power. GCO_mine is a lazy learning method based on Graph Cut Optimization (GCO), which considers both the local arrangements and the global structure of the data.OFSET_mine compares favourably to well established composite techniques. Itex hibits high classification performance when applied to a wide range of benchmark medical datasets with variable sample size, dimensionality and imbalance ratios.When applying OFSET _mine on our RVA data, an accuracy of 99.52% is achieved. In addition, using OFSET, the hybrid solution of FiltADASYN and DD_Rank, with Random Forest on our RVA data produces risk group classifications with accuracy 99.68%. This not only reflects the success of the framework but also establishes RVAas a valuable cardiovascular risk predictor