An improved prediction method for diabetes based on a feature-based least angle regression algorithm

Existing diabetes prediction algorithms have a number of shortcomings, most notably low accuracy and poor generalizability. In this paper, we propose a method based on feature weights to improve diabetes prediction that combines the advantages of traditional least angle regression (LARS) algorithms and principal component analysis (PCA) algorithms.First of all, a principal component analysis algorithm is used to obtain the characteristic independent variables found in typical diabetes prediction regression models. Each of these variables is assigned its own characteristics. After this, the original variable correlation is multiplied by the weight of the variable obtained using principal component analysis to obtain a new degree of correlation. This new correlation is used to optimize the forward direction and variable selection of a least angle regression solution before calculating the regression coefficients for the new model. An experiment using the Pima Indians Diabetes dataset provided by the University of California was performed to validate the proposed algorithm. The experimental results show that the algorithm improved the approximation speed for the dependent variables and the accuracy of the regression coefficients. It was also able to select the key characteristic variables for diabetes prediction whilst simplifying the standard diabetes prediction model. Thus, it may help with the provision of more accurate diabetes prevention and treatment measures in the future.