Adam optimized Logistic Regression to enhance heart disease diagnosis

Abstract

Disease prediction by Machine Learning (ML) models has been experimented a lot by researchers onto several heart disease datasets. Framingham heart disease data is relatable by medical practitioners as well as the features are interpretable by a layman. Logistic Regression is a binary classifier that is interpretable and a simplistic model. It utilizes sigmoid activation function and computes log loss for the cost function. To optimize the cost function, gradient descent optimiser is used which has static learning rate and converges a convex cost function. The work in this paper is in continuation with multiple probabilistic classifiers where Logistic Regression classifier with optimal threshold through Precision Recall curve (Model 1) demonstrated improved weighted F score, accuracy, precision, recall and fscore for imbalanced, missing valued and multicollinear data as Framingham. The Learning curve for Logistic Regression was neither biased nor had variance. Further, this experimentation failed to converge. Hence, in order to enhance performance further, adaptive moment optimiser has been used for Logistic Regression (Model 2) which accelerates training by smoothing learning and uses an adaptive learning rate per parameter. This stochastic optimisation of a non convex cost function has enhanced weighted Fscore further. The novel algorithm has also been tested on benchmarked Statlog, Cleveland and Cardio Vascular Disease (CVD) datasets with enhancement in performance as well. The results have been statistically evaluated by McNemar’s non parametric test and 5X2 Cross Validation (CV) paired t-test to validate that the novel model is different from Python’s implementation of Logistic Regression classifier.