Explainable Machine Learning in Cardiovascular Diagnostics

Sammanfattning: The major challenges in implementing machine learning models in medical applications stemfrom ethical and accountability concerns, which arise from the lack of insight and understandingof the models' inner workings and reasoning. This opaqueness has resulted in the emergenceof a new subfield of machine learning called Explainability, which aims to develop and deploymethods to gain insight into how input data is weighted and propagated through a machinelearning algorithm.This paper aims to examine the viability of certain explainability methods when applied tocardiovascular diagnostics. The machine learning models that were implemented andsubsequently evaluated include Logistic Regression, Decision Trees, and Random Forests.Methods such as Feature Importance plots, Lasso Regularization (L1 norm), and SequentialFeature Selection were applied to achieve better interpretation of these models.The results indicate that different models and forms of regularization prioritize various inputfeatures more heavily than others, even when trained on identical data. A consistent findingacross all models, except for Logistic Regression with Lasso regularization, was the ability tosignificantly reduce the dimensionality of the input feature space without substantial loss inmodel test performance. This allows for the isolation of specific features, thereby enhancinginsight into and improving a model's interpretability.Systolic and diastolic blood pressure along with cholesterol values were the two main inputfeatures that determined a patients cardiovascular diagnosis.