Impact of Class Weights and Feature Importance in Automated Stroke Detection

In this research paper, we do a parametric study of class weighting as a way to tackle imbalance during training. We then infer the most important features that should be taken into consideration for stroke prediction.

A stroke occurs when a blood vessel that carries oxygen and nutrients to the brain is either blocked by a clot or ruptures. Consequently, part of the brain is unable to obtain the blood (and oxygen) it needs, so brain cells die. This makes it important to be able to assess the probability of a stroke given features that are specific to patients so that they could take preventative measures in the future. Thus, the unpredictability and deadliness of strokes pose the following questions: Can we predict the occurrence of a stroke from few facts about the patient easily accessible by the doctor? What are the most important features for stroke prediction? In this project, we investigate the feasibility of using a supervised machine learning model to predict stroke occurrence. In practice, however, we faced challenges such as low prevalence and the imbalance in the available dataset, with many more negative than positive cases. In this research paper, we do a parametric study of class weighting as a way to tackle imbalance during training. We then infer the most important features that should be taken into consideration for stroke prediction. Assessing feature importance allows for patients to focus on two or three areas that may be contributing to their high probability for getting a stroke. The most significant result was that the most important feature that should be considered when determining the probability of an individual getting a stroke is age. However, there is no distinct second most important feature. Additionally, a non monotonic improvement with a class weight of 22.5 for positive cases in this dataset produces the most optimal results.

Explore More!

Research Paper