바로가기 메뉴 본문 바로가기 주메뉴 바로가기
  • 06-2bHave you mitigated the impact of features that may create bias?
    • There can be bias due to features such as discriminatory elements included in medical data. To prevent this, features like the composition of datasets are typically reviewed for any possibility of bias.

    • When using overseas medical datasets, conduct a review on whether dataset features are excessively biased towards a certain gender or focused on White people and people from developed countries. This is due to the fact that the results learned from the medical data may vary based on gender, ethnicity, and standard of living. For instance, if the training result from training data oriented to White people is used at a medical institution comprised of mostly Asian patients, the algorithm’s performance will significantly deteriorate. Therefore, pre-process data to mitigate the influence of the dataset’s structural features that can cause bias, such as race, on the model, or select training data attributes for race that is expected to occupy the main user of the model and system being developed, or use data composed of patient groups with diverse backgrounds whenever possible.

    • Simple approaches to mitigate biases include feature selection which excludes features that may lead to bias. The three methods for feature selection are filter, wrapper, and embedded methods, and these are methods used to analyze the statistical correlation of features in data to use features with high correlation coefficients or use subsets with great performance on certain features.

    • The methods, however, may not always be effective, as they can reinforce or create different biases when removing features associated with biases. An example is a study that suggests preparing diverse methods like reweighting can be better, because a model’s performance or accuracy deteriorates when racial variables are removed to mitigate racial bias in medical data [27]. Therefore, you must consider various techniques (e.g. reweighting, relabeling, variable blinding, sampling) to mitigate bias.

    • But there can be exceptions if you can mitigate bias during the training or if bias was intentional based on the AI system’s purpose.