Table 3 Relevant forms of algorithmic biases
From: Bias recognition and mitigation strategies in artificial intelligence healthcare applications
Bias | Definition | Example |
|---|---|---|
Aggregation bias | An inappropriate combination of distinct groups or populations during data pre-processing for model development leads to aggregation bias, where the model’s performance is only optimized for the majority. In other words, one single model is unlikely to suit all groups. | Machine learning clinical prediction studies may not adequately handle missing data, leading to aggregation bias. This uniquely affects model performance among subgroups less likely to have complete data resources43. |
Feature selection bias | When the set of features chosen to train a model do not adequately represent the underlying problem or are not equally relevant across all subpopulations within the dataset. | AI models developed for COVID-19 patient risk prediction, triage, and contact tracing have shown feature selection bias through inadequate representation or consideration of social determinants of health (SDOH), such as race, socio-economic status, and access to technology. The underrepresentation of these factors has led to biased outcomes, particularly those most vulnerable or marginalized95. |
