Table 1 Types and examples of bias in medical artificial intelligence
From: Machine learning in precision diabetes care and cardiovascular risk prediction
Bias type | Definition: “A bias arising from…” | Example |
|---|---|---|
Confirmation bias | A tendency to interpret data in a way that confirms our prior beliefs | A machine learning model confirms existing assumptions about certain broad phenotypic groups benefiting from a given therapy, potentially leading to unequal treatment and misdiagnosis |
Sampling bias | Non-random sampling which limits the generalizability of an algorithm | Enrolling patients who visit a particular clinic or location may not represent the broader diabetes population |
Algorithmic bias | The design and implementation of an algorithms that systematically discriminates against a given group | A blood pressure monitoring system that may provide consistently inaccurate readings for a given demographic group |
Aggregation bias | Drawing misleading conclusions about individuals from group data | Concluding all patients with type 2 diabetes and hypertension benefit from a given medication without considering individual variations |
Longitudinal data fallacy | Poor analysis of temporal data | Assessing quality of diabetes control and performing long-term risk prognostication using a single laboratory reading rather than long-term patterns |
Implicit bias | Unintentional embedding of underlying biases and prejudices in algorithms | A model that is trained using records from a specific racial or ethnic group may make inaccurate predictions and disproportionately misclassify individuals from other racial groups as having higher or lower risk of diabetic complications contributing to healthcare disparities |
User interaction bias | Both the user interface and the user's behavior | A diabetes management digital health app only collects voluntary input data, thus not capturing all relevant patient information |
Presentation bias | How information is displayed to users | A patient may miss important information on an app due to the information's placement at the bottom of the screen |
Emergent bias | Longitudinal changes in population, societal habits, norms, and practices over time | An outdated diabetes therapy might persist due to long-standing cultural beliefs |
Evaluation bias | The process of model evaluation | The effectiveness of a novel antihyperglycemic therapy is evaluated against a benchmark that favors a particular demographic |
Population bias | Differences in user characteristics between the training and the intended population | A diabetes management application initially tested among tech-savvy young adults may not adequately address the needs of older adults |