Figure 1.

The clinical machine learning development workflow (orange boxes) offers several opportunities (blue boxes) to evaluate and mitigate potential biases introduced by the data set or model. Preprocessing methods seek to adjust the existing data set to preempt biases resulting from inadequate data representation or labeling. In-processing methods impose fairness constraints as additional metrics optimized by the model during training or present data in a structured manner to avoid biases in the sampling process. Postprocessing methods account for model biases by adjusting model outputs or changing the way they are used.