Table 1.

Group fairness metrics encountered in this review.

Term	Description
AUROCa	Assesses overall classifier performance by measuring the TPRb and FPRc of a classifier at different thresholds.
Average odds	Compares the average of the TPR and FPR for the classification outcome between protected and unprotected groups.
Balanced accuracy	A measure of accuracy corrected for data imbalance, calculated as the average of sensitivity and specificity for a group.
Calibration	Assesses how well the risk score or probability predictions reflect actual outcomes.
Disparate impact	Measures deviation from statistical parity, calculated as the ratio of the rate of the positive outcome between protected and unprotected groups. Ideally, the disparate impact is 1.
Equal opportunity	For classification tasks in which one outcome is preferred over the other, equal opportunity is satisfied when the preferred outcome is predicted with equal accuracy between protected and unprotected groups. Ideally, the TPR or FNRd disparity between groups is 0.
Equalized odds	The TPR and FPR are equal between protected and unprotected groups.
Error rate	Compares the error rate of predictions, calculated as the number of incorrect predictions divided by the total number of predictions, between protected and unprotected groups. Ideally, the error rate disparity between groups is 0.
Statistical parity	Statistical parity (also known as demographic parity) is satisfied when the rate of positive outcomes is equal between protected and unprotected groups.

^aAUROC: area under the receiver operating characteristic curve.

^bTPR: true-positive rate.

^cFPR: false-positive rate.

^dFNR: false-negative rate.