Table 2 Classification performance measures
From: Machine learning to refine decision making within a syndromic surveillance service
Measure | Description |
|---|---|
accuracy | Proportion of correct predictions made by the classifier. |
Matthews correlation coefficient (MCC) | Calculated for each outcome separately. Varies between − 1 and 1, and is similar to a Pearson correlation. It is evaluated from all the elements of the confusion matrix. Gives a more balanced quantification of performance than accuracy as it considers how closely the predicted results follow the decisions in the test data. Other correlation measures exist, but the MCC is suited to asymmetric classes and multi-state systems [22]. |
Precision (positive predictive power) | Calculated for each outcome separately. Expresses the fraction of classifications that match the true outcome. True positives/(true positives + false positives). E.g. proportion of ‘Alerts’ produced by the classifier that were ‘Alerts’ in the risk assessment database. |
Recall (sensitivity) | Calculated for each outcome separately. Expresses the proportion of each outcome that is correctly returned by the classifier. True positives/(true positives + false negatives). E.g. proportion of ‘Alerts’ in the risk assessment database that were identified by the classifier. |