A default classification threshold of 0.5 may not provide the best balance between precision and recall. The F1 score, their harmonic mean, offers a criterion for choosing a threshold instead of guessing. section 4.6.5
Selecting a threshold
For a medical-text model, the best reported F1 score was 0.9559 at a threshold of 0.5999. Five-fold cross-validation produced an average accuracy of 0.9543, with similar results across folds. Even so, the high F1 score warrants a sanity check for potential data leakage; near-perfect scores on a small dataset can also indicate overfitting. section 4.6.5
Figure 4.11 Precision vs. Recall at Different Thresholds.

Compare precision and recall across thresholds before selecting the threshold with the best F1 score. section 4.6.5



















