Table 2.

Performance of several ML classifiers on the task of detecting instances of self-harm in the ED triage texts corpus

	PR AUC	Precision	Recall	F1 score
Naive Bayes	0.666 (±0.03)	0.550 (±0.03)	0.707 (±0.03)	0.618 (±0.02)
Logistic regression	0.799 (±0.03)	0.296 (±0.01)	0.952 (±0.01)	0.452 (±0.01)
k-nearest neighbor	0.463 (±0.05)	0.892 (±0.04)	0.276 (±0.04)	0.421 (±0.05)
Random forest	0.799 (±0.03)	0.843 (±0.03)	0.601 (±0.05)	0.702 (±0.04)
Gradient boosting	0.832 (±0.03)	0.855 (±0.04)	0.691 (±0.05)	0.764 (±0.03)
LSTM*	0.801 (±0.04)	0.560 (±0.08)	0.874 (±0.02)	0.682 (±0.05)

Note: Results are reported as mean (±95% confidence intervals) obtained from 10-fold cross-validation on the training set. For LSTM, we used 3-fold cross-validation.