Table 3 Results of our named entity evaluation using Bio_ClinicalBERT

From: Deep learning models can predict violence and threats against healthcare providers using clinical notes

  

Scoring strategy

  

Strict

Relaxed

Category

# In test set

P

R

F1

P

R

F1

Acute substance use

141

0.67

0.62

0.64

0.73

0.66

0.69

Aggressive behavior

263

0.63

0.71

0.67

0.67

0.77

0.72

Cognition

627

0.69

0.73

0.71

0.72

0.78

0.75

High utilization or incarceration

92

0.69

0.66

0.67

0.72

0.69

0.7

Mood symptoms

139

0.62

0.66

0.64

0.66

0.71

0.68

Noncompliance

228

0.64

0.65

0.64

0.69

0.7

0.69

Psychosis

66

0.63

0.73

0.68

0.67

0.77

0.72

Unmet needs or conflict

175

0.43

0.45

0.44

0.47

0.5

0.48

Total (micro avg.)

1463

0.64

0.67

0.65

0.68

0.72

0.7

  1. Strict scoring requires exact matching of start and end character indices and category, while relaxed scoring requires category matching and half or more of gold and predicted tokens to overlap.