Table 3 Performance on clinically-relevant utterances by patients.

From: Assessing the accuracy of automatic speech recognition for psychotherapy

PHQ

Keywordsa

Number of positives

True positives

False negatives

False positives

Sensitivity

Positive predictive value

1

Interest, interested, interesting, interests, pleasure

169

127

42

38

75%

77%

2

Depressed, depressing, feeling down, hopeless, miserable

74

63

11

12

85%

84%

3

Asleep, drowsy, sleepiness, sleeping, sleepy

114

85

29

19

75%

82%

4

Energy, tired

143

115

28

22

80%

84%

5

Overeat, overeating

5

3

2

0

60%

100%

6

Bad, badly, poorly

405

336

69

56

83%

86%

7

Mindfulness

11

9

2

0

82%

100%

8

Fidget, fidgety, restless, slow, slowing, slowly

39

28

11

13

72%

68%

9

Dead, death, depression, died, suicide

103

86

17

18

83%

83%

 

Weighted average

1063

852

211

178

80%

83%

  1. aFor each question of the Patient Health Questionnaire (PHQ-9), relevant keywords were identified by querying the Unified Medical Language System using each PHQ question to generate search terms. Each table row denotes a different question from the PHQ-9. Number of occurrences refer to how often the keywords appear in our transcribed therapy sessions. True positives refer to a correct transcription by the automatic speech recognition system. False negatives and false positives denote incorrect transcriptions. Sample size is denoted by the number of positives.