Table 2.
Subgroup analyses examining the influence of text representation, annotation source, model architecture, and text source on model performance in text-based depression estimation.
| Moderators | Models, n (%) | Point estimation (95% CI) | Q-value (df) | P value | |
| Text representation | 15 (100) | —a | 16.472 (1) | <.001 | |
|
|
Embedding-based | 5 (33.3) | 0.741 (0.648-0.812) | —a | <.001 |
|
|
Traditional features | 10 (66.7) | 0.514 (0.385-0.623) | —* | <.001 |
| Annotation source | 15 (100) | —a | 4.996 (1) | .03 | |
|
|
Clinician diagnosis | 8 (53.3) | 0.688 (0.554-0.787) | —a | <.001 |
|
|
Self-report scale | 7 (46.7) | 0.500 (0.340-0.631) | —a | <.001 |
| Model architecture | 15 (100) | —a | 22.595 (1) | <.001 | |
|
|
Deep | 6 (40) | 0.731 (0.660-0.789) | —a | <.001 |
|
|
Shallow | 9 (60) | 0.486 (0.352-0.599) | —a | <.001 |
| Text source | 15 (100) | —a | 3.003 (1) | .08 | |
|
|
Documentation | 8 (53.3) | 0.529 (0.381-0.650) | —a | <.001 |
|
|
Transcribed speech | 7 (46.7) | 0.687 (0.521-0.803) | —a | <.001 |
aNot applicable.