Table 2.
Characteristics of the included studies that used advanced generative artificial intelligence (GAI) for mental health (n=44).
| Characteristics | Studies | References | |||
| Number of authors, mean (SD; range) | 3.75 (1.66; 1-8) | —a | |||
| Authors, n (%) | |||||
|
|
1 | 11 (25) | [44-46,74-81] | ||
|
|
1-5 | 16 (36) | [43,47-50,82-92] | ||
|
|
≥5 | 17 (39) | [51-53,72,73,93-104] | ||
| Study phase, n (%) | |||||
|
|
Development and validation of GAI techniques | 37 (84) | [43,44,46,47,49,51-53,72,73,75,77-83,85-101,103,104] | ||
|
|
Efficacy of GAI models | 7 (16) | [45,48,50,74,76,84,102] | ||
| Mental health outcome, n (%) | |||||
|
|
General mental health | 28 (64) | [43,46,47,51-53,72,77,79-85,87-95,99,101,102,104] | ||
|
|
Depression | 9 (20) | [45,48,49,73,75,76,78,98,100] | ||
|
|
Anxiety | 1 (2) | [104] | ||
|
|
Anxiety and depression | 5 (11) | [74,86,96,97,103] | ||
|
|
Bipolar disorder | 2 (5) | [74,97] | ||
|
|
Eating disorder | 1 (2) | [84] | ||
|
|
Schizophrenia | 1 (2) | [84] | ||
|
|
PTSDb | 2 (5) | [84,101] | ||
| Data sources , n (%) | |||||
|
|
Private data | 2 (5) | [49,82] | ||
|
|
Social media data | 9 (20) | [43,48,72,85,87,94,98,101,103] | ||
|
|
Publicly available data set | 19 (43) | [44,46,47,50,51,53,77,79-81,83,86,88,89,96,97,99,100,102] | ||
|
|
Other sources | 14 (32) | [45,52,73-76,78,84,90-93,95,104] | ||
| Type of techniquesc, n (%) | |||||
|
|
GPT-2d | 8 (18) | [53,72,77,79,81,88,89,98] | ||
|
|
GPT-3 or GPT3.5 | 13 (30) | [45,48,49,73,74,84,87,90,93,95,99,101,102] | ||
|
|
GPT-4 | 1 (2) | [76] | ||
|
|
DialoGPT | 1 (2) | [47] | ||
|
|
GANse | 6 (14) | [44,51,75,78,83,100] | ||
|
|
PanGu 350 M/WenZhong-110M (a type of LLMf in Chinese) | 1 (2) | [94] | ||
|
|
HyperCLOVA (a type of LLM) | 1 (2) | [50] | ||
|
|
Markov chain | 1 (2) | [92] | ||
|
|
LSTMg | 11 (25) | [44,46,53,75,85,86,90,96,100,103,104] | ||
|
|
GRUh | 1 (2) | [91] | ||
|
|
Midjourneyi | 1 (2) | [52] | ||
| Delivery mode , n (%) | |||||
|
|
Text | 32 (73) | [43,45-49,72-74,76-78,80-82,84,85,87-90,93-103] | ||
|
|
Audio | 9 (20) | [43,50,53,75,79,86,91,92,104] | ||
|
|
Image | 4 (9) | [44,51,52,83] | ||
| Groups receiving benefit , n (%) | |||||
|
|
General population | 29 (66) | [47,48,50-53,74,75,77-86,88-93,95,97-99,104] | ||
|
|
Clinicians and mental health professionals | 3 (7) | [45,76,94] | ||
|
|
Students and youth | 2 (5) | [43,46] | ||
|
|
Children | 1 (2) | [44] | ||
|
|
Peer supporters | 1 (2) | [72] | ||
|
|
Suicide gatekeepers | 1 (2) | [73] | ||
|
|
Not reported | 7 (16) | [49,87,96,100-103] | ||
aNot applicable.
bPTSD: posttraumatic stress disorder.
cIndicates the models or components used in generative tasks.
dGPT: generative pretrained transformer.
eGAN: generative adversarial network.
fLLM: large language model.
gLSTM: long short-term memory.
hGRU: gated recurrent unit.
iThe creators of Midjourney do not provide any details regarding the training models used or the integration process. Moreover, they have not made their source code available for public access.