Background: Up to 13% of adolescents suffer from depressive disorders. Despite the high psychological burden, adolescents rarely decide to contact child and adolescent psychiatric services. To provide a low-barrier alternative, our long-term goal is to develop a chatbot for early identification of depressive symptoms. To test feasibility, we followed a two-step procedure, a) collection and linguistic analysis of psychiatric interviews with healthy adolescents and adolescents with depressive disorders and training of classifiers for detection of disorders from their answers in interviews, and b) generation of additional adolescent utterances via Chat GPT to improve the previously created model.

Methods: For step a), we collected standardized interviews with 53 adolescents, n = 40 with and n = 13 without depressive disorders. The transcribed interviews comprised 4,077 question-answer-pairs, with which we predicted the clinical rating (depressive/non-depressive) with use of a feedforward neural network that received BERT (Bidirectional Encoder Representations from Transformers) vectors of interviewer questions and patient answers as input. For step b), we used the answers of all 53 interviews to instruct Chat GPT to generate new similar utterances.

Results: In step a), the classifier based on BERT was able to discriminate answers by adolescents with and without depression with accuracies up to 97% and identified commonly used words and phrases. Evaluating the quality of utterances generated in step b), we found that prompt engineering for this task is difficult as Chat GPT performs poorly with long prompts and abstract descriptions of expectations on appropriate responses. The best approach was to cite original answers from the transcripts in order to optimally mimic the style of language used by patients and to find a practicable compromise between the length of prompts that Chat GPT can handle and the number of examples presented in order to minimize literal repetitions in Chat GPT’s output.

Conclusion: The results indicate that identifying linguistic patterns in adolescents’ transcribed verbal responses is promising and that Chat GPT can be leveraged to generate a large dataset of interviews. The main benefit is that without any loss of validity the synthetic data are significantly easier to obtain than interview transcripts.