Exploring GPT-based models for depression detection in adolescents using naturalistic speech: A proof-of-concept study

This study investigates whether GPT-4o, a Large Language Model (LLM), can identify depression from daily, naturalistic non-symptom-specific speech transcripts; how its performance varies with input volume; and how it compares to other LLMs. Thirty Brazilian adolescents (aged 16–19) completed a 14-day remote data collection using IDEABot, a WhatsApp-based chatbot. Half of the participants ( n = 15, 8 girls) had a confirmed Major Depressive Disorder (MDD) diagnosis established by a child and adolescent psychiatrist using gold-standard procedures. GPT-4o was prompted to classify participants as depressed or non-depressed based on daily transcribed audio messages in Brazilian Portuguese. Performance was evaluated across varying input lengths (number of days), and a secondary analysis compared GPT-4o with 12 alternative GPT-based models. GPT-4o achieved an overall accuracy of 86.6% (95% CI 70.3%, 95.6%), with 100% sensitivity and 73.3% specificity (95% CI 55.6%, 86.2%). Classification remained stable across days, and increasing input length yielded only modest improvements. No statistically significant differences were observed between models, although GPT-4o showed numerically higher performance than newer and larger models. As a proof-of-concept study using a purposively selected sample representing opposite ends of the symptom spectrum, these findings are preliminary. Nonetheless, they suggest that LLMs may support depression detection from naturalistic conversational text, extending prior work based on structured assessments to more ecologically valid contexts. This study highlights both the promise and current limitations of AI-based approaches for mental health screening. • GPT-4o identified depression using naturalistic daily speech transcripts. • High accuracy and sensitivity achieved in Brazilian adolescents. • Classification performance was stable across varying amounts of speech input. • GPT-4o performs similarly to newer and larger GPT-based models. • LLMs may support scalable, low-burden depression screening.