REVIEW
Evaluation of ChatGPT for NLP-based Mental Health Applications
Computation and Language · Submitted 28 Mar 2023 · 16 references
Pith's verdict: A thin but honest zero-shot baseline report; the numbers are checkable, but the depression negative class has no stated provenance and that is the one thing that could undercut the headline 0.86.
T0 · AI review
Why Pith says this
Highlight any text to comment on it.
Referee report
The paper
The paper claims
An off-the-shelf chat model, given only a one-line prompt, classifies social media posts for stress and depression well above a majority-class guess, and does noticeably worse on a five-way suicidality task.
The reading
What carries the argument
The mechanism is minimal by design: a fixed natural-language prompt of the form 'Which of the classes can the following post be assigned to? Only reply with answers from one of the classes: CLASS NAMES', issued through the ChatComplete API with the gpt-3.5-turbo backend, with the first response taken as the prediction and compared against the dataset's existing annotations. The evaluation apparatus — weighted F1 for the multi-class case, balanced accuracy, and confusion matrices against a majority-class baseline — is what carries the argument, since the baseline is the only thing standing between the numbers and a claim of trivial class prior exploitation.
Extended reading notes
Core claim
The central claim is that zero-shot classification by ChatGPT substantially exceeds a dominant-class baseline on all three mental health detection tasks, and that this gap is evidence of a usable latent ability to infer mental state from social media text. Concretely, the paper reports weighted F1 of 0.73 for two-class stress detection (baseline 0.35), 0.86 for two-class depression detection (baseline 0.60), and 0.37 for five-class suicidality detection (baseline 0.19), with balanced accuracies of 0.73, 0.85 and 0.33. The confusion matrices show where the ability thins out: on the five-way suicidality task most inputs collapse into the Suicidal Ideation category, and the author notes that the same collapse appears in a supervised CNN model trained on the same data, which he reads as evidence that the class boundaries themselves overlap rather than as a pure failure of the model. The author also places the numbers against in-domain supervised results — a fine-tuned BERT reaching 0.81 on stress, and a feature-engineered neural model reaching 0.93 on depression — and argues that since his numbers are zero-shot, closing that gap by fine-tuning is a plausible next step rather than a hope.
Load-bearing premise
The claim rests on the assumption that each dataset's positive and negative classes differ in the mental health construct being measured rather than in topic, community, or writing style, which matters most for the depression task where the negative posts are described only as coming from a non-depression category.
Editorial extensions
If this is right
- If the reported gaps hold, a general chat model can serve as a zero-effort first-pass filter for stress and depression text before any labelled data or trained model is available.
- The suicidality result sets a realistic expectation: five-way severity schemes are where prompt-only classification breaks down, so applications needing that granularity still require supervised training or expert annotation.
- Because the model was evaluated zero-shot, the gap to fine-tuned BERT (0.81 on stress) and to feature-engineered models (0.93 on depression) becomes a concrete target for adaptation rather than a ceiling.
- The agreement between ChatGPT's confusion pattern and that of a supervised CNN on the same suicidality data suggests the five-class label scheme itself may have overlapping boundaries, independent of which model is applied.
- Re-annotating the stress and depression datasets with multiple experts would test whether model errors track genuine annotator disagreement rather than model failure.
Reading between the lines
- The headline depression number is the one most exposed to dataset construction: if the 548 negative posts come from unrelated subreddits, a strong F1 can be earned by detecting community topic and writing style rather than depressive content, so a same-community negative control would be the decisive check.
- A prompt-sensitivity sweep is cheap and would sharpen the claim considerably: if F1 moves by several points across paraphrases of the same class list, the reported numbers describe one prompt rather than the model's ability.
- Repeated sampling per post would convert the single-response point estimate into a distribution, and the spread itself may be informative — high-variance posts are plausibly the ones near the class boundary.
- Because the model was never shown examples, its errors on the suicidality task are a partial signal about where human category boundaries are thin, which is useful annotation feedback independent of any deployment plan.
Original abstract
Large language models (LLM) have been successful in several natural language understanding tasks and could be relevant for natural language processing (NLP)-based mental health application research. In this work, we report the performance of LLM-based ChatGPT (with gpt-3.5-turbo backend) in three text-based mental health classification tasks: stress detection (2-class classification), depression detection (2-class classification), and suicidality detection (5-class classification). We obtained annotated social media posts for the three classification tasks from public datasets. Then ChatGPT API classified the social media posts with an input prompt for classification. We obtained F1 scores of 0.73, 0.86, and 0.37 for stress detection, depression detection, and suicidality detection, respectively. A baseline model that always predicted the dominant class resulted in F1 scores of 0.35, 0.60, and 0.19. The zero-shot classification accuracy obtained with ChatGPT indicates a potential use of language models for mental health classification tasks.
The compiled record
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
The abstract, sentence by sentence
Large language models (LLM) have been successful in several natural language understanding tasks and could be relevant for natural language processing (NLP)-based mental health application research. In this work, we report the performance of LLM-based ChatGPT (with gpt-3.5-turbo backend) in three text-based mental health classification tasks: stress detection (2-class classification), depression detection (2-class classification), and suicidality detection (5-class classification). We obtained annotated social media posts for the three classification tasks from public datasets. Then ChatGPT API classified the social media posts with an input prompt for classification. We obtained F1 scores of 0.73, 0.86, and 0.37 for stress detection, depression detection, and suicidality detection, respectively. A baseline model that always predicted the dominant class resulted in F1 scores of 0.35, 0.60, and 0.19. The zero-shot classification accuracy obtained with ChatGPT indicates a potential use of language models for mental health classification tasks.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | Large language models (LLM) have been successful in several natural language understanding tasks and could be relevant for natural language processing (NLP)-based mental health application research. | background 1.00 | 0.05 | 0.05 | hedged possibility 0.83 |
| s1 | In this work, we report the performance of LLM-based ChatGPT (with gpt-3.5-turbo backend) in three text-based mental health classification tasks: stress detection (2-class classification), depression detection (2-class classification), and suicidality detection (5-class classification). | contribution 0.93 | 0.93 | 0.05 | asserted 0.98 |
| s2 | We obtained annotated social media posts for the three classification tasks from public datasets. | method 1.00 | 0.92 | 0.02 | asserted 1.00 |
| s3 | Then ChatGPT API classified the social media posts with an input prompt for classification. | method 1.00 | 0.89 | 0.02 | asserted 0.99 |
| s4 | We obtained F1 scores of 0.73, 0.86, and 0.37 for stress detection, depression detection, and suicidality detection, respectively. | result 1.00 | 0.96 | 0.13 | asserted 1.00 |
| s5 | A baseline model that always predicted the dominant class resulted in F1 scores of 0.35, 0.60, and 0.19. | result 0.99 | 0.84 | 0.03 | asserted 1.00 |
| s6 | The zero-shot classification accuracy obtained with ChatGPT indicates a potential use of language models for mental health classification tasks. | interpretation 1.00 | 0.83 | 0.86 | hedged possibility 1.00 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
Figures
Figures reproduced from arXiv:2303.15727 by the authors.
References
How the literature uses it
35 arXiv papers cite this in 55 sentences
Papers citing it · 40 sentences without a confident label
Uses its method or tool 1 paper
-
We used the python OpenAI library to invoke the Chat Complete API of ChatGPT by using gpt-3.5-turbo backend as in [ref] .
Underneath the Numbers: Quantitative and Qualitative Gender Fairness in LLMs for Depression Prediction · Large Language Models · 1.00
Uses its data or values 1 paper
-
Lamichhane's study evaluate the performance of LLM-based ChatGPT in various mental health classification tasks, reporting F1 scores of 0.73 for stress detection, 0.86 for depression detection, and 0.37 for suicidality detection [ref] .
Towards a Psychological Generalist AI: A Survey of Current Applications of Large Language Models and Future Prospects · Task-specific evaluations in psychology · 0.99
Cites it as support 1 paper
-
The manual prompt-based approach has shown typical few-shot prompting capabilities in the medical text translation, text augmentation, generation, summarization, and other tasks, with significant improvements over baseline models on public datasets [ref] .
Prompt Engineering for Healthcare: Methodologies and Applications · Few-shot Prompting · 0.98
Cites it as background 9 papers
-
Closer to our study, Lamichhane [ref] and Amin [ref] tested the performance of ChatGPT (GPT-3.5) on multiple classification tasks (stress, depression, and suicide detection).
Mental-LLM: Leveraging Large Language Models for Mental Health Prediction via Online Text Data · LLM and Health Applications · 0.83
-
Lamichhane et al. [ref] have evaluated the use of ChatGPT (gpt-3.5) to accomplish three mental health-related classification tasks, namely stress detection, depression detection, and suicidal detection.
Underneath the Numbers: Quantitative and Qualitative Gender Fairness in LLMs for Depression Prediction · LLMs for Mental Health Applications · 0.87
-
Recent studies have applied large language models (LLMs) to mental health assessment, marking a shift from task-specific classifiers toward general-purpose models capable of handling multiple mental health tasks within a unified framework [ref] .
Mental-R1: Aligning LLM Reasoning for Mental Health Assessment · Mental Health Assessment · 0.90
Discussion
Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.