Depression Subtype Classification Benchmark on 14,983 Tweets

14,983 English-language tweets are stratified into six clinically grounded categories: five depression subtypes and a no-depression class. The dataset was curated to benchmark few-shot prompting against fine-tuned encoders for tweet-level classification, with fine-tuned RoBERTa-large achieving an accuracy of 0.954.

Use Cases

• Benchmark few-shot LLMs against fine-tuned encoders for classifying tweets into the six depression subtype categories.
• Analyze per-class precision and recall for nuanced subtypes like postpartum, psychotic, and atypical depression.
• Examine confusion matrices to identify common misclassification patterns, such as major depression being labeled as bipolar.
• Use the stratified class labels to train a transformer encoder for high-recall detection of major depression.

Strengths

• 14,983 tweets provide a substantial sample for model benchmarking.
• Six clinically grounded categories enable fine-grained depression subtyping analysis.
• Benchmark results include macro-F1 scores ranging from 0.73–0.77 for few-shot models to 0.94–0.96 for fine-tuned models.
• Dataset is stratified, supporting balanced evaluation across classes.

Limitations

• Data is limited to English-language tweets from X (formerly Twitter), restricting geographic and linguistic generalizability.
• The 18.7 KB file size indicates the primary data is a description document, not the raw tweet dataset, which may require separate sourcing.
• Labeling bias and the noisy, short nature of social media text present inherent classification challenges.

Provenance

Source

Rawan AlSaad via figshare.

Collection Method

Curated dataset of tweets stratified into six clinically grounded categories.

Freshness

Last updated March 2026.