crich011/roberta-postoperative-sentiment-twitter

Postoperative Patient Sentiment

Model Details

• Base model: cardiffnlp/twitter-roberta-base-sentiment-latest
• Task: 3-class sentiment classification
• Labels: negative, mixed/neutral, positive, assigned by two physicians training as orthopedic surgeons

Intended Use

Sentiment classification of postoperative patient comments for research and quality-improvement workflows.

Training

• Fine-tuning set: 400 comments total (200 spine, 200 arthroplasty)
• Hyperparameters: 4 epochs, learning rate 2e-5, batch size 8 (train) / 16 (eval), max length 256

Evaluation (Holdout)

Cohort	n	Macro F1	Weighted F1	Accuracy
Spine	205	0.684667	0.756538	0.756098
Arthroplasty	191	0.666545	0.737915	0.769634

Limitations

• Developed in a specific institutional context; external generalization is not guaranteed.
• Fine-tuning labels are based on physician rater judgement and may reflect rater variability.
• Not intended as a standalone clinical decision tool.