Healthcare Text Classification

Prompts

This paper uses three task-specific prompt templates (Table 2). and are replaced at runtime with the task label and input text respectively.

Social Media Tasks

Used for breast cancer, medication regimen changes, adverse pregnancy outcomes, and potential COVID-19 cases:

Read the following tweet. Answer if the user is self-reporting  and explain why.
Tweet:
Your answer should state "the tweet is self-reporting " or "the tweet is not self-reporting potential".

Stigma Labeling (Clinical Note)

Read the following clinical note.
Clinical note:
Answer if this note involves language about the patient which could result in the stigmatization or negative labeling of a patient, which could lead to further status loss/discrimination in the context of the patient-provider relationship. Your answer should state "the note involves ..." or "the note does not involve ...".

Medication Change Discussion (Clinical Note)

Read the following clinical note.
Clinical note:
Answer if the presence of a medication change is discussed. Your answer should state "the presence of a medication change is discussed" or "the presence of a medication change is not discussed".

Usage Notes

This prompt is from the paper “Benchmarking Open-Source Large Language Models on Healthcare Text Classification Tasks” (Guo & Sarker, 2025).

• Models benchmarked: GEMMA-3-27B-IT, LLaMA3-70B, LLaMA4-109B, DeepSeek-R1-Distill-LLaMA-70B, and DeepSeek-V3-0324-UD-Q2_K_XL.
• Tasks: Six binary classification tasks — four involving social media data (breast cancer, medication regimen changes, adverse pregnancy outcomes, potential COVID-19 cases) and two involving clinical notes (stigma labeling, medication change discussion).
• Setting: All evaluation experiments conducted in zero-shot settings.
• Key finding: DeepSeek-V3 emerged as the strongest overall performer, achieving the highest F1 scores in four out of six tasks.