Detecting caregivers of dementia family members on social media networks is becoming increasingly important for healthcare support. This paper presents our system for detecting dementia family caregivers on Twitter/X as part of the Social Media Mining for Health (#SMM4H) Shared Task 3. Using the Gemma 2-2B decoder-only transformer model with 2 billion parameters, we implemented 8-bit quantization and Low-Rank Adaptation (LoRA) to efficiently fine-tune on a dataset of 6,724 tweets. Our approach achieved an F1 score of 0.966 on the test set, marginally outperforming the previous state-of-the-art BERTweet-Large baseline and achieving the highest score in this task. The results suggest that larger, newer models as well as different model architectures can improve performance in this task, potentially opening avenues for better identification of caregivers, who may benefit from support resources on social media platforms.

PDF