Concept-Aware Learning Framework for Detecting Common Symptoms in Social Media Text: Model Development and Evaluation Study.

BACKGROUND

Social media offers valuable information for public health surveillance, but detecting valid symptom expressions in noisy and informal text remains challenging. Fine-tuned language models can achieve strong test accuracy while relying on spurious patterns, limiting their reliability in real-world applications.

OBJECTIVE

This study aimed to develop a more robust symptom detection model by incorporating surveillance-relevant attributes into the training objective and introducing a counterfactual-based evaluation framework to assess attribute sensitivity.

METHODS

We proposed a concept-aware learning framework (CALF), which combines multiattribute counterfactual data augmentation with multitask learning. We compared standard fine-tuning (baseline) with data-level interventions, a model-level intervention, hybrid strategies, and prompt-based methods across multiple encoders and LLMs for data augmentation. Performance was evaluated on the following 3 datasets: the MedWeb test using macro F1-score (m-F1), the MedWeb checklist using flip accuracy (FA), and a supplementary test set using F1-score for flu-or-fever detection (f-F1). FA measures sensitivity to surveillance-relevant attributes, and the supplementary test score serves as an indicator for out-of-distribution performance.

RESULTS

The baseline achieved an m-F1-score of 0.871 on the test set but a low overall FA on the checklist (FA=0.342). It also has limited sensitivity to recovery status, timing of occurrence, and uncertainty of expression. Compared with the baseline, CALF has a comparable MedWeb test score (m-F1=0.880 vs 0.871; P=.18), but a substantially higher FA (FA=0.830 vs 0.342; P=.001) and supplementary test score (f-F1=0.856 vs 0.797; P=.007). CALF also improved sensitivity to recovery, timing, and uncertainty across pretrained encoders. The gains were generally larger when GPT-4.1 (OpenAI) was used for data augmentation. Under GPT-4.1, CALF often achieved the highest overall FA among encoder-based methods. When Qwen2.5-7B-Instruct (Alibaba Cloud) was used, an alternative hybrid strategy may achieve comparable or slightly better results. Checklist performance was strongly associated with supplementary test performance (ρ=0.811), whereas MedWeb test performance showed weak association with supplementary test performance (ρ=-0.042). Prompting GPT-4.1 achieved the highest FA overall (FA=0.920) and the highest supplementary test score (f-F1=0.910), but a lower test set score than the baseline (m-F1=0.776 vs 0.871).

CONCLUSIONS

Standard held-out metrics alone do not adequately characterize robustness to dataset-specific or spuriously correlated features. Models may perform well on the test set while remaining insensitive to attributes that influence whether a symptom mention is relevant to surveillance. The MedWeb checklist complements conventional evaluation by measuring sensitivity to controlled attribute changes and may indicate out-of-distribution performance. CALF improved sensitivity to surveillance-relevant attributes while maintaining strong classification performance, supporting concept-aware learning as a practical strategy for developing more robust encoder-based models.