Summarizing Online Patient Conversations Using Generative Language Models: Experimental and Comparative Study
Bibliographic record
Abstract
BACKGROUND: Social media is acknowledged by regulatory bodies (eg, the Food and Drug Administration) as an important source of patient experience data to learn about patients' unmet needs, priorities, and preferences. However, current methods rely either on manual analysis and do not scale, or on automatic processing, yielding mainly quantitative insights. Methods that can automatically summarize texts and yield qualitative insights at scale are missing. OBJECTIVE: The objective of this study was to evaluate to what extent state-of-the-art large language models can appropriately summarize posts shared by patients in web-based forums and health communities. Specifically, the goal was to compare the performance of different language models and prompting strategies on the task of summarizing documents reflecting the experiences of individual patients. METHODS: In our experimental and comparative study, we applied 3 different language models (Flan-T5, Generative Pretrained Transformer [GPT], GPT-3, and GPT-3.5) in combination with various prompting strategies to the task of summarizing posts from patients in online communities. The generated summaries were evaluated with respect to 124 manually created summaries as a ground-truth reference. As evaluation metrics, we used 2 standard metrics from the field of text generation, namely, Recall-Oriented Understudy for Gisting Evaluation (ROUGE) and BERTScore, to compare the automatically generated summaries to the manually created reference summaries. RESULTS: Among the zero-shot prompting-based large language models investigated, GPT-3.5 performed better than the other models with respect to the ROUGE metrics, as well as with respect to BERTScore. While zero-shot prompting seems to be a good prompting strategy, overall GPT-3.5 in combination with directional stimulus prompting in a 3-shot setting had the best results with respect to the aforementioned metrics. A manual investigation of the summarization of the best-performing method showed that the generated summaries were accurate and plausible compared to the manual summaries. CONCLUSIONS: Taken together, our results suggest that state-of-the-art pretrained language models are a valuable tool to provide qualitative insights about the patient experience to better understand unmet needs, patient priorities, and how a disease impacts daily functioning and quality of life to inform processes aimed at improving health care delivery and ensure that drug development focuses more on the actual priorities and unmet needs of patients. The key limitations of our work are the small data sample as well as the fact that the manual summaries were created by 1 annotator only. Furthermore, the results hold only for the examined models and prompting strategies, potentially not generalizing to other models and strategies.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.009 | 0.040 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.003 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".