Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-2098
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-21T10:42:25.056836Z
status: published
lens: g_space
sector: science
headline: Using Artificial Intelligence in Qualitative Lupus Pregnancy Research: The Potential of Large Language Models and Prompting Strategies
dek: Objective Our objective was to evaluate the ability of different large language models (LLMs) to translate and analyze qualitative interviews describing pregnancy experiences among women with systemic lupus erythematosus (SLE). An additional objective was to compare themes identified by LLMs with those identified by human researchers. Methods We analyzed 26 semi‐structured interviews originally conducted in Portuguese by Rodrigues et al. among women with SLE exploring their perinatal care experiences. Three prom…
gain_title: Four LLMs generated themes closely aligned with human researchers from Portuguese interviews with women with lupus, and few-shot prompting gave the most comprehensive and coherent theme extraction.
problem_title: (none)
trace_subject: (none)
gain_reading: Four LLMs generated themes closely aligned with human researchers from Portuguese interviews with women with lupus, and few-shot prompting gave the most comprehensive and coherent theme extraction.
gain_evidence: LLMs generated similar themes that closely aligned with those identified by human researchers | few‐shot prompting achieved the most comprehensive and coherent theme extraction
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers gave 26 semi-structured Portuguese interviews about perinatal care among women with systemic lupus erythematosus to four language models, using three prompting styles to produce key themes. An LLM-as-a-judge framework then scored thematic coverage, unity and semantic similarity against the themes human researchers had found.

The results suggest LLMs could serve as supportive tools in qualitative health research, including work with non-English data and varied patient perspectives. The authors frame them as complements to human interpretation, and the evidence rests on one interview set judged by another model rather than a broader test.
limitation: The study judged LLM output with an LLM-as-a-judge framework on a single set of 26 interviews, and the authors position LLMs only as complements to human interpretation.
tag: Evidence-backed gain
key_points: The study used 26 semi-structured interviews originally conducted in Portuguese with women with SLE about perinatal care. | GPT-4o, Claude Sonnet 4.5, Gemini 2.5 Flash and DeepSeek R were tested with zero-shot, few-shot and chain-of-thought prompting. | LLM outputs showed high thematic unity and semantic similarity under LLM-as-a-judge evaluation. | Claude Sonnet 4.5 showed the most consistent performance across the models tested. | Zero-shot prompting, though widely used, produced lower thematic results than few-shot prompting.
rundown: The evaluation had two aims: testing how well LLMs translate and analyze qualitative interviews about pregnancy in women with lupus, and comparing the themes they found with those identified by human researchers.

Each of four models was run on the transcripts under zero-shot, few-shot and chain-of-thought prompts. Outputs were scored on thematic coverage, thematic unity and semantic similarity.

Across models and strategies the themes closely matched the human-identified ones. Few-shot prompting consistently beat zero-shot and chain-of-thought, and Claude Sonnet 4.5 was the most consistent model. The authors conclude LLM-assisted analysis may complement human interpretation and support inclusion of non-English data.
sources:
- peer_reviewed | Arthritis Care & Research | https://doi.org/10.1002/acr.80193 | 2026-10-20
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

72fc9e26b05655d8bf45a98e772b8f44f69bfa26e3dffc85cb254d5c47f726ef

previous

0000000000000000000000000000000000000000000000000000000000000000