Performance–efficiency trade-offs of large language models for structured extraction from Chinese clinical free text: a benchmark study of model scale, reasoning mode, and architecture

Background Large language models (LLMs) may support structured extraction from clinical free text, but evaluation should consider accuracy, output validity, and inference efficiency. Objective To evaluate LLM performance–efficiency trade-offs in Chinese clinical text extraction across model scale, reasoning mode, temperature, architecture, and text-only versus multimodal configurations. Methods This single-center retrospective benchmark sampled 100 Chinese clinical reports per task: left ventricular ejection fraction (LVEF) extraction, carotid intima-media thickness (IMT) extraction, and coronary computed tomography angiography (CTA) classification. The same 300 reports were evaluated under 23 conditions, yielding 6900 outputs. Outcomes included output validity and all-report accuracy (exact-match for LVEF and IMT; classification for CTA). Class-specific precision, recall, and F1, macro-F1, and the Matthews correlation coefficient (MCC) were calculated among valid CTA responses. Local inference efficiency measures included tokens, time to first token, and end-to-end latency. Results Four designated default configurations achieved 99.7% overall accuracy and 100.0% validity. CTA accuracy was 99.0%, with valid-response macro-F1 of 0.984–0.985 and MCC of 0.969–0.970, although false-positive and false-negative errors differed. Reasoning increased Qwen3-0.6B overall accuracy from 31.0% to 88.7% and CTA MCC from 0.025 to 0.599. For Qwen3-32B, overall accuracy increased from 98.7% to 99.7%, CTA MCC from 0.890 to 0.970, and median end-to-end latency from 3.746 to 32.097 s. Temperature changes offered no consistent improvement across CTA metrics. Dense and mixture-of-experts configurations had closely spaced CTA metrics but different latency and token profiles. Text-only and multimodal Qwen configurations showed no consistent relative advantage across two temperatures. The primary dense-only parameter-scale association was not significant after multiplicity correction, and report-level estimates were sensitive to model composition. Conclusions Several evaluated configurations achieved high LVEF and IMT exact-match accuracy, CTA classification performance, and output validity on three predefined tasks. Class-specific metrics identified error differences that accuracy alone did not capture. Selection should consider all-report performance, output validity, class-specific errors, and inference requirements under intended deployment conditions.