Reasoning Modes Boost Language Model Accuracy in Chinese Clinical Text Extraction

Designated default configurations of large language models achieved 99.7 percent overall accuracy and complete output validity in a retrospective benchmark evaluating structured extraction from Chinese clinical text. The investigation assessed model performance and efficiency trade-offs across differences in parameter scale, reasoning mode, sampling temperature, architecture, and input modality.

The single-center retrospective study examined 100 Chinese clinical reports across each of three predefined clinical tasks: left ventricular ejection fraction extraction, carotid intima-media thickness extraction, and coronary computed tomography angiography classification. Across all three tasks, the same 300 reports were processed under 23 distinct experimental conditions, generating a total dataset of 6,900 model outputs.

Evaluated outcomes encompassed output validity alongside all-report accuracy, measured through exact matches for numerical ejection fraction and intima-media thickness values and through classification accuracy for coronary angiography reports. Among valid angiography responses, investigators calculated class-specific precision, recall, and F1 scores, as well as macro-F1 and the Matthews correlation coefficient. Efficiency measures recorded during local inference included token generation counts, time to first token, and total end-to-end latency.

Across four designated default configurations, coronary computed tomography angiography classification reached 99.0 percent accuracy, yielding macro-F1 values between 0.984 and 0.985 and Matthews correlation coefficients between 0.969 and 0.970 among valid responses. Despite high overall classification scores, distinct distributions of false-positive and false-negative errors occurred across the tested configurations.

Activating reasoning modes produced marked gains in analytical performance alongside substantial increases in computational runtime. In the compact Qwen3-0.6B model, reasoning improved overall accuracy from 31.0 percent to 88.7 percent while elevating the coronary angiography Matthews correlation coefficient from 0.025 to 0.599. For the larger Qwen3-32B model, reasoning increased overall accuracy from 98.7 percent to 99.7 percent and lifted the Matthews correlation coefficient from 0.890 to 0.970, while median end-to-end latency climbed from 3.746 seconds to 32.097 seconds.

Varying sampling temperature provided no uniform improvement across coronary angiography evaluation metrics. In architectural evaluations, dense models and mixture-of-experts configurations demonstrated closely aligned classification metrics while displaying divergent token outputs and latency profiles. Similarly, comparisons between text-only and multimodal configurations revealed no consistent relative advantage across two evaluated temperatures.

Statistical evaluation indicated that the primary association between parameter scale and performance in dense-only models was not significant after correction for multiplicity. In addition, report-level performance estimates exhibited sensitivity to the underlying model composition.

The benchmark demonstrated that multiple configurations achieved high extraction accuracy, classification performance, and output validity on the predefined clinical tasks. Because class-specific metrics revealed error patterns that overall accuracy failed to detect, the researchers concluded that model selection must weigh overall performance and output validity against class-specific error tolerances and local inference efficiency requirements.

Subject of Research: Performance and efficiency trade-offs of large language models in Chinese clinical text extraction

Article Title: 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

Article References: Hu, B., Wang, F., Lan, C., Liu, P., Yu, Y., Zhou, L., Li, N., Wu, P., & Yang, H. (2026). 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. Frontiers in Medicine, 13, Article 1942113. https://doi.org/10.3389/fmed.2026.1942113

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1942113

Keywords: large language models, Chinese clinical text, clinical natural language processing, structured information extraction, left ventricular ejection fraction, carotid intima-media thickness, coronary computed tomography angiography, reasoning mode, model scale, inference efficiency, end-to-end latency, time to first token, output validity, Matthews correlation coefficient, macro-F1, mixture of experts, multimodal models