Background Clinical document classification is a critical process in healthcare, converting unstructured medical texts into standardized ICD-10 diagnoses. This process faces challenges due to the complex and varied nature of medical language, which includes domain specific terminology, abbreviations, and unique writing styles across institutions. Additionally, privacy regulations and limited high quality annotated datasets hinder the development of robust models. LLMs have emerged as a transformative technology in healthcare, improving the efficiency and accuracy of tasks like clinical document classification by leveraging advanced natural language understanding. Objective The objective of this study is to evaluate the performance and consistency of LLMs in binary classification clinical discharge summaries based on ICD-10 codes. By leveraging both reasoning and non-reasoning LLMs, the study aims to determine how effectively these models can identify and classify clinical patterns in a binary context, providing insights into their potential for improving automated clinical coding accuracy and enhancing decision support in healthcare settings. Methods This study used a balanced subset of the MIMIC-IV dataset, comprising 3,000 discharge summaries including 150 positive and 150 negative samples for each of the top 10 ICD-10 codes. The summaries were tokenized using cTAKES, which converted clinical narratives into structured SNOMED codes, capturing contextual details such as affirmation or negation. Eight LLMs, including four reasoning (Qwen QWQ, Deepseek Reasoner, GPT o3 Mini, Gemini 2.0 Flash Thinking) and four non-reasoning models (Llama 3.3, GPT 4o Mini, Gemini 2.0 Flash, Deepseek Chat), were evaluated over three experimental runs. Final predictions were determined using majority voting across the runs to assess accuracy, F1 score, and consistency. Results Among the eight evaluated LLMs, reasoning models demonstrated superior performance in ICD-10 classification, achieving an average accuracy of 71% and an F1 score of 67%, compared to 68% accuracy and 60% F1 score for non-reasoning models. Gemini 2.0 Flash Thinking achieved the highest accuracy at 75% and F1 score at 76%, while GPT 4o Mini had the lowest performance 64% accuracy, and 47% F1 score. Consistency analysis revealed that non-reasoning models exhibited higher stability of 91% average consistency than reasoning models of 84%. Performance variations across ICD-10 codes highlighted strengths in identifying well defined conditions but challenges in classifying abstract diagnostic categories. Conclusion The evaluation of reasoning and non-reasoning LLMs in ICD-10 classification highlights a trade-off between accuracy and consistency. Reasoning models achieved higher classification accuracy and F1 scores, excelling in complex clinical cases, while non-reasoning models demonstrated superior stability across repeated trials. These findings suggest that a hybrid approach, leveraging the strengths of both model types, could optimize automated clinical coding by balancing accuracy and reliability. Future research should explore multi-label classification, domain specific fine tuning, and ensemble modeling to enhance performance and generalizability in real-world healthcare applications.

Can reasoning LLMs enhance clinical document classification?

期刊论文
出版模式：

开放获取
发表日期：

2026年1月7日

查看全文

Akram Mustafa,
Usman Naseem,
Mostafa Rahimi Azghadi

摘要

Background

Clinical document classification is a critical process in healthcare, converting unstructured medical texts into standardized ICD-10 diagnoses. This process faces challenges due to the complex and varied nature of medical language, which includes domain specific terminology, abbreviations, and unique writing styles across institutions. Additionally, privacy regulations and limited high quality annotated datasets hinder the development of robust models. LLMs have emerged as a transformative technology in healthcare, improving the efficiency and accuracy of tasks like clinical document classification by leveraging advanced natural language understanding.

Objective

The objective of this study is to evaluate the performance and consistency of LLMs in binary classification clinical discharge summaries based on ICD-10 codes. By leveraging both reasoning and non-reasoning LLMs, the study aims to determine how effectively these models can identify and classify clinical patterns in a binary context, providing insights into their potential for improving automated clinical coding accuracy and enhancing decision support in healthcare settings.

Methods

This study used a balanced subset of the MIMIC-IV dataset, comprising 3,000 discharge summaries including 150 positive and 150 negative samples for each of the top 10 ICD-10 codes. The summaries were tokenized using cTAKES, which converted clinical narratives into structured SNOMED codes, capturing contextual details such as affirmation or negation. Eight LLMs, including four reasoning (Qwen QWQ, Deepseek Reasoner, GPT o3 Mini, Gemini 2.0 Flash Thinking) and four non-reasoning models (Llama 3.3, GPT 4o Mini, Gemini 2.0 Flash, Deepseek Chat), were evaluated over three experimental runs. Final predictions were determined using majority voting across the runs to assess accuracy, F1 score, and consistency.

Results

Among the eight evaluated LLMs, reasoning models demonstrated superior performance in ICD-10 classification, achieving an average accuracy of 71% and an F1 score of 67%, compared to 68% accuracy and 60% F1 score for non-reasoning models. Gemini 2.0 Flash Thinking achieved the highest accuracy at 75% and F1 score at 76%, while GPT 4o Mini had the lowest performance 64% accuracy, and 47% F1 score. Consistency analysis revealed that non-reasoning models exhibited higher stability of 91% average consistency than reasoning models of 84%. Performance variations across ICD-10 codes highlighted strengths in identifying well defined conditions but challenges in classifying abstract diagnostic categories.

Conclusion

The evaluation of reasoning and non-reasoning LLMs in ICD-10 classification highlights a trade-off between accuracy and consistency. Reasoning models achieved higher classification accuracy and F1 scores, excelling in complex clinical cases, while non-reasoning models demonstrated superior stability across repeated trials. These findings suggest that a hybrid approach, leveraging the strengths of both model types, could optimize automated clinical coding by balancing accuracy and reliability. Future research should explore multi-label classification, domain specific fine tuning, and ensemble modeling to enhance performance and generalizability in real-world healthcare applications.