Large Language Models Revolutionize Medical ICD Auto-Coding: From PLM-ICD to Next-Generation Intelligent Coding Systems

This study focuses on using state-of-the-art medical large language models (LLMs) to improve the accuracy, interpretability, and effectiveness of ICD auto-coding in unstructured clinical records, and conducts a multi-dimensional comparative analysis with the existing baseline method PLM-ICD. The study will evaluate model performance from three core dimensions: accuracy (micro-F1, macro-F1, AUPRC), interpretability (attention mechanism, generative explanation), and practical application effects (inference speed, resource consumption, etc.), aiming to provide new technical directions for the automation of medical ICD coding.

In the modern healthcare system, ICD coding is a key link connecting clinical diagnosis and treatment with medical management. However, traditional manual coding is costly and inefficient, making it difficult to handle massive electronic medical record data. Pre-trained language models (PLMs) such as PLM-ICD have brought hope for automated coding, but with the development of large language models, whether they can achieve a qualitative leap has become a research focus. PLM-ICD uses models like BERT to extract features and predict ICD codes, while the powerful capabilities of LLMs are expected to solve the limitations of existing methods.

The technical architecture of PLM-ICD includes:

1. Text Encoding Layer: Uses BERT or medical domain variants (e.g., ClinicalBERT, BioBERT) as encoders to learn semantic representations of medical terms;
2. Label-Aware Attention Mechanism: For multi-label classification tasks, learns specific attention vectors for each ICD code to extract relevant information;
3. Hierarchical Code Structure Utilization: Leverages the hierarchical structure of ICD (e.g., A00→A00.0) to ensure reasonable code combinations through hierarchical classification.

Medical LLMs (e.g., Med-PaLM, Meditron) have three major advantages over PLM-ICD:

1. Extended Context Understanding: Supports longer token inputs (e.g., 4096+), enabling complete processing of long clinical records and capturing cross-paragraph associations;
2. Rich Medical Knowledge Reserve: Pre-training covers massive medical literature and guidelines, enabling understanding of deep knowledge such as disease associations and diagnostic criteria;
3. Generative Reasoning Capability: Can generate coding explanations, confidence notes, and even interactive clarification of issues, improving user experience.

The experiment uses the MIMIC-III/IV dataset (de-identified intensive care unit records and ICD code annotations). The evaluation protocol includes:

• Time-sensitive data partitioning (training/validation/test separated in chronological order);
• Performance reporting on the test set after hyperparameter tuning on the validation set;
• Significance tests to verify performance improvements;
• Error analysis to identify failure modes. The comparison models cover medical LLMs of different scales (7B-70B parameters) and training strategies (pre-training, instruction fine-tuning, etc.).

Expected outcomes include:

1. Technical Contribution: Establish performance benchmarks for medical LLMs in ICD coding tasks, revealing advantages and limitations;
2. Practical Guide: Assist medical institutions in evaluating and selecting coding solutions, covering model selection, deployment costs, etc.;
3. Open-Source Contribution: Publicize code, models, and experimental records to promote community collaboration and reproduction. These outcomes will drive the progress of medical AI coding technology, benefiting medical institutions and patients.

Applying LLMs to ICD coding faces challenges:

1. Computational Resource Requirements: Inference costs are higher than PLMs; need to explore model compression, knowledge distillation, etc., to reduce overhead;
2. Coding Consistency Assurance: Need to combine rule engines to ensure codes comply with ICD rules (e.g., code pairing/mutual exclusion);
3. Continuous Learning and Adaptation: Need to respond to medical knowledge updates and ICD version revisions (e.g., ICD-9→10→11) to achieve rapid system adaptation. Future research will focus on these directions.