Language Models for ICD Coding: Benchmarking Generative and Retrieval Approaches on Multimodal Clinical Data

The International Classification of Diseases (ICD) is a standard system for coding clinical diagnoses and procedures. Automated ICD Coding assigns such codes to unstructured medical reports, representing a complex task due to the high ontological dimensionality and the semantic gap between clinical narratives and standardized codes. Existing solutions employ discriminative and retrieval approaches based on label attention mechanisms and ontological graphs, or generative methods that directly output codes as sequences via Large Language Models. These approaches present critical limitations: the lack of error traceability, with direct assignments devoid of hierarchical paths; the lack of domain constraints, whereby generative models produce hallucinations or combine taxonomic versions between ICD-9 and ICD-10; and unimodal isolation, which ignores chest X-rays. This thesis proposes a systematic comparative study for multimodal ICD coding, integrating discharge letters and radiographic images on a large scale. We compare three architectural paradigms: (i) generative fine-tuning with multimodal decoders and constrained Prefix Tree decoding; (ii) zero-shot top-down hierarchical prompting; and (iii) contrastive Bi-Encoder Retrieval (MedSigLIP and PubMedBERT). The experimentation is conducted on the MIMIC-GEN-RAG multimodal dataset through a quantitative framework that evaluates multi-label classification, @K ranking metrics, and hierarchical correctness. Experimental results demonstrate that our Bi-Encoder Retrieval significantly outperforms existing literature baselines in diagnostic recall and efficiency. Concurrently, Prefix Tree-constrained generation eliminates nosological hallucinations, ensuring strict taxonomic adherence and high precision on primary diagnoses. Furthermore, multimodal integration enhances accuracy on visible thoracic findings, establishing the effectiveness of our framework for automated clinical coding applications.