Multimodal Generative Information Retrieval of Chest X-Rays Grounded on ICD-9 Taxonomy

The International Classification of Diseases (ICD) is a World Health Organisation standardised system for coding diagnoses and medical procedures. Automatic ICD coding assigns codes to clinical documentation but is a complex task due to hierarchical taxonomies, multiple comorbidities, and semantic gaps between clinical language and standardised codes. Existing solutions employ either retrieval-based approaches, which match clinical text to code databases using attention mechanisms and external knowledge, or generative methods that directly produce ICD codes as textual outputs through large language models. Current approaches face critical limitations: error traceability and domain guardrailing. Existing systems provide direct code assignments without exposing hierarchical reasoning paths, while generative models frequently produce codes from unintended taxonomic versions despite explicit constraints. We address these limitations by introducing the first hybrid retrieval-generative model for ICD coding that approaches the task as a hierarchical generative process. Our method utilises the ICD taxonomic tree structure to guide predictions, ensuring domain adherence while providing explicit hierarchical reasoning paths. We train our hybrid model using a custom dataset built from MIMIC-CXR JPG and MIMIC-IV, where inputs are radiology reports and chest X-ray images, and outputs are hierarchical ICD code sequences. Performance is evaluated using Recall@K and Mean Reciprocal Rank metrics. Our hybrid approach demonstrates strong performance enhancing explainability through hierarchical retrieval, with cluster matching analysis confirming the model's ability to maintain coherent relationships within the ICD taxonomy. These results challenge conventional retrieval and generative models in specialised medical domains and establish our approach's effectiveness for automatic ICD coding applications.