Work

Automated Medical Coding Using Supervised Deep Learning and Large Language Models

Year: 2026
Type: dissertation
Abstract: Automated medical coding assigns International Classification of Diseases (ICD) codes to free-format patient notes. It remains challenging due to limited labeled data, lengthy clinical narratives, and the imbalanced distribution of frequent and rare codes. This thesis investigates automated ICD code prediction using both supervised deep learning and Large Language Model (LLM)-based approaches on MIMIC-IV. In the first part, supervised encoder-decoder architectures are evaluated to study transformer-based encoding and domain-specific pretraining. A CNN baseline is compared with transformer-based encoders, including BERT, BioBERT, Bio_Clinical-BERT, and RoBERTa-PM, across ICD-9 and ICD-10 under Full-Code and Top-50 settings. Transformer-based models consistently outperform the CNN baseline, while domain-specific pretraining further improves performance, with RoBERTa-PM achieving the best results. The second part investigates the capabilities and limitations of LLMs. We develop a prompting pipeline without fine-tuning that decomposes coding into two subtasks: disease/procedure identification and ICD mapping. We compare zero-shot, few-shot, zero-shot Chain-of-Thought (CoT), few-shot CoT, and Retrieval-Augmented Generation (RAG) strategies on MIMIC-IV discharge summaries over the full ICD-10 code set. While LLMs underperform on common codes, they outperform supervised models on rare codes, achieving an 8.1% relative Macro-F1 improvement over the PLM-ICD baseline (0.239 vs. 0.221). The largest gain stems from combining zero-shot CoT for identification with few-shot hierarchical guidance for mapping, while integrating RAG yields the best results (Micro-F1 = 0.4004, Macro-F1 = 0.2389). Overall, transformer-based supervised models remain most effective for ICD prediction, whereas structured prompting and retrieval augmentation substantially improve rare-code prediction where labeled data and fine-tuning resources are limited.
Language: English
Open Access status: green
OpenAlex ID: W7221492162
Cites: 35
Cited by:
Related to: 10