A Unified Review of Deep Learning for Automated Medical Coding

TL;DR

The paper addresses automated medical coding by proposing a unified encoder-decoder framework that decomposes models into encoders, deep-architecture building blocks, decoders, and auxiliary information. It surveys diverse encoder types (RNNs, CNNs, transformers, graphs, hierarchical) and decoder designs (FC, attention, hierarchical, multitask, few-shot, autoregressive), along with the use of external data such as code descriptions, code hierarchies, ontologies, and chart data. It also covers benchmarking data (notably MIMIC) and evaluation metrics, discusses real-world practice across countries, and outlines challenges (data quality, interpretability, privacy) with concrete future directions including LLMs, multimodal integration, and robust handling of long documents. The work aims to facilitate cross-method comparisons, provide a practical roadmap for researchers and practitioners, and advance clinically usable, knowledge-aware, and privacy-conscious medical coding systems. Overall, it highlights that a modular, knowledge-enhanced, and human-in-the-loop approach is key to scalable, interpretable, and robust automated medical coding in real-world healthcare settings.

Abstract

Automated medical coding, an essential task for healthcare operation and delivery, makes unstructured data manageable by predicting medical codes from clinical documents. Recent advances in deep learning and natural language processing have been widely applied to this task. However, deep learning-based medical coding lacks a unified view of the design of neural network architectures. This review proposes a unified framework to provide a general understanding of the building blocks of medical coding models and summarizes recent advanced models under the proposed framework. Our unified framework decomposes medical coding into four main components, i.e., encoder modules for text feature extraction, mechanisms for building deep encoder architectures, decoder modules for transforming hidden representations into medical codes, and the usage of auxiliary information. Finally, we introduce the benchmarks and real-world usage and discuss key research challenges and future directions.

Figures (13)

• Figure 1: A medical coding model maps an example clinical note to the corresponding ICD procedure and diagnosis codes.
• Figure 2: The distribution of ICD codes in the MIMIC-III dataset curated by mullenbach2018explainable.
• Figure 3: An illustration of the unified encoder-decoder framework for automated medical coding
• Figure 4: Illustrations of representative neural text encoders. (a) The RNN encoder captures sequential dependency. (b) The CNN encoder extracts local features. (c) The BERT encoder encodes contextualized information.
• Figure 5: Graph encoders on the constructed medical graph. a) GMAN applies the GCN layers to encode disease hierarchy and disease-finding causal relations. b) CMGE uses the GAT as the graph encoder on a multi-granularity graph.