PaperlayerView exact revision on arXiv

Source-linked AI summary

Automated Clinical Coding: What, Why, and Where We Are?

Hang Dong, Matúš Falis, William Whiteley, Beatrice Alex, Joshua Matterson, Shaoxiong Ji, Jiaoyan Chen, Honghan Wu

arXiv:2203.11092v3cs.CLcs.AI

TL;DR

Clinical coding is difficult because it requires interpreting complex records and mapping them to codes through a standard process. This paper reviews automated coding research and its challenges, finding that current deep-learning methods remain limited and that knowledge-based approaches may need to complement them.

Abstract

from arXiv · show

Clinical coding is the task of transforming medical information in a patient's health records into structured codes so that they can be used for statistical analysis. This is a cognitive and time-consuming task that follows a standard process in order to achieve a high level of consistency. Clinical coding could potentially be supported by an automated system to improve the efficiency and accuracy of the process. We introduce the idea of automated clinical coding and summarise its challenges from the perspective of Artificial Intelligence (AI) and Natural Language Processing (NLP), based on the literature, our project experience over the past two and half years (late 2019 - early 2022), and discussions with clinical coding experts in Scotland and the UK. Our research reveals the gaps between the current deep learning-based approach applied to clinical coding and the need for explainability and consistency in real-world practice. Knowledge-based methods that represent and reason the standard, explainable process of a task may need to be incorporated into deep learning-based methods for clinical coding. Automated clinical coding is a promising task for AI, despite the technical and organisational challenges. Coders are needed to be involved in the development process. There is much to achieve to develop and deploy an AI-based automated system to support coding in the next five years and beyond.

Main Text

Clinical coding converts clinicians’ medical records into structured codes used to provide consistent, comparable information. This paper introduces automated clinical coding and surveys its challenges and research directions.

Main Text

  • Clinical coding transforms medical records, usually free text written by clinicians, into structured codes in systems such as ICD-10.In Scotland, coding also contributes to the national SMR01 dataset.Source: Main Text · p. 2
  • Manual coding requires experts to interpret records, select codes from large classifications, and follow a standard process and regularly updated guidelines.Training an expert clinical coder in the NHS usually takes months or longer.Source: Main Text · p. 2
  • Automated clinical coding applies AI techniques such as NLP and machine learning to support the coding task.The paper surveys technical challenges based on literature, project experience, and discussions with coding practitioners and clinicians.Source: Main Text · p. 2

Why do we need automated clinical coding?

Manual coding takes substantial time and can be delayed by backlogs; coding errors and under-coding also remain concerns. Computer-assisted coding may improve the efficiency and accuracy of human coding.

Why do we need automated clinical coding?

  • NHS Scotland coders handle about 60 cases a day, yet departments of 25 to 30 coders usually code over 20,000 cases monthly and may still face backlogs lasting months or longer.A cited example reports a backlog exceeding a year.Source: Why do we need automated clinical… · p. 2
  • Coding errors may reflect incomplete patient data, subjectivity in code selection, limited expertise, or data-entry mistakes.Reported UK coding accuracy averaged around 83%, with studies ranging from 50% to 98%.Source: Why do we need automated clinical… · p. 2
  • Scottish coding accuracy is high but imperfect, and under-coding occurs for around 20% of common conditions.The 2019–2020 assessment reported 92.5% accuracy for 3-digit codes and 88.8% for 4-digit codes of main conditions.Source: Why do we need automated clinical… · p. 2
  • A qualitative literature review reports that computer-assisted coding could improve the accuracy, quality, and efficiency of manual coding.The authors see AI as having potential to support clinical coders.Source: Why do we need automated clinical… · p. 2

Why is automated coding a complex problem to solve?

Automated coding must handle difficult clinical text, complex and changing classification systems, and the practical challenges of integrating AI into healthcare workflows. These demands extend beyond matching words to codes.

Why is automated coding a complex problem to solve?

  • Clinical coding requires natural-language understanding and linking clinical text to knowledge representations such as ICD-10.Clinical records can be lengthy, variably structured, notational, incomplete, and spread across multiple document types.Source: Why is automated coding a complex… · p. 3
  • Clinical records use terse abbreviations and symbols, and relevant information may be missing or dispersed across discharge, radiology, and pathology documents.MIMIC-III discharge summaries average around 1,500 words.Source: Why is automated coding a complex… · p. 3
  • Classification systems are large, hierarchical, and regularly updated, requiring automated coding to adapt to changing structures and codes.ICD-10-CM has around 68,000 diagnosis codes; ICD-11-MMS includes around 17,000 unique codes and more than 120,000 codable terms.Source: Why is automated coding a complex… · p. 3
  • Moving to semi-automated coding raises sociotechnical questions about coder trust, system interaction, error handling, and changing roles and skills.The paper frames national healthcare transition as potentially more challenging than the technical issues themselves.Source: Why is automated coding a complex… · p. 3

How to solve automated clinical coding: symbolic or neural AI?

Automated clinical coding has used both rule-based and learned approaches, with deep learning now prominent. The paper argues that knowledge-based methods should complement deep learning to better represent the standard coding process.

How to solve automated clinical coding: symbolic or neural AI?

  • Symbolic AI represents practice through symbols and rules, whereas neural approaches learn mappings from patient information to codes from training data.The paper describes symbolic AI as having struggled to scale to complex real-world scenarios.Source: How to solve automated clinical… · p. 3
  • Earlier rule-based systems used regular expressions, logic expressions, and keywords, but extending rules across many codes and their relationships is not straightforward and can be time-consuming.Rule-based matching can achieve high precision but low recall, and has supported human coding efficiency.Source: How to solve automated clinical… · pp. 3–4
  • Deep learning has been applied to automated coding since around 2017 and is now a prominent approach, often framing coding as multi-label classification.Some studies instead use concept extraction or Named Entity Recognition and Linking (NER+L).Source: How to solve automated clinical… · p. 4
  • Deep learning avoids expert rules and handcrafted text features and can perform better overall when trained on sufficient data.The paper notes that this approach learns from data rather than explicitly modeling the standard coding process.Source: How to solve automated clinical… · p. 4
  • Knowledge-augmented deep learning has mainly used target-ontology definitions and hierarchies, leaving other ontologies, richer ontology information, and coding guidelines underused.The authors argue for combining knowledge-based and deep-learning approaches; extracting and representing local coding guidelines requires coding-expert input.Source: How to solve automated clinical… · p. 4

How do state-of-the-art deep learning models work so far?

Current deep learning systems have not reached high performance across the full MIMIC-III coding task, and face major challenges with unseen labels, coding reasoning, and long clinical documents.

How do state-of-the-art deep learning models work so far?

  • Under 60% Micro-F1 was the best reported score on the full 8,932 ICD-9 codes in MIMIC-III.MIMIC-III is an older US intensive-care dataset and is not representative of documents available in the UK or other regions.Source: How do state-of-the-art deep learning… · p. 4
  • Around 5,000 MIMIC-III codes appear fewer than 10 times in training, and over 50% never appear; vanilla models fail completely on unseen labels.The passage also identifies the imbalance and scale of multi-label classification as challenges.Source: How do state-of-the-art deep learning… · p. 4
  • Deep learning learns text–code associations without explicitly modeling coders’ reasoning across potentially contradictory sources or guideline-based decisions.Knowledge-based techniques may support this reasoning and potentially improve performance and explainability.Source: How do state-of-the-art deep learning… · p. 4
  • Transformer input limits create a needle-in-the-haystack problem; even approaches processing up to 4,096 tokens remain insufficient for clinical notes.The passage also notes that clinical notes contain redundancy, which may impede code prediction.Source: How do state-of-the-art deep learning… · p. 5

What are the potential challenges to address for automated clinical coding?

Automated clinical coding faces technical challenges in learning rare codes, adapting to changing terminologies, and handling diverse clinical records, alongside the need for explainable decisions and coder involvement. The review also identifies limitations in current models and the need to tailor systems to their intended purposes and settings.

Model performance

  • Current BERT-based approaches do not outperform CNN-based methods for clinical-coding multi-label classification, except in one study.The authors suggest BERT may be inefficient at modelling concept-level information and long documents.Source: What are the potential challenges to… · p. 5

Gold-standard coding datasets

  • Systems need large, openly available expert-labelled datasets, since MIMIC-III may be under-coded and its models may not generalise across note styles, lengths, or languages.Expert-labelled datasets are also needed for different uses of clinical codes, including decision-making, diagnosis, and epidemiology.Source: What are the potential challenges to… · p. 5

Heterogeneous, incomplete, and noisy sources

  • Coding should draw on all relevant patient documents and may include structured data or radiographs, but real-world records are often incomplete, noisy, and inconsistently formatted.Most recent studies instead use discharge summaries.Source: What are the potential challenges to… · p. 5

Explainability

  • Label-wise attention highlights mostly show associations rather than causality, and their usefulness to coders still needs evaluation.The authors suggest integrating symbolic representations of coding steps with deep learning as a more inherently explainable direction.Source: What are the potential challenges to… · p. 5

Human-in-the-loop learning

  • Coder feedback, including corrections, highlights, and rules, may require repeated system updates; active learning can select data for annotation feedback.Examples include interfaces in MedCATTrainer and SemEHR, and active learning has been evaluated in automated coding to potentially reduce human annotations.Source: What are the potential challenges to… · pp. 5–6

Few-shot and zero-shot learning

  • Few-shot and zero-shot coding remain difficult: low-frequency codes on MIMIC-III achieve below or around 40% recall@K, with K=10 or 15.The authors note that code descriptions, properties, relations, and coding rules may help bridge the gap between seen and unseen codes.Source: What are the potential challenges to… · p. 6

Adaptation and deployment contexts

  • Models must adapt to terminology changes such as ICD-10 to ICD-11, whose poly-hierarchy and postcoordination may require new learning paradigms and ontology matching.Systems also need tailoring to purposes and country contexts; proprietary solutions are difficult to compare technically because their research access and inner workings are usually unavailable.Source: What are the potential challenges to… · pp. 6–7

NER+L

  • NER+L links codes to text and may help with long documents and few- or zero-shot codes, but requires contextual understanding and still needs code summarisation under coding guidelines.It is less adopted than multi-label classification, and more benchmarking is needed; combining the approaches remains an open direction.Source: What are the potential challenges to… · pp. 6–7

Conclusion

The paper reviews automated clinical coding from AI-research and clinical-coding perspectives, summarising deep-learning challenges and future research directions. It concludes that technical and organisational challenges remain, including involving coders in development and deployment.

Conclusion

  • The paper reviews automated clinical coding from AI-research and clinical-coding perspectives, summarises recent deep-learning challenges, and positions future research directions.Source: Conclusion · p. 7
  • Deployment also faces organisational challenges, and coder involvement in model development and deployment is essential but may be difficult to secure.Coders are often occupied with their coding work; the authors call for further research support in medical informatics and computer science.Source: Conclusion · p. 7
  • The authors cannot estimate concretely how far human-centred, explainable, intelligent, and robust automated coding is from practice, but identify a clearer path and challenges to address.They look forward to further advances over the next five years and beyond.Source: Conclusion · pp. 7–8
Loading 2203.11092v3…