Automated Clinical Coding for Primary Care: Revolutionizing Efficiency and Accuracy

Introduction: The Advent of Automated Clinical Coding in Primary Care

Clinical coding, the intricate process of translating medical information from patient records into standardized codes, is crucial for healthcare data analysis and management. This task, traditionally performed manually by expert coders, is inherently complex and time-intensive. In primary care settings, where patient volume is high and efficiency is paramount, the demands on clinical coding are particularly significant. Automated clinical coding systems, leveraging Artificial Intelligence (AI) and Natural Language Processing (NLP), present a transformative solution to enhance both the efficiency and accuracy of this critical process within primary care.

The primary goal of clinical coding is to ensure consistent and comparable clinical data across different healthcare providers and over time. This standardized data is essential for various purposes, including public health monitoring, healthcare planning, policy development, and epidemiological research. In primary care, accurate and efficient coding directly impacts patient care management, resource allocation, and the overall effectiveness of healthcare delivery. While the US healthcare system utilizes these codes extensively for billing, their value extends far beyond financial transactions, informing crucial aspects of healthcare operations and research globally. For an overview of clinical coding principles in the UK’s National Health Service (NHS), resources like “Clinical coding for non coders” by NHS Digital provide valuable insights.

Manual clinical coding is a demanding cognitive task. Expert coders must meticulously analyze extensive patient documentation, often in the form of free-text clinical notes, and select the most appropriate codes from comprehensive classification systems such as ICD-10 (International Classification of Diseases, Tenth Revision). Consider the complexity of ICD-10-CM, used in the US, which contains approximately 68,000 diagnosis codes. Coders must adhere to constantly evolving coding guidelines and standards to maintain consistency. Public Health Scotland, for instance, regularly updates and publishes coding guidelines. Training a proficient clinical coder can take months, highlighting the expertise and time investment required for this role, as noted in studies examining international clinical coding training programs.

Automated clinical coding offers a compelling alternative, employing AI techniques, particularly NLP and machine learning, to streamline and enhance this process. It represents a significant advancement in computer-assisted coding (CAC), aiming to harness the power of AI to handle the growing volume of healthcare data. The application of AI in healthcare, especially through machine learning and NLP, is rapidly gaining traction, promising intelligent data processing and improved clinical workflows. Automated clinical coding in primary care stands out as a high-potential AI application, capable of revolutionizing administrative and research aspects of patient record management within these busy healthcare settings. Recent years have witnessed a surge in research exploring deep learning approaches for automated clinical coding, as highlighted in several comprehensive reviews.

Despite the promising advancements in automated clinical coding, particularly with deep learning, it’s crucial to acknowledge that the task remains far from fully resolved, especially within the nuanced environment of primary care. Drawing upon ongoing research and discussions with clinical coding professionals and clinicians across the UK, this article aims to delve into the specific technical challenges of automated clinical coding in primary care, particularly concerning deep learning methodologies. Furthermore, it proposes future research directions to address these challenges and unlock the full potential of AI in this domain.

Fig. 1. Illustrative example of manual and automated clinical coding processes, showcasing potential interactions (dashed arrows) between clinical coders and automated systems, using ICD-9-CM codes from a clinical note within the MIMIC-III dataset.

Note: Data and coding formats depicted are specific to the MIMIC-III dataset and may not reflect data structures or coding systems used in other regions, such as the UK, where data may be less structured and standardized discharge summaries are not universally adopted.

The Compelling Need for Automated Clinical Coding in Primary Care

Several critical factors underscore the urgent need for automated clinical coding solutions, especially within the demanding context of primary care. Firstly, manual coding is exceptionally time-consuming. While a clinical coder in NHS Scotland might process approximately 60 cases daily, even this efficient pace translates to a considerable time investment per case. In bustling primary care practices, the sheer volume of patient encounters can quickly lead to substantial backlogs in coding, sometimes stretching to months or even years, as evidenced by reports of coding delays.

Secondly, manual coding is susceptible to errors. Incomplete patient information, subjective interpretations in code selection, variations in coding expertise, and simple data entry mistakes can all contribute to inaccuracies. Studies have indicated varying accuracy rates in manual coding, highlighting the inherent challenges in maintaining perfect consistency. Even in systems with high reported accuracy, like in Scotland, under-coding and imperfections persist. Computer-assisted coding, particularly automated systems, offers a pathway to mitigate these errors, potentially improving the accuracy, quality, and efficiency of coding processes, as suggested by literature reviews. The integration of advanced AI technologies like NLP holds immense promise in enhancing the support automated coding systems can provide to primary care clinical coders, specifically by directly contributing to accurate clinical code assignment.

Navigating the Complexities of Automated Coding in Primary Care

While the prospect of automated clinical coding in primary care is highly appealing, the task presents significant hurdles for computer-based systems. The very nature of clinical coding – text analysis, summarization, and classification into codes – mirrors the complexities of Natural Language Understanding (NLU), a long-standing challenge within AI. Effectively linking the nuances of human language to structured knowledge representations like ICD-10 is a non-trivial undertaking. Furthermore, clinical coding in primary care presents unique challenges compared to general NLU tasks:

Approaches to Automated Clinical Coding: Symbolic AI, Neural Networks, and Hybrid Models in Primary Care

The field of AI has historically been shaped by two primary approaches: symbolic, knowledge-based systems and neural network-based approaches, which have evolved into deep learning. When applied to clinical coding, these approaches manifest in distinct ways. Symbolic AI aims to replicate the rule-based processes that expert clinical coders employ, using symbols and rules to represent coding standards and guidelines. Conversely, neural network and deep learning approaches focus on learning complex patterns directly from data, aiming to map patient information to the correct medical codes through intricate function approximation.

Historically, symbolic AI, dominant from the 1950s to the early 1980s, struggled to scale to real-world complexity, particularly in handling the intricacies of natural language. Neural networks re-emerged in the mid-1980s with the rise of machine learning. Deep learning then became the dominant paradigm after 2011, and continues to evolve rapidly.

In the context of automated clinical coding, while research dates back approximately 50 years, deep learning-based methods are a relatively recent development. Prior to deep learning, most systems relied on rule-based approaches (regular expressions, logic rules, keywords) combined with feature engineering for text classification. Rule-based systems, while offering precision in specific scenarios, struggled with scalability and the sheer volume and complexity of coding rules and code variations. This led to the integration of machine learning techniques, such as Decision Trees and Support Vector Machines (SVM), to enhance rule-based systems with textual feature analysis. While rule-based methods, like regular expressions, can achieve high precision (though often lower recall) and improve coding efficiency as support tools, they are not sufficient on their own for comprehensive automated coding.

The application of deep learning to automated coding gained momentum around 2017, leading to a surge in research. Deep learning offers the advantage of not requiring explicit rule engineering or manual feature crafting, learning directly from data and achieving impressive performance in many cases. Most deep learning approaches frame clinical coding as a multi-label classification problem, while some explore concept extraction or Named Entity Recognition and Linking (NER + L) approaches. While deep learning is currently the predominant method, a compelling argument exists for incorporating knowledge-based approaches to complement deep learning, creating hybrid systems. Knowledge-augmented deep learning, an emerging trend, seeks to integrate knowledge graphs and ontologies into deep learning models, often through embedding techniques, or by directly incorporating hierarchical code structures and relationships. However, current knowledge integration often remains limited to the target ontology (e.g., ICD-9) and its hierarchical structure, neglecting the vast resources of other clinical ontologies (UMLS, SNOMED-CT, etc.) and codified coding guidelines. Leveraging these richer knowledge sources, especially coding standards and guidelines, presents a significant opportunity to enhance deep learning models, although it requires effectively extracting and representing knowledge from these sources, which often necessitates input from expert clinical coders.

Understanding State-of-the-Art Deep Learning Models in Clinical Coding

“Coding tasks involving complex reasoning, such as those in which disparate pieces of information must be connected, are a difficult challenge for current NLP systems.” – Kukafka et al., a sentiment echoed by Stanfill et al.

Clinical coding serves as a rigorous testing ground for contemporary AI, especially for machine learning and deep learning applied to NLP. The challenges of this task have spurred research in areas such as text representation learning, multi-task learning, zero-shot learning, meta-learning, and multi-modal learning. However, a fully realized deep learning-based clinical coding system remains an ongoing pursuit. Current state-of-the-art Micro-F1 scores on the full set of ICD-9 codes using the MIMIC-III dataset remain below 60%. The MIMIC-III discharge summaries, while widely used for benchmarking, are coded with the older ICD-9-CM version and represent data from intensive care units in the US collected over a decade ago, potentially limiting their representativeness for contemporary primary care settings globally.

The dominant deep learning approach involves training a complex function to map clinical notes to sets of codes, framed as a multi-label classification problem. However, this approach faces several limitations when applied to clinical coding:

Key Challenges and Future Directions for Automated Clinical Coding in Primary Care

Interestingly, current BERT-based deep learning models have not consistently outperformed CNN-based methods for multi-label clinical coding classification, with some exceptions. This may be attributed to BERT’s potential limitations in effectively capturing concept-level information, which is often represented in keywords or short phrases rather than complex token relationships within long documents, and in processing lengthy clinical documents efficiently.

Crucially, manual clinical coding relies heavily on a standardized process incorporating rules and guidelines specific to healthcare systems, such as code priority, hypothetical mentions, code definitions, and mutual exclusion rules. Future automated systems, particularly deep learning-based ones, must integrate knowledge reasoning with ontologies and rule-based approaches to achieve improved, more explainable, and clinically relevant results.

Drawing from research in clinical coding, several technical challenges warrant focused attention. Explainability and effective handling of few- and zero-shot learning scenarios are particularly relevant to multi-label classification approaches and may be mitigated by NER + L methodologies.

While multi-label classification is a common approach, Named Entity Recognition and Linking (NER + L), exemplified by tools like MedCAT and research in rare disease identification using SemEHR, offers an alternative, though less explored in recent literature. NER + L, grounded in clinical information extraction, benefits from recent deep learning advancements. NER + L offers inherent explainability by linking codes to specific text segments and is better suited to handling long documents. However, extracted codes still require summarization and adherence to coding standards. NER + L methods may also improve coding for rare or unseen codes (few-shot/zero-shot) by identifying relevant concepts within clinical notes. A potential drawback of NER + L is the need for contextual understanding (negation, temporality, experiencer), which is less critical in multi-label classification. Combining multi-label classification and NER + L approaches within a hybrid system design warrants further investigation. Initial attempts to integrate NER + L concepts through text enrichment or multi-task learning have not yet shown significant performance gains over multi-label classification alone, indicating a need for further research. Using NER + L and ontologies to augment clinical notes with synonyms or sibling code names to enhance few- and zero-shot coding is another promising direction. Ranking ICD-10 codes extracted via NER + L for billing code prediction also offers a potentially superior approach to addressing few- and zero-shot challenges compared to multi-label classification. More comprehensive benchmarking of NER + L-enhanced methods is needed for comparative evaluation.

Automated clinical coding systems need to be tailored to specific purposes (billing vs. research) and contexts (countries, healthcare settings). For billing, the focus is often on predicting Diagnosis-Related Groups (DRGs) or Healthcare Resource Groups (HRGs), which involve a smaller, grouped set of codes and can potentially be predicted prior to detailed ICD coding. For research, systems require broader classification systems with finer granularity for tasks like case detection and phenotyping, necessitating the use of terminologies like SNOMED CT, ORDO, ICD-11, and customized terminologies. NER + L systems with rule-based inference can be particularly valuable for phenotyping when labeled data is scarce. Automated coding systems can also be integrated with clinical outcome prediction models (e.g., readmission, mortality) in end-to-end deep learning frameworks. Furthermore, the evaluation metrics may vary depending on the application; case detection in research may prioritize precision (PPV) over recall (sensitivity). Country-specific factors, such as the “Note Bloat” phenomenon in the US, where redundant information is copied into notes, also impact system design. De-duplication of clinical notes has been shown to improve prediction task performance. Billing and insurance models also introduce country-specific considerations for system design.

Beyond healthcare institutions and academia, industry organizations play a crucial role. Collaborations between industry and academia are increasingly common. The integration of NLP components into EHR systems like Epic, such as the CogStack project’s collaboration with University College London Hospital (UCLH), demonstrates the growing trend of embedding automated coding tools into clinical workflows. Industry-developed NER + L APIs from companies like Amazon, Microsoft, and Google are also emerging, offering commercial solutions for clinical concept extraction. Numerous technology companies, including Deloitte, Optum, Capita, and CHKS, provide proprietary (semi-)automated clinical coding solutions. Start-ups like AKASA are also actively developing deep learning-based automated coding systems, reporting state-of-the-art performance. These diverse efforts highlight the significant commercial and clinical interest in automated clinical coding and its potential to transform healthcare.

Conclusion: Charting the Future of Automated Clinical Coding in Primary Care

This article has explored the landscape of automated clinical coding, particularly within primary care, from the perspectives of AI researchers and clinical coding professionals. It has outlined the importance of this task, the challenges posed by deep learning approaches, and key directions for future research.

While technical challenges are significant, organizational factors are equally critical for successful deployment. Engaging clinical coders in the development and implementation process is essential, as emphasized in reviews of computer-assisted coding. However, the demanding nature of coders’ daily work can make their engagement in system testing and development challenging. Continued research support in medical informatics and computer science is crucial to address both technical and organizational hurdles.

The question remains: How far are we from achieving human-centered, explainable, intelligent, and robust automated clinical coding systems for primary care that can handle the complexities of real-world scenarios? While a precise timeline is difficult to predict, a clearer understanding of the challenges and promising research directions is emerging. The growing body of research and development efforts in both academia and industry suggests that significant advancements in AI-assisted clinical coding for primary care are likely in the next five years and beyond, paving the way for practical applications in the near future.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to the original article.

Supplementary information

Reporting Summary (2.8MB, pdf)

Acknowledgements

Please refer to the original article for acknowledgements.

Author contributions

Please refer to the original article for author contributions.

Data availability

The authors declare that all data supporting the findings of this perspective article are available within the paper.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Contributor Information

Hang Dong, Email: [email protected].

Honghan Wu, Email: [email protected].

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-022-00705-7.

References

• 1.Public Health Scotland. National Data Catalogue. General acute inpatient and day case – Scottish Morbidity Record (SMR01). https://www.ndc.scot.nhs.uk/National-Datasets/data.asp?SubID=5 (2020).
• 2.American Academy of Professional Coders (AAPC). What is medical coding? https://www.aapc.com/medical-coding/medical-coding.aspx (2022).
• 3.NHS Digital. Clinical coding for non coders. https://hscic.kahootz.com/gf2.ti/f/762498/30719205.1/PPSX/-/Coding_for_non_coders_automaticnew.ppsx (2017).
• 4.Enrico, C. In Guide to Health Informatics Ch. 24 (Taylor & Francis Group, 2015).
• 5.National Center for Health Statistics. International Classification of Diseases, (ICD-10-CM/PCS) transition – background. https://www.cdc.gov/nchs/icd/icd10cm_pcs_background.htm (2015).
• 6.Public Health Scotland. Terminology Services. Scottish Clinical Coding Standards. https://www.isdscotland.org/Products-and-services/Terminology-services/Clinical-coding-guidelines/ (2022).
• 7.Otero Varela, L. et al. International Classification of Diseases clinical coding training: an international survey. Health Inf. Manag.10.1177/18333583221106509 (2022) [DOI] [PMC free article] [PubMed]
• 8.Stanfill MH, Williams M, Fenton SH, Jenders RA, Hersh WR. A systematic literature review of automated clinical coding and classification systems. J. Am. Med Inf. Assoc. 2010;17:646–651. doi: 10.1136/jamia.2009.001024. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 9.Campbell S, Giadresco K. Computer-assisted clinical coding: a narrative review of the literature on its benefits, limitations, implementation and impact on clinical coding professionals. HIM J. 2020;49:5–18. doi: 10.1177/1833358319851305. [DOI] [PubMed] [Google Scholar]
• 10.Jiang, F. et al. Artificial intelligence in healthcare: past, present and future. Stroke Vasc. Neurol. 10.1136/svn-2017-000101 (2017) [DOI] [PMC free article] [PubMed]
• 11.Kaur, R., Ginige, J. A. & Obst, O. AI-based ICD coding and classification approaches using discharge summaries: A systematic literature review. Expert Syst. Appl. 118997 10.1016/j.eswa.2022.118997 (2022).
• 12.Ji, S., Sun, W., Dong, H., Wu, H. & Marttinen, P. A unified review of deep learning for automated medical coding. Preprint at arXivhttp://arxiv.org/abs/2201.02797 (2022).
• 13.Teng, F. et al. A review on deep neural networks for ICD coding. In IEEE Transactions on Knowledge and Data Engineering 1–19 (IEEE, 2022)
• 14.Alonso V, et al. Problems and barriers during the process of clinical coding: a Focus Group Study of coders’ perceptions. J. Med. Syst. 2020;44:62. doi: 10.1007/s10916-020-1532-x. [DOI] [PubMed] [Google Scholar]
• 15.Burns EM, et al. Systematic review of discharge coding accuracy. J. Public Health. 2012;34:138–148. doi: 10.1093/pubmed/fdr054. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 16.Public Health Scotland. Data quality assurance. Assessment of SMR01 Data Scotland Report 2019 V1. https://beta.isdscotland.org/media/7465/assessment-of-smr01-data-scotland-report-2019-v1.pdf (2019).
• 17.Wooldridge, M. The Road to Conscious Machines: The Story of AI (Penguin UK, 2020).
• 18.Russell, S. J. & Norvig, P. Artificial Intelligence: A Modern Approach, Global Edition (Pearson, 2021).
• 19.Mullenbach, J., Wiegreffe, S., Duke, J., Sun, J. & Eisenstein, J. Explainable prediction of medical codes from clinical text. In Proc. 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) 1101–1111 (Association for Computational Linguistics, 2018).
• 20.Johnson AEW, et al. MIMIC-III, a freely accessible critical care database. Sci. Data. 2016;3:1–9. doi: 10.1038/sdata.2016.35. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 21.Barrows Jr, R. C., Busuioc, M. & Friedman, C. Limited parsing of notational text visit notes: ad-hoc vs. NLP approaches. In Proc. AMIA Symposium 51 (American Medical Informatics Association, 2000). [PMC free article] [PubMed]
• 22.World Health Organization. ICD-11 for Mortality and Morbidity Statistics (WHO, 2022).
• 23.World Health Organization. WHO’s new International Classification of Diseases (ICD-11) comes into effect. https://www.who.int/news/item/11-02-2022-who-s-new-international-classification-of-diseases-(icd-11)-comes-into-effect (2022).
• 24.Gaebel W, Stricker J, Kerst A. Changes from ICD-10 to ICD-11 and future directions in psychiatric classification. Dialogues Clin. Neurosci. 2020;22:7–15. doi: 10.31887/DCNS.2020.22.1/wgaebel. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 25.Chute CG. The rendering of human phenotype and rare diseases in ICD-11. J. Inherit. Metab. Dis. 2018;41:563–569. doi: 10.1007/s10545-018-0172-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 26.World Health Organization. ICD-11 Reference Guide. 2.10 Precoordination and postcoordination. https://icdcdn.who.int/icd11referenceguide/en/html/index.html#precoordination-and-postcoordination (2022).
• 27.Bengio Y, Lecun Y, Hinton G. Deep learning for AI. Commun. ACM. 2021;64:58–65. doi: 10.1145/3448250. [DOI] [Google Scholar]
• 28.Dinwoodie HP, Howell RW. Automatic disease coding: the ‘fruit-machine’ method in general practice. Br. J. Prev. Soc. Med. 1973;27:59–62. doi: 10.1136/jech.27.1.59. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 29.Farkas, R., & Szarvas, G. Automatic construction of rule-based ICD-9-CM coding systems. *BMC Bioinformatics*9, 1–9 (2008). [DOI] [PMC free article] [PubMed]
• 30.Zhou L, Cheng C, Ou D, Huang H. Construction of a semi-automatic ICD-10 coding system. BMC Med. Inform. Decis. Mak. 2020;20:1–12. doi: 10.1186/s12911-020-1085-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 31.Shi, H., Xie, P., Hu, Z., Zhang, M. & Xing, E. P. Towards automated ICD coding using deep learning. Preprint at arXivhttps://arxiv.org/abs/1711.04075 (2017).
• 32.Karimi, S., Dai, X., Hassanzadeh, H. & Nguyen, A. Automatic diagnosis coding of radiology reports: a comparison of deep learning and conventional classification methods. in BioNLP 2017 328–332 (Association for Computational Linguistics, 2017).
• 33.acadTags. Awesome-medical-coding-NLP. https://github.com/acadTags/Awesome-medical-coding-NLP (2022).
• 34.Nam, J., Kim, J., Loza Mencía, E., Gurevych, I. & Fürnkranz, J. In Machine Learning and Knowledge Discovery in Databases (eds. Calders, T., Esposito, F., Hüllermeier, E. & Meo, R.) 437–452 (Springer, 2014).
• 35.Kraljevic, Z. et al. Multi-domain clinical natural language processing with MedCAT: the medical concept annotation toolkit. Artif. Intelligence Med. 117, 102083 (2021). [DOI] [PubMed]
• 36.Wiegreffe, S., Choi, E., Yan, S., Sun, J. & Eisenstein, J. Clinical concept extraction for document-level coding. In Proc. 18th BioNLP Workshop and Shared Task 261–272 (Association for Computational Linguistics, 2019)
• 37.Rios, A. & Kavuluru, R. Few-shot and zero-shot multi-label learning for structured label spaces. In Proc. 2018 Conference on Empirical Methods in Natural Language Processing 3132–3142 (Association for Computational Linguistics, 2018). [PMC free article] [PubMed]
• 38.Teng F, Yang W, Chen L, Huang L, Xu Q. Explainable prediction of medical codes with knowledge graphs. Front. Bioeng. Biotechnol. 2020;8:867. doi: 10.3389/fbioe.2020.00867. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 39.Xie, X., Xiong, Y., Yu, P. S. & Zhu, Y. EHR coding with multi-scale feature attention and structured knowledge graph propagation. In Proc. 28th ACM International Conference on Information and Knowledge Management 649–658 (ACM, 2019).
• 40.Cao, P. et al. Hypercore: hyperbolic and co-graph representation for automatic ICD coding. In Proc. 58th Annual Meeting of the Association for Computational Linguistics 3105–3114 (Association for Computational Linguistics, 2020).
• 41.Falis, M. et al. Ontological attention ensembles for capturing semantic concepts in ICD code prediction from clinical text. In Proc. Tenth International Workshop on Health Text Mining and Information Analysis (LOUHI 2019) 168–177 (Association for Computational Linguistics, 2019).
• 42.Falis, M., Dong, H., Birch, A. & Alex, B. CoPHE: a count-preserving hierarchical evaluation metric in large-scale multi-label text classification. In Proc. 2021 Conference on Empirical Methods in Natural Language Processing 907–912 (Association for Computational Linguistics, 2021).
• 43.Kukafka R, Bales ME, Burkhardt A, Friedman C. Human and automated coding of rehabilitation discharge summaries according to the international classification of functioning, disability, and health. J. Am. Med. Inform. Assoc. 2006;13:508–515. doi: 10.1197/jamia.M2107. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 44.Ji S, Hölttä M, Marttinen P. Does the magic of BERT apply to medical code assignment? A quantitative study. Computers Biol. Med. 2021;139:104998. doi: 10.1016/j.compbiomed.2021.104998. [DOI] [PubMed] [Google Scholar]
• 45.Sun, W., Ji, S., Cambria, E. & Marttinen, P. Multitask balanced and recalibrated network for medical code prediction. ACM Trans. Intelligent Syst. Technol. 10.1145/3563041 (2022)
• 46.Chalkidis, I. et al. An empirical study on large-scale multi-label text classification including few and zero-shot labels. In Proc. 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) 7503–7515 (Association for Computational Linguistics, 2020).
• 47.Wang, R. et al. Meta-LMTC: meta-learning for large-scale multi-label text classification. In Proc. 2021 Conference on Empirical Methods in Natural Language Processing 8633–8646 (Association for Computational Linguistics, 2021).
• 48.Xu, K. et al. Multimodal machine learning for automated ICD coding. In Machine Learning for Healthcare Conference 197–215 (PMLR, 2019).
• 49.Liu, Y., Cheng, H., Klopfer, R., Gormley, M. R. & Schaaf, T. Effective convolutional attention network for multi-label clinical document classification. In Proc. 2021 Conference on Empirical Methods in Natural Language Processing 5941–5953 (Association for Computational Linguistics, 2021).
• 50.Kim, B. H., & Ganapathi, V. Read, attend, and code: pushing the limits of medical codes prediction from clinical notes by machines. In Machine Learning for Healthcare Conference 196–208 (PMLR, 2021).
• 51.Yuan, Z., Tan, C., & Huang, S. Code synonyms do matter: multiple synonyms matching network for automatic ICD coding. In Proc. 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) 808–814 (Association for Computational Linguistics, 2022).
• 52.Huang, C. W., Tsai, S. C., & Chen, Y. N. PLM-ICD: automatic ICD coding with pretrained language models. In Proc. 4th Clinical Natural Language Processing Workshop 10–20 (Association for Computational Linguistics, 2022).
• 53.Terminology and Classifications Delivery Service, National Health Service Digital. National Clinical Coding Standards ICD-10 5th Edition. https://classbrowser.nhs.uk/ref_books/ICD-10_2021_5th_Ed_NCCS.pdf (2021).
• 54.Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: pre-training of deep bidirectional transformers for language understanding. In Proc. 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) 4171–4186 (Association for Computational Linguistics, 2019).
• 55.Feucht, M., Wu, Z., Althammer, S. & Tresp, V. Description-based label attention classifier for explainable ICD-9 classification. In Proc. Seventh Workshop on Noisy User-generated Text (W-NUT 2021) 62–66 (Association for Computational Linguistics, 2021).
• 56.Yogarajan, V., Pfahringer, B., Smith, T., & Montiel, J. In Artificial Neural Networks and Machine Learning – ICANN 2022 (eds. Pimenidis, E., Angelov, P., Jayne, C., Papaleonidas, A. & Aydin, M.) 209–221 (Springer Nature Switzerland, 2022).
• 57.Michalopoulos, G., Malyska, M., Sahar, N., Wong, A. & Chen, H. ICDBigBird: a contextual embedding model for ICD code classification. In Proc. 21st Workshop on Biomedical Language Processing 330–336 (Association for Computational Linguistics, 2022).
• 58.Liu, J., Capurro, D., Nguyen, A. & Verspoor, K. “Note Bloat” impacts deep learning-based NLP models for clinical prediction tasks. J. Biomed. Inform. 133, 104149 (2022) [DOI] [PubMed]
• 59.Searle T, Ibrahim Z, Teo J, Dobson R. Estimating redundancy in clinical text. J. Biomed. Inform. 2021;124:103938. doi: 10.1016/j.jbi.2021.103938. [DOI] [PubMed] [Google Scholar]
• 60.Gao, S. et al. Limitations of transformers on clinical text classification. IEEE J. Biomed. Health Inform. 25, 3596–3607 (2021). [DOI] [PMC free article] [PubMed]
• 61.Dong H, Suárez-Paniagua V, Whiteley W, Wu H. Explainable automated coding of clinical notes using hierarchical label-wise attention networks and label embedding initialisation. J. Biomed. Inform. 2021;116:103728. doi: 10.1016/j.jbi.2021.103728. [DOI] [PubMed] [Google Scholar]
• 62.Searle, T., Ibrahim, Z. & Dobson, R. Experimental evaluation and development of a silver-standard for the MIMIC-III clinical coding dataset. In Proc. 19th SIGBioMed Workshop on Biomedical Language Processing 76–85 (Association for Computational Linguistics, 2020).
• 63.Baumel, T., Nassour-Kassis, J., Cohen, R., Elhadad, M. & Elhadad, N. In Workshops at the Thirty-Second AAAI Conference on Artificial Intelligence. 409-416 (2018).
• 64.Searle, T., Kraljevic, Z., Bendayan, R., Bean, D. & Dobson, R. MedCATTrainer: a biomedical free text annotation interface with active learning and research use case specific customisation. In Proc. 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP): System Demonstrations 139–144 (Association for Computational Linguistics, 2019).
• 65.Wu H, et al. SemEHR: a general-purpose semantic search system to surface semantic data from clinical notes for tailored care, trial recruitment, and clinical research. J. Am. Med. Inform. Assoc. 2018;25:530–537. doi: 10.1093/jamia/ocx160. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 66.Dong, H. et al. Rare disease identification from clinical notes with ontologies and weak supervision. In 2021 43rd Annual International Conference of the IEEE Engineering in Medicine Biology Society (EMBC) 2294–2298 (EMBC, 2021). [DOI] [PubMed]
• 67.Dong, H. et al. Ontology-based and weakly supervised rare disease phenotyping from clinical notes. Preprint at http://arxiv.org/abs/2205.05656 (2022).
• 68.Ferreira, M. D. et al. Active learning for medical code assignment. In Workshops from ACM Conference on Health, Inference, and Learning (CHIL) 2021. Preprint at arXivhttp://arxiv.org/abs/2104.05741 (2021).
• 69.Chen, J. et al. Knowledge-aware zero-shot learning: survey and perspective. In Proc. Thirtieth International Joint Conference on Artificial Intelligence, IJCAI 2021 4366–4373 (IJCAI, 2021).
• 70.Falis, M. Blood is thicker than water, a hierarchical evaluation metric for document classification. https://www.ltg.ed.ac.uk/blood-is-thicker-than-water/ (2021).
• 71.Healthcare Cost and Utilization Project (HCUP). Clinical classifications software (CCS) for ICD-9-CM. https://www.hcup-us.ahrq.gov/toolssoftware/ccs/ccs.jsp (2017).
• 72.Hahn U, Oleynik M. Medical information extraction in the age of deep learning. Yearb. Med. Inform. 2020;29:208–220. doi: 10.1055/s-0040-1702001. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 73.Falis, M., Dong, H., Birch, A. & Alex, B. Horses to zebras: ontology-guided data augmentation and synthesis for ICD-9 coding. In Proc. 21st Workshop on Biomedical Language Processing 389–401 (Association for Computational Linguistics, 2022).
• 74.DeYoung, J., Shing, H.-C., Kong, L., Winestock, C. & Shivade, C. Entity anchored ICD coding. Accepted to American Medical Informatics Association (AMIA) 2022 Annual Symposium. Preprint at arXivhttp://arxiv.org/abs/2208.07444 (2022).
• 75.Liu J, Capurro D, Nguyen A, Verspoor K. Early prediction of diagnostic-related groups and estimation of hospital cost by processing clinical notes. NPJ Digital Med. 2021;4:1–8. doi: 10.1038/s41746-021-00474-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 76.Donnelly K. SNOMED-CT: The advanced terminology and coding system for eHealth. Stud. Health Technol. Inform. 2006;121:279. [PubMed] [Google Scholar]
• 77.Vasant, D. et al. ORDO: an ontology connecting rare disease, epidemiology and genetic data. In Bio-Ontology @ ISMB 2014. 1-4. https://www.researchgate.net/publication/281824026_ORDO_An_Ontology_Connecting_Rare_Disease_Epidemiology_and_Genetic_Data (2014).
• 78.Alex B, et al. Text mining brain imaging reports. J. Biomed. Semant. 2019;10:1–11. doi: 10.1186/s13326-019-0211-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 79.Ford E, Carroll JA, Smith HE, Scott D, Cassell JA. Extracting information from the text of electronic medical records to improve case detection: a systematic review. J. Am. Med. Inform. Assoc. 2016;23:1007–1015. doi: 10.1093/jamia/ocv180. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 80.Rannikmäe K, et al. Developing automated methods for disease subtyping in UK Biobank: an exemplar study on stroke. BMC Med. Inform. Decis. Mak. 2021;21:1–9. doi: 10.1186/s12911-021-01556-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 81.Lovelace, J., Hurley, N. C., Haimovich, A. D. & Mortazavi, B. J. Dynamically extracting outcome-specific problem lists from clinical notes with guided multi-headed attention. In Machine Learning for Healthcare Conference 245–270 (PMLR, 2020).
• 82.Rannikmäe K, et al. Accuracy of identifying incident stroke cases from linked health care data in UK Biobank. Neurology. 2020;95:e697–e707. doi: 10.1212/WNL.0000000000009924. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 83.Noor K, et al. Deployment of a free-text analytics platform at a UK National Health Service Research Hospital: CogStack at University College London Hospitals. JMIR Med. Inform. 2022;10:e38122. doi: 10.2196/38122. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 84.King’s College Hospital NHS Foundation Trust. CogStack wins an artificial intelligence in health and care. https://www.kch.nhs.uk/news/public/news/view/34965 (2021).
• 85.Amazon Web Services. ICD-10-CM linking. https://docs.aws.amazon.com/comprehend-medical/latest/dev/ontology-icd10.html (2022).
• 86.Azure. What is text analytics for health in Azure Cognitive Service for Language? https://docs.microsoft.com/en-us/azure/cognitive-services/language-service/text-analytics-for-health/overview?tabs=ner (2022).
• 87.Google Cloud. Healthcare natural language API. https://cloud.google.com/healthcare-api/docs/concepts/nlp (2022).
• 88.Bodnari, A. Healthcare gets more productive with new industry-specific AI tools. https://cloud.google.com/blog/topics/healthcare-life-sciences/now-in-preview-healthcare-natural-language-api-and-automl-entity-extraction-for-healthcare (2020).
• 89.Amazon Web Services. Announcing ICD-10-CM and RxNorm ontology linking for Amazon Comprehend Medical. https://aws.amazon.com/about-aws/whats-new/2019/12/announcing-icd-10-cm-rxnorm-ontology-linking-amazon-comprehend-medical/ (2019).
• 90.Miranda, M. Automated clinical coding. The AI-based solution to address the critical shortage of clinical coders. https://www2.deloitte.com/au/en/blog/consulting-blog/2020/automated-clinical-coding.html (2020).
• 91.Optum Inc. Enterprise computer-assisted coding (CAC). https://www.optum360.com/solutions/coding-and-documentation/coding-and-cdi-technology/enterprise-cac.html (2022).
• 92.Capita plc. Creating better health outcomes with automated clinical coding. https://www.capita.com/expertise/industry-specific-services/health-services/healthcare-business-operations/clinical-coding/automated-clinical-coding (2022).
• 93.CHKS. Automated clinical coding. https://www.chks.co.uk/Clinical-coding (2022).
• 94.Mace, S. Making medical coding better and faster with artificial intelligence. Medical Technology Schools. https://www.medicaltechnologyschools.com/health-information-technology/medical-coding-and-artificial-intelligence (2021).

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Reporting Summary (2.8MB, pdf)

Data Availability Statement

The authors declare that all data supporting the findings of this perspective article are available within the paper.