phdassistance

  • Have any questions?

Natural Language Processing of Clinical Notes for Automated ICD-10 Coding Accuracy

 Info: Natural Language Processing of Clinical Notes for Automated ICD-10 Coding Accuracy Topics I phdassistance.com

Published: 17th August in Natural Language Processing of Clinical Notes for Automated ICD-10 Coding Accuracy Topics I phdassistance.com

Share this:

Facebook
Twitter
Whatsapp
Linkedin

Related Services

Our academic writing and marking services can help you

Introduction

The increasing number of EHRs demands effective and precise coding methods for clinical cases. Manual ICD-10 coding is usually associated with low speed and the inability to cope with long and complicated clinical texts. Modern advancements in artificial intelligence and NLP for Automated ICD-10 Coding offer possibilities to identify useful information and code diseases more effectively. Yet, the problems of large coding spaces, data imbalance, ambiguous terminology, long clinical documents, and the lack of model interpretability remain open to investigation.

Proposed PhD Topic 1: Transformer-Based Natural Language Processing for Improving ICD-10 Coding from Long and Complex Clinical Notes
Background Context:

The growing use of electronic health records has offered the chance of automation in the process of extracting diagnoses from unstructured medical texts. The use of Natural Language Processing in Medical Coding will enable the translation of clinical notes into coded diseases, thereby alleviating the burden of manual work and ensuring consistency. But coding is quite difficult when the clinical condition is uncommon. Lee & Im (2025) studied the automatic classification of ICD based on clinical text and found promising results in various stages of cause-of-death classification. Nevertheless, the researchers’ results have revealed that there is greater difficulty in classifying rare illnesses and underrepresented categories owing to scarcity of training samples. Another finding of the researchers was the existence of ambiguity and confusion in clinical language, as well as similarity between diseases. Such findings give rise to the possibility of undertaking doctoral research and developing a special framework that will enhance classification accuracy of rare and clinically important diseases while keeping performance on common ones at the same high level.

PhD-Level Verification:

This research shows the variability of model performance on rare compared to common diagnostic categories. This is a clear research gap when it comes to developing classification methods that are optimised for rare clinical conditions. A PhD research project can focus on developing strategies for data augmentation, transfer learning, and class-specific learning for rare ICD-10 categories.

Research Questions:
  • How can clinical text models improve ICD-10 classification for rare diseases?
  • Which learning strategies can reduce errors caused by limited examples?
  • Can specialised training approaches improve performance across both common and rare diagnostic categories?
    • Contributions at the PhD-Level:
  • Development of a rare-condition-aware clinical text classification framework.
  • Development of strategies for learning from limited examples.
  • Identification of linguistic factors associated with rare-code misclassification.
    • Suggested Readings:

    Lee, S., & Im, G. (2025). Machine learning for automated cause-of-death classification from 2021 to 2022 in Korea: development and validation of an ICD-10 prediction model.

    Automated Medical Coding
    Proposed PhD Topic 2: Addressing Rare Disease and Imbalanced ICD-10 Classes Through Active Learning for Automated Medical Coding
    Background Context:

    The incorporation of AI into documentation in the healthcare industry has brought about the use of Automated Medical Coding, especially in situations where there are huge numbers of electronic documents that need to be processed in an efficient manner. However, accurate predictions might not be enough in practical settings within the medical settings. Medical coders, physicians, auditors, and health administrators should understand the process behind an automatically coded document. The systematic review on the use of artificial intelligence in ICD coding done by Mousavi Baigi et al. (2025) revealed poor interpretability and transparency as the key problems. They pointed out that even if the predictions produced using deep learning models are highly accurate, very little information is generated regarding the clinical evidence behind such predictions. The need for transparency, interdisciplinary collaboration, and auditable coding was stressed. Such gaps in the existing literature present the scope for developing an architecture where, along with predicting the proper classification, it is possible to determine the clinical evidence behind it.

    PhD-Level Verification:

    This research demonstrates the variability in the performance of models when dealing with rare classes versus common ones. This represents a research gap in developing classification models optimised for rare diseases. The following research ideas could be proposed in a PhD study: data augmentation for rare ICD-10 categories; transfer learning for rare ICD-10 categories; class-specific learning for rare ICD-10 categories.

    Research Questions:
  • How can explainable AI improve transparency in automated clinical coding?
  • Which forms of explanation are most useful to professional coders?
  • Can interpretable predictions improve user confidence and verification accuracy?
    • PhD-Level Contributions:
  • Development of an explainable clinical coding framework.
  • Identification of textual evidence supporting each predicted code.
  • Development of an auditable decision-support interface.
    • Suggested Readings:

    Mousavi Baigi, S. F., Sarbaz, M., Darroudi, A., et al. (2025). Artificial Intelligence-based Automated International Classification of Diseases Coding: A Systematic Review.

    Proposed Dissertation topic 3: Explainable AI for Transparent and Auditable ICD-10 Coding from Clinical Narratives
    Background Context:

    Clinical documentation is full of facts that do not always get understood by applying the keyword approach to the documentation. It could mean that the fact stated within the note is about a condition present now, something diagnosed in the past, a disease that is suspected, negative findings, or something that happened before being admitted to the hospital. Clinical Notes NLP can help identify and interpret these different forms of clinical information by considering the context in which medical terms are used. Gaspar et al. (2025) evaluated the effectiveness of NLP in detecting instances of bleeding in hospital discharge notes against traditional ICD-10-based methods. The results showed that narrative processing provided clinically relevant information that would be challenging to detect using structured codes only. The NLP method also managed to discern clinically relevant and severe bleeding cases, as well as historical ones, while the rule-based one failed to represent temporal information effectively. Therefore, the results indicate the importance of contextual analysis of clinical narratives as an area for further doctoral study due to the ambiguous nature of the meaning of some medical terms depending on temporal or clinical context.

    PhD Level Verification:

    According to Gaspar et al., NLP can capture clinically significant data in discharge summaries that are not captured by conventional ICD-10 methods. This study was also able to prove the importance of differentiating between historical events and severity during identification of the need for further temporal reasoning and validation. As PhD level, this gives the chance of exploring the contextual understanding of clinical narratives which include temporal information, negations, and historical versus current diseases. The suggested research can help in developing a context-aware classification model as well as identifying the types of errors that occur in context.

    Research Questions:
  • How can contextual information improve diagnostic classification from clinical narratives?
  • How can models distinguish current conditions from historical or negated conditions?
  • Does contextual reasoning reduce incorrect diagnostic classifications?
    • PhD-Level Contributions:
  • Development of a context-aware clinical text framework.
  • Integration of temporal and negation information into diagnostic classification.
  • Development of mechanisms for detecting contextual ambiguity.
    • Suggested Readings:

    Gaspar, F., Zayene, M., Coumau, C., et al. (2025). Natural Language Processing and ICD-10 Coding for Detecting Bleeding Events in Discharge Summaries: Comparative Cross-Sectional Study.

    Proposed Dissertation Topic 4: Context-Aware Clinical Notes NLP for Distinguishing Diagnoses, Symptoms, History, and Clinical Events in ICD-10 Coding
    Background Context:

    Large-scale clinical classification requires a lot of labelled medical documents, but making high-quality labels is expensive since trained health care specialists must analyse the documents and assign corresponding disease codes, especially since some disease groups can be quite frequent and others rare. Kaur and Ginige have analysed different algorithms to perform automated coding and have emphasised the role of machine learning and deep learning methods to assign ICD codes from clinical data. This reanalysis can serve as a starting point for further study in the development of better learning strategies that can take advantage of both labelled and unlabeled medical data. Active learning is one example where the model is allowed to pick the most valuable data to annotate by the expert. This approach has the potential to decrease the need for annotation of all data and to improve the process of learning difficult and rare diagnostic classes.

    PhD-Level Verification:

    Kaur & Ginige have shown how useful the algorithms can be when applied to ICD coding tasks, providing an initial background for future work on more effective and efficient machine learning-based coding. ICD-10 Coding Automation reduce the manual effort involved in assigning diagnostic codes. One important issue that requires further research is the problem of generating large enough labeled clinical datasets. PhD research may involve the application of active learning for identifying the most useful clinical documents to label manually, and evaluating the improvements in annotation and classification of ICD-10 codes achieved in this way.

    Research Questions:
  • How can active learning reduce the amount of expert-labelled data required?
  • Which clinical records provide the greatest value for model training?
  • Can active learning improve classification of infrequent diagnostic categories?
    • Contributions at the PhD-Level:
  • Development of an active-learning framework for clinical classification.
  • Intelligent selection of records requiring expert annotation.
  • Reduction in manual annotation requirements.
    • Suggested Readings:

    Kaur, R., & Ginige, J. A. (2018). Comparative analysis of algorithmic approaches for auto-coding with ICD-10-AM and ACHI.

    Proposed Dissertation Topic 5: Human-in-the-Loop Natural Language Processing for Reliable and Generalisable Automated ICD-10 Coding
    Background Context:

    The large number of categories in ICD-10 makes classification a complex problem for AI. There is a possibility that medical documents must be classified from general diseases to very specific diagnostic codes, where similar terms can occur for related codes. This means that considering each code as a separate class will make the problem even more complex. Moons et al. have explored deep learning algorithms that could classify clinical data into ICD codes and emphasised the importance of hierarchical objectives when it comes to handling big coding space and a small sample size. AI Medical Coding therefore requires approaches that can understand relationships between different levels of the ICD-10 classification system. Moons et al.’s study indicated that hierarchical methods may assist in helping the model learn about relationships between classification levels. But noisy clinical documentation and increasing coding hierarchy complexity continue to be challenging issues for automatic large-scale classification. That opens the way for building a hierarchical structure in which the model learns about the disease group, category, subcategory, and finally ICD code.

    PhD-Level Verification:

    Moons et al. showed the possibilities of deep neural networks and hierarchical learning in automatic classification according to ICD codes. This shows that one may use hierarchical relations for solving the problems of large-scale classification. PhD level research might include an analysis of hierarchical approaches on noisy data sets. The experiment would be conducted to analyze the performance of the flat and hierarchical approaches through performance on a code level and hierarchical errors to evaluate the benefit from using ICD-10 hierarchy.

    Research Questions:
  • How can hierarchical deep learning improve large-scale ICD-10 classification?
  • Can hierarchical prediction reduce errors between closely related codes?
  • How does hierarchical modelling perform when clinical training data are limited or noisy?
    • PhD-Level Contributions:
  • Development of a hierarchical ICD classification framework.
  • Integration of the ICD hierarchy into a deep learning architecture.
  • Development of strategies for reducing errors between related diagnostic codes.
    • Suggested Readings:

    Moons et al. (2020). Deep neural networks for automatically classifying clinical data into ICD codes.

    Need assistance finalising your dissertation topic in natural language processing in clinical notes? Selecting a strong, researchable topic can be challenging — but you don’t have to do it alone.
    Our research consultants can help refine your ideas, identify literature gaps, and guide you toward a topic that aligns with current academic trends and your programme requirements.
    Contact us to begin one-on-one topic development and refinement with PhdAssistance.com Research Lab.

    FAQs:

    1. How can NLP automate ICD-10 coding?
      NLP extracts relevant clinical information from medical documentation and maps it to appropriate ICD-10 codes. This can reduce manual coding effort and improve the speed and consistency of coding.
    2. How does NLP analyse clinical notes for ICD-10 coding?
      NLP processes clinical terms, diagnoses, symptoms, context, and relationships within unstructured notes. It then uses these extracted features to predict relevant ICD-10 codes.
    3. Can AI improve ICD-10 coding accuracy?
      Yes, AI can identify clinical information that may be missed through conventional rule-based coding approaches. However, accuracy depends on data quality, model design, coding complexity, and clinical validation.
    4. Which NLP techniques are used for automated medical coding?
      Common approaches include text classification, word embeddings, transformers, deep learning, and encoder-decoder models. These techniques help models understand clinical language and associate it with relevant diagnostic codes.
    5. How accurate is automated ICD-10 coding?
      Accuracy varies depending on the dataset, coding categories, clinical notes, and model used. Research shows promising results, but rare codes, imbalanced data, and complex narratives can reduce performance.
    6. What are the key challenges in NLP-based ICD-10 coding?
      Major challenges include lengthy clinical notes, rare and imbalanced codes, ambiguous terminology, limited labelled data, and explainability. Ensuring reliable performance across different healthcare settings also remains an important research challenge.

    Share this:

    Facebook
    Twitter
    Whatsapp
    Linkedin

    Cite this work

    Study Resources

    Free resources to assist you with your university studies!

    Research Questions