Using machine learning for automated de-identification and clinical coding of free text data in electronic medical records

Download files
Access & Terms of Use
open access
Copyright: Liu, Leibo
The widespread adoption of Electronic Medical Records (EMRs) in hospitals continues to increase the amount of patient data that are digitally stored. Although the primary use of the EMR is to support patient care by making all relevant information accessible, governments and health organisations are looking for ways to unleash the potential of these data for secondary purposes, including clinical research, disease surveillance and automation of healthcare processes and workflows. EMRs include large quantities of free text documents that contain valuable information. The greatest challenges in using the free text data in EMRs include the removal of personally identifiable information and the extraction of relevant information for specific tasks such as clinical coding. Machine learning-based automated approaches can potentially address these challenges. This thesis aims to explore and improve the performance of machine learning models for automated de-identification and clinical coding of free text data in EMRs, as captured in hospital discharge summaries, and facilitate the applications of these approaches in real-world use cases. It does so by 1) implementing an end-to-end de-identification framework using an ensemble of deep learning models; 2) developing a web-based system for de-identification of free text (DEFT) with an interactive learning loop; 3) proposing and implementing a hierarchical label-wise attention transformer model (HiLAT) for explainable International Classification of Diseases (ICD) coding; and 4) investigating the use of extreme multi-label long text transformer-based models for automated ICD coding. The key findings include: 1) An end-to-end framework using an ensemble of deep learning base-models achieved excellent performance on the de-identification task. 2) A new web-based de-identification software system (DEFT) can be readily and easily adopted by data custodians and researchers to perform de-identification of free text in EMRs. 3) A novel domain-specific transformer-based model (HiLAT) achieved state-of-the-art (SOTA) results for predicting ICD codes on a Medical Information Mart for Intensive Care (MIMIC-III) dataset comprising the discharge summaries (n=12,808) that are coded with at least one of the most 50 frequent diagnosis and procedure codes. In addition, the label-wise attention scores for the tokens in the discharge summary presented a potential explainability tool for checking the face validity of ICD code predictions. 4) An optimised transformer-based model, PLM-ICD, achieved the latest SOTA results for ICD coding on all the discharge summaries of the MIMIC-III dataset (n=59,652). The segmentation method, which split the long text consecutively into multiple small chunks, addressed the problem of applying transformer-based models to long text datasets. However, using transformer-based models on extremely large label sets needs further research. These findings demonstrate that the de-identification and clinical coding tasks can benefit from the application of machine learning approaches, present practical tools for implementing these approaches, and highlight priorities for further research.
Persistent link to this record
Link to Publisher Version
Link to Open Access Version
Additional Link
Conference Proceedings Editor(s)
Other Contributor(s)
Corporate/Industry Contributor(s)
Publication Year
Resource Type
Degree Type
PhD Doctorate
UNSW Faculty
download public version.pdf 9.94 MB Adobe Portable Document Format
Related dataset(s)