Developing Delirium Prediction Models in Clinical Settings: Integrating Machine Learning and Natural Language Processing with Patient Data, Clinical Notes, and Antipsychotics Data
Loading...
Citations
Altmetric:
Altmetric::
Download
Associated Records
Authors
File Licence
Editors
Date
2023-12
Source Title
Journal Title
Journal ISSN
Volume Title
Publisher
Melbourne Authors
Affiliation
Computing and Information Systems
Document Type
PhD thesis
Subject Code
Citation
Access Status
Open Access
DOI
Access URL
Open Access URL
Open Access at PMC
Linked Resource URL
NHMRC Grant Code
ARC Grant Code
Description
© 2024 Sobia Amjad
Abstract
This thesis details the effort to improve delirium identification. Delirium is an acute confusion in 11% to 42% of hospitalised patients and 60-80% of elderly patients. It is a serious neuropsychiatric syndrome frequently underdiagnosed, resulting in extended hospital stays, higher mortality rates and increased health care costs. Furthermore, because it is a cognitive condition, there is no clear definition, leading to high levels of overlap with other neuropsychiatric disorders.
Electronic health records (EHR) data has significantly advanced health research, offering automated tools for predictive analysis. The objective is to leverage EHR to develop machine learning models (ML), including those employing natural language processing (NLP) techniques for delirium prediction. The ML model assesses whether a patient will likely receive an ICD-10 code for delirium within 24 to 48 hours of hospital presentation, encouraging timely interventions for patients at risk in the wards. The data collected from tertiary-level hospitals explores the key features that contribute to the performance of predictive models. The study focuses on optimising early delirium prediction, thereby refining the diagnostic process for delirium and enhancing both the timeliness and accuracy of results.
The first study, "Machine Learning-Based Delirium Prediction in Hospitalized Patients Using Routinely Available Administrative and Clinical Data," developed models using two hospital data, proving the generalizability of our predictive models. This study worked with data from around 200,000 patients. The challenge was that delirium-positive cases were rare, less than 5\%. Thus, we directed our research towards using language in clinical notes to identify delirium. The second study, "Advancing Delirium Classification: A Clinical Notes-based Natural Language Processing-Supported Machine Learning Model," applied NLP to classify delirium at the clinical notes level. This method allowed us to explore delirium-suggestive words within the notes, thus enhancing classification accuracy beyond traditional dictionary-based approaches.
After applying methods on structured clinical data and notes, we implemented those methods for the patient-level delirium prediction. Also, we incorporated antipsychotic data into our prediction model in the third study entitled "EHR-Based Delirium Prediction: A Unified Data-driven Model with Clinical Data, Notes, and Antipsychotics". However, just like we faced a challenge with the rare cases of delirium in our first project, we also dealt with the issue of having limited data on antipsychotics and notes. Despite these challenges, we managed to navigate the complex nature of healthcare data. We developed a model that leverages all available data from the initial 24 to 48 hours of hospital presentation, which delivered promising results. Our study also looked carefully at how to use drug-related information without letting it create bias in our delirium predictions.
Classification techniques included Logistic Regression, Extreme gradient boosting, Support Vector Machines, and Random Forests. The Logistic regression model is discussed in main studies due to its common use in medical research and acceptance within the broader artificial intelligence community. We employed visualizations to explain predictions and highlight the transparency of our predictive models. Moreover, we presented visual graphs for each study to explain the model and demonstrate their reliability for delirium prediction to gain the trust of healthcare professionals.
Notes
Terms of Use
Keywords
Machine learning, Disease prediction, Document classification, Natural language processing, Delirium