TL;DR
This paper proposes a novel approach to predict hospital-induced delirium by integrating structured electronic health record (EHR) data with unstructured clinical notes using machine learning and natural language processing (NLP). The goal is to create a more accurate and robust prediction model that can be integrated into EHR systems to provide real-time decision support and improve patient safety.
What Problem Does This Paper Address?
Hospital-induced delirium is a common, costly, and potentially preventable condition affecting a significant portion of hospitalized older adults. Existing prediction models often rely solely on structured EHR data, which may miss crucial contextual information present in unstructured clinical notes, such as nurses' observations. This limits their accuracy and ability to identify at-risk patients effectively. The paper aims to bridge this gap by developing a comprehensive prediction model that leverages both structured and unstructured data, leading to better identification and prevention of hospital-induced delirium.
Methods
The study employs a multi-faceted approach to develop and validate prediction models for hospital-induced delirium. Key methods include:
- Computable Phenotype Development: Creating a standardized, automated algorithm to define "hospital-induced delirium" as the outcome measure, going beyond simple ICD codes to incorporate nursing assessments.
- Structured Data Modeling: Developing a baseline logistic regression model using expert-identified structured data variables. Machine learning techniques like decision trees, bagging, random forests, and adaptive boosting will also be used to build predictive models from structured data.
- Text Mining:
- Supervised and Unsupervised NLP: Utilizing these techniques to identify risk factors, signs, and symptoms of delirium from nursing notes. This involves creating a keyword thesaurus and training annotators to label text.
- Named Entity Recognition (NER): Employing conditional random fields (CRF) to identify and classify specific clinical concepts within text.
- Topic Modeling (LDA): Using latent Dirichlet allocation to discover previously unknown terms related to delirium and generate hypotheses about risk factors or early signs.
- Integrated Model Development: Translating text data into numerical vectors using Word2Vec models and combining these with structured data features. This integrated dataset will then be used to train prediction models using logistic regression with LASSO and other machine learning approaches.
- Validation: Models will be evaluated using 10-fold cross-validation and externally validated on a separate dataset. The study adheres to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement.
Datasets
The study will utilize data from approximately 332,230 encounters from two academic medical centers in North Florida, spanning from January 2012 to May 2021. This dataset includes demographic information, inpatient and outpatient clinical encounter data, diagnoses, procedures, lab results, medications, nursing assessments, and perioperative anesthesia information system data. The data is sourced from the Integrated Data Repository (IDR), which aggregates information from various clinical and administrative systems, including the Epic electronic medical record.
Key Insights and Ideas
- The Power of Unstructured Data: The core idea is that unstructured clinical notes, particularly those from nurses, contain invaluable contextual information that structured data alone misses, significantly improving the accuracy of delirium prediction.
- Bridging the Gap: The study aims to create a "computable phenotype" that accurately captures the outcome of delirium, moving beyond simple coding to a more nuanced definition informed by clinical practice.
- Hybrid Modeling: Combining traditional logistic regression with advanced machine learning and NLP techniques offers a powerful way to extract insights from complex, multi-modal EHR data.
- Learning Health System Infrastructure: The project aims to build a research-data infrastructure that can process and analyze clinical text data in near real-time, paving the way for immediate clinical decision support.
Key Results
The paper is a protocol, so specific quantitative results are not yet available. However, the expected outcomes include:
- Development and validation of a computable phenotype for hospital-induced delirium.
- Creation of multiple prediction models: an expert-informed logistic regression model, models using structured data, and an integrated model combining structured and unstructured data.
- A durable, high-performing research-data infrastructure capable of processing and analyzing clinical text data in near real-time.
- Findings will be disseminated at scientific conferences and in peer-reviewed journals.
Limitations and Future Work
Limitations:
- EHR Data Quality: The authors acknowledge that EHR data can have issues with completeness, accuracy, and complexity. They plan to mitigate this by regularly consulting with their IC task force to understand how clinicians document data.
- Interpretability vs. Accuracy: Machine learning models, while potentially more accurate, can be less interpretable than traditional models. The study will attempt to balance performance gains with interpretability.
Future Work:
- The developed models have the potential to be integrated into EHR systems to provide point-of-care decision support for preventing harm and improving care quality.
- The research-data infrastructure developed in this study will support ongoing research on hospital-induced delirium and can be scaled to other health systems and research networks like OneFlorida.
- The findings will inform future studies and improve clinical decision-making for hospital-induced delirium.
Summary updated