Knowledge Graph Embeddings for ICU readmission prediction | JoVE Visualize

Area of Science:

Biomedical Informatics
Machine Learning
Clinical Decision Support

Background:

Intensive Care Unit (ICU) readmissions pose significant health risks and financial burdens.
Existing machine learning models for predicting ICU readmissions often overlook the contextual meaning of clinical data.
Ontologies and Knowledge Graphs offer a way to integrate scientific knowledge and context into data analysis.

Purpose of the Study:

To develop a novel approach for predicting 30-day ICU readmissions by enriching Electronic Health Record (EHR) data with semantic annotations.
To leverage ontologies and Knowledge Graphs to create contextualized representations of patient data for improved predictive modeling.
To evaluate the performance of the proposed method against baseline and state-of-the-art approaches.

Main Methods:

Enriched the MIMIC-III dataset with patient-oriented annotations linked to biomedical ontologies.
Constructed a Knowledge Graph representing patient data within the framework of biomedical ontologies.
Developed machine learning models utilizing Knowledge Graph embeddings for ICU readmission risk prediction, including variants targeting different time points during ICU stays.

Main Results:

The developed predictive models achieved a mean Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.827 and an Area Under the Precision-Recall Curve (AUC-PR) of 0.691.
The proposed approach demonstrated superior performance compared to baseline and state-of-the-art methods.
The approach has the potential to identify a significant portion of ICU patients at risk of readmission.

Conclusions:

Integrating ontologies and Knowledge Graphs into clinical machine learning applications provides a powerful method for incorporating scientific context.
This approach enables the creation of unified data representations from diverse EHR information types.
The study highlights the significant potential impact of semantic enrichment and Knowledge Graph embeddings in clinical decision-making for preventing ICU readmissions.