Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Prediction of 30-day hospital readmission with clinical notes and EHR information

Tiago Almeida, Plinio Moreno, Catarina Barata

TL;DR

This paper addresses predicting 30-day hospital readmission by integrating structured EHR data and unstructured clinical notes. It proposes a scalable multimodal GraphSAGE-based GNN where each admission is a node, features come from demographic data, ICD embeddings for diagnoses/procedures, lab abnormalities, and BioClinicalBERT-derived note embeddings, with edges built via cosine similarity using FAISS. On the MIMIC-IV dataset, the model achieves AUROC of 0.727 and balanced accuracy of 0.667, outperforming logistic regression and MLP baselines, and illustrating the value of multimodal fusion. Compared to state-of-the-art multimodal readmission methods, the approach attains competitive performance on a larger-scale graph, while noting limitations in missing data handling and lack of temporal modeling.

Abstract

High hospital readmission rates are associated with significant costs and health risks for patients. Therefore, it is critical to develop predictive models that can support clinicians to determine whether or not a patient will return to the hospital in a relatively short period of time (e.g, 30-days). Nowadays, it is possible to collect both structured (electronic health records - EHR) and unstructured information (clinical notes) about a patient hospital event, all potentially containing relevant information for a predictive model. However, their integration is challenging. In this work we explore the combination of clinical notes and EHRs to predict 30-day hospital readmissions. We address the representation of the various types of information available in the EHR data, as well as exploring LLMs to characterize the clinical notes. We collect both information sources as the nodes of a graph neural network (GNN). Our model achieves an AUROC of 0.72 and a balanced accuracy of 66.7\%, highlighting the importance of combining the multimodal information.

Prediction of 30-day hospital readmission with clinical notes and EHR information

TL;DR

This paper addresses predicting 30-day hospital readmission by integrating structured EHR data and unstructured clinical notes. It proposes a scalable multimodal GraphSAGE-based GNN where each admission is a node, features come from demographic data, ICD embeddings for diagnoses/procedures, lab abnormalities, and BioClinicalBERT-derived note embeddings, with edges built via cosine similarity using FAISS. On the MIMIC-IV dataset, the model achieves AUROC of 0.727 and balanced accuracy of 0.667, outperforming logistic regression and MLP baselines, and illustrating the value of multimodal fusion. Compared to state-of-the-art multimodal readmission methods, the approach attains competitive performance on a larger-scale graph, while noting limitations in missing data handling and lack of temporal modeling.

Abstract

High hospital readmission rates are associated with significant costs and health risks for patients. Therefore, it is critical to develop predictive models that can support clinicians to determine whether or not a patient will return to the hospital in a relatively short period of time (e.g, 30-days). Nowadays, it is possible to collect both structured (electronic health records - EHR) and unstructured information (clinical notes) about a patient hospital event, all potentially containing relevant information for a predictive model. However, their integration is challenging. In this work we explore the combination of clinical notes and EHRs to predict 30-day hospital readmissions. We address the representation of the various types of information available in the EHR data, as well as exploring LLMs to characterize the clinical notes. We collect both information sources as the nodes of a graph neural network (GNN). Our model achieves an AUROC of 0.72 and a balanced accuracy of 66.7\%, highlighting the importance of combining the multimodal information.

Paper Structure

This paper contains 14 sections, 1 figure, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Background
  4. Neural Network architectures and Deep Learning Models
  5. Graph Neural Networks
  6. Proposed Approach
  7. Dataset
  8. Data Analysis and Pre-processing
  9. Readmission Model
  10. Experimental Setup
  11. Experimental Framework
  12. Model Training and Evaluation
  13. Results and Discusssion
  14. Conclusions

Figures (1)

  • Figure 1: Data analysis and pre-processing overview. Demographic features are encoded into a fixed length vector. ICD codes from the Diagnoses and Procedures tables were converted to text and sent to BioClinicalBERT to obtain the embeddings. Lab events were encoded into a fixed length vector. Clinical notes were sent to BioClinicalBERT to extract embeddings. The node features collect all the vectors from each group.