Combining Structured and Free-text Electronic Medical Record Data for Real-time Clinical Decision Support

Emilia Apostolova, Tony Wang, Tim Tschampel, Ioannis Koutroulis, Tom VelezACL 2019

TL;DR

This paper introduces "Patient Context Vectors" (PCVs) to fuse structured (like vital signs and ICD codes) and unstructured (clinical notes) Electronic Medical Record (EMR) data for real-time clinical decision support. By representing patient history and conditions in a unified, low-dimensional vector, they significantly improve the prediction of rare, life-threatening conditions like ARDS.

What Problem Does This Paper Address?

Early identification of critical conditions like Sepsis, Septic Shock, and Acute Respiratory Distress Syndrome (ARDS) is crucial in critical care. However, these conditions are often rare, complex, and their risk factors are scattered across a patient's EMR, including structured data (vitals, labs) and unstructured free-text notes (nursing notes, radiology reports). Traditional alert systems often rely solely on structured data, missing vital context from clinical notes. Furthermore, the variability in charting and data availability in real-time poses a significant challenge for building effective clinical decision support (CDS) systems. This paper aims to bridge this gap by creating a comprehensive patient representation that leverages all available EMR data for timely risk prediction.

Methods

The core innovation is the Patient Context Vector (PCV). PCVs are generated in two ways:

  1. From ICD Codes: Existing ICD (International Classification of Diseases) codes are converted into low-dimensional embeddings (similar to word embeddings). These embeddings are then averaged to create a PCV that summarizes the patient's diagnostic history.
  2. From Clinical Texts: A Convolutional Neural Network (CNN) is trained to predict a PCV directly from free-text nursing notes. This captures information not necessarily captured by ICD codes.

These two PCVs (one from ICD codes, one from notes) can be used independently or combined. For prediction, these PCVs are then simply appended to the existing structured data (vital signs, lab results) to train predictive models. The authors experimented with Gradient Boosting Machine (GBM) and Distributed Random Forest (DRF) models for ARDS prediction.

Datasets

The study utilized the MIMIC-III Intensive Care Unit (ICU) database. This publicly available dataset contains deidentified health-related data for over 40,000 patients, including over 2 million free-text clinical notes and over 650,000 diagnosis codes. The data spans multiple ICU types and includes vital signs, lab results, diagnosis codes, and free-text notes. No new benchmark datasets were introduced; the focus was on leveraging an existing, rich EMR dataset.

Key Insights and Ideas

  • Unifying Structured and Unstructured Data: The PCV approach elegantly bridges the gap between structured and unstructured EMR data, creating a single, interpretable representation of a patient's condition.
  • ICD Codes as a Rich Source: Even though ICD codes are often seen as simple labels, their embeddings capture complex relationships and can serve as a powerful summary of a patient's medical history.
  • Clinical Notes as a Real-time Goldmine: Nursing notes, in particular, contain a wealth of real-time information that might not be immediately coded, making them invaluable for timely prediction.
  • Real-time Feasibility: The PCVs perform well even when generated from only the first half of a patient's ICD codes and notes, suggesting their suitability for real-time CDS where full data might not yet be available.
  • Embedding Importance: The variable importance plots show that certain dimensions of the PCVs are as predictive as key vital signs, highlighting their significant contribution.

Key Results

  • Combining structured data with PCVs derived from either ICD codes or clinical notes significantly improved ARDS prediction performance compared to using structured data alone.
  • The best results were achieved when combining structured data with PCVs derived from both the first half of ICD codes and the first half of nursing notes.
  • For example, using the GBM model, the "Baseline + Notes Patient Context Vector" achieved an AUC of 93.59, Recall of 56.35, and F1-score of 61.01, outperforming the baseline (AUC 90.42, Recall 41.76, F1 51.68). The combined approach ("Baseline + first half of notes/ICD") further boosted performance to AUC 93.59, Recall 56.35, and F1 61.01. (Note: The table shows identical results for "Baseline + Notes Patient Context Vector" and "Baseline + first half of notes/ICD" for GBM, which might be a typo or indicate very similar performance).
  • The DRF model also showed similar improvements, with the combined approach achieving an AUC of 92.61, Recall of 57.02, and F1-score of 58.98.

Limitations and Future Work

  • Real-time Availability: While the study showed good performance with partial data, the exact real-time availability of ICD codes and notes can still vary, impacting immediate prediction.
  • Model Complexity: The CNN model for predicting PCVs from notes, while effective, adds complexity to the overall system.
  • Generalizability: The study focused on ARDS. Future work could explore the generalizability of PCVs to predicting other rare and complex conditions.
  • Combining PCVs: The paper suggests that combining the PCVs from ICD codes and notes might yield even better results, which could be a direction for future research.
  • Interpretability of Embeddings: While variable importance is shown, further work could delve into the specific clinical meaning of individual embedding dimensions.

Summary updated

BibTeX
@inproceedings{apostolova2019combining,
  author = {Emilia Apostolova and Tony Wang and Tim Tschampel and Ioannis Koutroulis and Tom Velez},
  title = {{Combining Structured and Free-text Electronic Medical Record Data for Real-time Clinical Decision Support}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure