Risk stratification of ICU patients using topic models inferred from unstructured progress notes.

Area of Science:

• Medical Informatics
• Computational Linguistics
• Clinical Decision Support

Background:

• Accurate prediction of hospital mortality in Intensive Care Units (ICUs) is crucial for patient management.
• Existing risk stratification models often rely solely on structured physiologic data.
• Unstructured clinical notes contain valuable information that is underutilized in predictive models.

Purpose of the Study:

• To develop and evaluate a novel approach for ICU patient risk stratification.
• To integrate the topic structure derived from clinical concepts in nursing notes with physiologic data for improved hospital mortality prediction.
• To assess the performance enhancement of the SAPS-I algorithm by incorporating extracted clinical topics.

Main Methods:

• Utilized Hierarchical Dirichlet Processes (HDP), a non-parametric topic modeling technique, to identify co-occurring UMLS clinical concepts in nursing notes.
• Extracted clinical "topics" from unstructured nursing notes of 14,739 adult ICU patients from the MIMIC II database.
• Combined the learned topic structure with physiologic data (SAPS-I) for hospital mortality prediction.

Main Results:

• The learned topic structure from the first 24-hour ICU nursing notes significantly improved hospital mortality prediction performance.
• The Area Under the Curve (AUC) for predicting mortality using physiologic data and nursing text notes was 0.82.
• Physiologic data alone with the SAPS-I algorithm achieved an AUC of 0.72, indicating a substantial improvement with the novel approach.

Conclusions:

• Integrating learned clinical topics from nursing notes with physiologic data enhances ICU patient risk stratification.
• The novel approach significantly improves the performance of the baseline SAPS-I algorithm for hospital mortality prediction.
• This method offers a promising avenue for leveraging unstructured clinical data to improve patient outcomes.