A Visualization of Evolving Clinical Sentiment Using Vector Representations of Clinical Notes.

Area of Science:

• Computational linguistics
• Medical informatics
• Natural Language Processing

Background:

• Clinical notes contain valuable information on patient care and outcomes.
• Analyzing sentiment and language evolution in clinical text can provide insights into healthcare trends.
• Previous studies have not comprehensively visualized sentiment shifts across demographic and outcome variables.

Purpose of the Study:

• To visualize the evolution of clinical language and sentiment in intensive care unit (ICU) notes.
• To examine how language and sentiment differ across patient categories: time in hospital, age, mortality, gender, and race.
• To identify patterns in clinical text associated with patient outcomes and demographics.

Main Methods:

• Utilized seven years of unstructured free text notes from the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC) database.
• Partitioned text data by demographic and outcome categories.
• Generated high-dimensional vector space representations and visualized them using t-Distributed Stochastic Neighbor Embedding (tSNE) and Principal Component Analysis (PCA).
• Inferred sentiment by analyzing the proximity of keywords to other terms within the vector space.

Main Results:

• Observed distinct differences in clinical note sentiment related to time, outcome, and demographics.
• Found a decrease in cluster homogeneity and complexity over time for patients with poor outcomes.
• Detected greater positive sentiment in notes concerning females, unmarried patients, and patients of African ethnicity.

Conclusions:

• Clinical language and sentiment are dynamic and influenced by patient characteristics and outcomes.
• The study provides novel visualizations of sentiment evolution in critical care notes.
• Findings highlight potential biases or differing communication patterns related to patient demographics and prognosis.