Data-Driven Rule Mining and Representation of Temporal Patterns in Physiological Sensor Data.

Area of Science:

• Sensor data analytics
• Clinical informatics
• Pattern recognition

Background:

• Qualitative pattern mining is crucial for analyzing complex sensor data streams.
• Understanding temporal relationships in physiological data aids clinical decision-making.

Purpose of the Study:

• To extract and represent temporal relations of prototypical patterns in clinical data streams using rule mining.
• To develop a novel similarity method for validating mined temporal rules.
• To utilize natural language generation for explicit textual representation of temporal patterns.

Main Methods:

• Leveraging rule mining techniques on physiological time series data (heart rate, respiration rate, blood pressure).
• Introducing a novel similarity method for comparing rule sets to validate extracted temporal rules.
• Employing natural language generation to translate mined temporal rules into textual descriptions.

Main Results:

• Temporal rules were successfully mined from the MIMIC database, correlating with specific clinical conditions like respiratory failure, angina, and sepsis.
• A novel similarity method demonstrated effectiveness in rule set validation.
• The extracted rule sets were found to be distinct for different clinical conditions, highlighting their specificity.

Conclusions:

• The proposed data-driven approach effectively extracts and represents clinically relevant temporal patterns from sensor data.
• The method provides distinct, textually represented insights into various clinical conditions, aiding in their identification and differentiation.