Using the Diagnostic Odds Ratio to Select Patterns to Build an Interpretable Pattern-Based Classifier in a Clinical...

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Clinical Data Mining

Background:

• Intensive care burn units (ICBUs) generate extensive patient data over time.
• Multivariate sequential patterns can model patient evolution for survival prediction.
• Standard pattern discovery yields many irrelevant patterns for classification.

Purpose of the Study:

• To employ the diagnostic odds ratio (DOR) for selecting multivariate sequential patterns in clinical classification.
• To evaluate DOR-based pattern selection against frequency-based methods like Jumping Emerging Patterns.

Main Methods:

• Utilized temporal data from 465 ICBU patients over 5 days.
• Applied multivariate sequential patterns with two discretization methods.
• Compared four DOR application strategies for pattern selection, including thresholding and non-overlapping confidence intervals (CIs).

Main Results:

• Discretization significantly impacts classification accuracy and model interpretability.
• Combining threshold and non-overlapping CIs yielded the fewest, smallest patterns with acceptable accuracy.
• The best trade-off model used a JRIP classifier with 5 patterns, achieving 56.32% specificity and 0.767 AUC.

Conclusions:

• Sequential patterns effectively leverage temporal ICBU data for patient survival classification.
• Diagnostic odds ratio (DOR) is a suitable metric for selecting discriminative and interpretable patterns.
• The proposed DOR-based method enhances the utility of sequential patterns in clinical settings.