Discretization of continuous features in clinical datasets.

Area of Science:

• Health Informatics
• Machine Learning
• Data Science

Background:

• Electronic medical records (EMRs) provide vast clinical data for secondary health information uses.
• Discretization is a crucial preprocessing step for machine learning classification of EMR data.

Purpose of the Study:

• To evaluate the performance of six distinct discretization strategies (supervised and unsupervised) on EMR data.
• To compare the classification accuracy of discretized EMR data against original continuous data.

Main Methods:

• Classified laboratory (arterial blood gas) and physiologic (cardiac output) data from intensive care unit patients.
• Applied decision trees and Naïve Bayes classifiers to data partitioned using two supervised and four unsupervised discretization methods.

Main Results:

• Supervised discretization methods demonstrated superior accuracy and consistency compared to unsupervised approaches.
• Among unsupervised methods, equal frequency and k-means partitioning performed well; equal width partitioning was less accurate.
• Supervised methods resulted in larger decision trees.

Conclusions:

• No single discretization method is universally optimal for all EMR data.
• Discretization performance is influenced by class labels and the number of intervals chosen.
• Supervised methods offer high accuracy for specific tasks, while unsupervised methods provide adaptable discretized data for multiple uses.