Robust clinical marker identification for diabetic kidney disease with ensemble feature selection

Area of Science:

• Medical Informatics
• Nephrology
• Machine Learning

Background:

• Diabetic kidney disease (DKD) is a major complication of diabetes, causing significant illness and death.
• Identifying risk factors for DKD is crucial for early intervention and improved patient outcomes.

Purpose of the Study:

• To develop and validate an ensemble feature selection approach for identifying robust DKD risk factors using electronic medical records (EMRs).
• To discover novel predictive features for DKD onset that may be overlooked by current models.

Main Methods:

• A retrospective cohort of 15,645 adult patients with type 2 diabetes was analyzed.
• Compared 3 machine-learning-based embedded feature selection methods with 6 feature ensemble techniques.
• Utilized gradient boosting machine (GBM) with weighted mean rank ensemble for feature selection.

Main Results:

• The GBM ensemble model achieved an AUC of 0.82 (internal) and 0.71 (external validation).
• Identified 440 robust and predictive features from over 84,000 clinical features.
• Key features included labs, vital signs, medications, orders, and diagnoses.

Conclusions:

• The proposed ensemble feature selection framework effectively identifies a robust set of DKD risk factors from EMR data.
• This approach facilitates unbiased knowledge discovery for DKD, potentially improving prediction models.