Fusing Data Mining, Machine Learning and Traditional Statistics to Detect Biomarkers Associated with Depression

Area of Science:

• Epidemiology
• Biomarkers
• Data Mining

Background:

• Large-scale epidemiological datasets offer potential for identifying disease biomarkers.
• Machine learning and data mining techniques show promise in analyzing complex health data.
• This study addresses challenges in variable selection, including missing data and survey design.

Purpose of the Study:

• To identify key biomarkers associated with depression using a hybrid methodology.
• To illustrate a novel approach for variable selection in epidemiological research.
• To leverage machine learning and statistical modeling for biomarker discovery.

Main Methods:

• A three-step hybrid methodology combining multiple imputation, machine learning boosted regression, and logistic regression was employed.
• Data from the National Health and Nutrition Examination Study (2009-2010) was analyzed, including 67 biomarkers and various covariates.
• A final weighted multiple logistic regression model controlled for confounders and moderators.

Main Results:

• Machine learning initially identified 21 potential biomarkers for depression.
• A hybrid approach refined this to three significant biomarkers: red cell distribution width, serum glucose, and total bilirubin.
• Interactions between total bilirubin and ethnicity/smoking status were significant.

Conclusions:

• A hybrid methodology integrating machine learning and traditional statistics effectively identifies depression biomarkers.
• The identified biomarkers (red cell distribution width, serum glucose, total bilirubin) warrant further investigation.
• This approach accounts for missing data and complex survey designs in epidemiological studies.