Feature Extraction for Heroin-Use Classification Using Imbalanced Random Forest Methods.

Area of Science:

• Data Science
• Public Health
• Computational Statistics

Background:

• The National Survey on Drug Use and Health (NSDUH) dataset contains numerous responses and features.
• Identifying key predictors for heroin use classification is crucial for public health interventions.
• Existing methods for applying random forest (RF) to imbalanced medical data are not well-defined.

Purpose of the Study:

• To identify important features within the NSDUH dataset for classifying heroin use.
• To apply and compare different RF classification techniques for imbalanced medical datasets.
• To establish a method for feature extraction from imbalanced datasets with numerous predictors.

Main Methods:

• Three distinct RF classification techniques were applied to the 2016 NSDUH data.
• Models were compared using the area under the precision-recall curve (AUPRC) to determine the best performing method.
• Variable importance scores (VIS) were analyzed for stability and to identify key features influencing heroin use classification.

Main Results:

• RF with random oversampling achieved the best performance with an AUPRC of 0.5437.
• Features related to other drug use were the most important category (average z-scored VIS = 1.66).
• Cocaine (z-scored VIS = 11.05) and crack usage (6.51) were the most significant individual predictors; marijuana use before age 18 was also notable (3.11).

Conclusions:

• Random forest with random oversampling is an effective method for feature extraction from imbalanced medical datasets.
• Other drug use, particularly cocaine and crack, alongside early marijuana initiation, are strong indicators of heroin use.
• This study provides a validated methodology for utilizing RF in analyzing large, imbalanced health survey data.