JMIR Publications

ABSTRACT

Background:

Scalable, cost-effective tools are essential for mental health screening. Machine learning offers innovative ways to enhance accuracy and optimize existing instruments for assessing anxiety and depression severity.

Objective:

This study applied decision tree models to refine mental health screening by identifying the most predictive components of anxiety and depression severity in Brazil.

Methods:

Data from 20,585 adults across all 27 Brazilian states and 3,000+ cities, covering urban and rural areas, were analyzed. Mental health symptoms were assessed using the GAD-7 and PHQ-9 scales. Decision trees, built using the CART algorithm, identified the most predictive items. Sociodemographic variables were also tested as predictors.

Results:

The models derived concise decision rules using two (GAD-7) or three (PHQ-9) items, achieving high accuracy for Minimum/Mild and Severe severity levels (86.1%/85.1% for GAD-7, 81.7%/78.8% for PHQ-9) with AUC > 0.900. Moderate severity had lower reliability (51.5% GAD-7; 66.8% PHQ-9; AUC = 0.728/0.776). Sociodemographic factors had minimal predictive value. Findings suggest a balance between efficiency and precision.c

Conclusions:

Machine learning facilitates streamlined mental health screening by optimizing severity classification, improving scalability while maintaining reliability. These results reinforce the potential of decision-tree models to enhance screening efficiency in public health applications.