CART machine learning peeling pathways for anxiety and depression severity classification in population screening

Background Regular population-level monitoring of anxiety and depression requires brief assessment strategies that reduce respondent burden while retaining classification performance. Objective This study aimed to: (1) apply CART decision tree peeling to the GAD-7 and PHQ-9 scales in a nationwide sample; (2) derive abbreviated decision rules that reproduce the severity categories defined by the full scales; and (3) examine multiple classification pathways for epidemiological mental health surveillance. Methods This cross-sectional study analyzed data from 51,486 Brazilian adults recruited online through social media across more than 3500 municipalities between 2024 and 2025. An iterative CART root-removal (peeling) procedure identified alternative abbreviated pathways that reproduced the severity categories defined by the full GAD-7 and PHQ-9 scales. Retained rules were evaluated in independent validation datasets, and a retrospective item response theory-based computerized adaptive testing simulation provided a benchmark. Results GAD-7 exploration yielded 940 rules (CS = 0.823, SD = 0.038), including 134 retained three-node rules across severity levels (29 none, 32 mild, 41 moderate, and 32 severe). For the PHQ-9, 1266 rules were identified (CS = 0.800, SD = 0.039), including 509 four-node rules (95 none, 144 mild, 185 moderate/moderately severe, and 85 severe). Conclusions CART peeling identified multiple abbreviated pathways that reproduced GAD-7 and PHQ-9 severity categories. Using fewer than half of the original items, these pathways retained high classification performance relative to the full-scale categories. External validation supported their performance across independent datasets, indicating potential utility for more frequent and flexible epidemiological monitoring.