CART Machine Learning Peeling Pathways for Anxiety and Depression Severity Classification in Population Screening (Preprint)

BACKGROUND Monitoring anxiety and depression at the population level requires tools that are not only precise but also concise, thereby ensuring their feasibility for repeated use without overwhelming respondents. Although validated scales such as the GAD-7 and PHQ-9 are widely used, repeatedly administering them in population-based screening surveys, alongside other measures, may lead to respondent fatigue and limit adherence. The identification of multiple, abbreviated classification pathways that maintain high robustness and interpretability, may enhance the feasibility of large-scale monitoring in contexts of continuous surveillance and repeated assessments over time. OBJECTIVE This study had three objectives: (1) apply CART decision tree models based on a peeling approach to the GAD-7 and PHQ-9 scales in a national sample, (2) identify sets of abbreviated decision rules (defined as simplified classification pathways using fewer items than the original scale) for each level of anxiety and depression severity, and (3) combine these abbreviated decision rules into multiple classification pathways for use in epidemiological mental health surveillance. METHODS This cross-sectional study included 51,486 Brazilian adults from over 3,500 municipalities nationwide, in 2024 and 2025. CART models were estimated using the decision tree peeling procedure, based on progressively removing the root to explore alternative severity classification pathways. The set of rules was generated through an iterative process, which was determined by a stop criteria based on a composite score (CS) created from the average of balanced accuracy and F1–macro indexes, adopting a cutoff of higher than 0.6. The robustness of the findings was also examined through external validation in two independent samples, including a German-speaking, population-based sample (PHQ-9) and a German oncological, clinical sample (GAD-7). RESULTS The initial exploration of the GAD-7 identified 940 unique rules (CS = 0.823 ± 0.038). Of these, 134 final rules with three features were selected and distributed across severity levels (29 pathways for no symptoms, 32 for mild cases, 41 for moderate cases, and 32 for severe cases). For the PHQ-9, 1,266 unique rules were identified (CS = 0.800 ± 0.039), resulting in 509 final rules with four features. These were also stratified by severity: 95 pathways were identified for no symptoms, 144 for mild cases, 185 for moderate/moderately severe cases, and 85 for severe cases. No sociodemographic variable remained among the final rules. CONCLUSIONS The CART peeling approach identified multiple robust and abbreviated decision pathways for classifying anxiety and depression severity. These pathways used fewer than half of the original scale items and achieved performance comparable to total score models. The validation of these alternative pathways in independent samples provides a practical basis for epidemiological mental health surveillance, as it reduces respondent burden and supports more frequent and flexible monitoring of symptom changes over time.