Title : Does greater model complexity improve prediction of sustained viral suppression? Comparing CART and random forest among women living with HIV in Brazil
Abstract:
Background: Sustained viral suppression is a key goal of HIV care, yet clinical, psychosocial, and programmatic vulnerabilities may interact in complex ways among women living with HIV. Machine learning approaches can help identify individuals at risk of unfavorable treatment outcomes. However, whether greater algorithmic complexity translates into improved predictive performance remains uncertain, particularly when interpretability is essential for healthcare decision-making.
Objective: To compare the predictive performance and interpretability of Classification and Regression Trees (CART) and Random Forest models for identifying cisgender women living with HIV at risk of failing to achieve sustained viral suppression.
Methods: This observational study used secondary data from 236 cisgender women living with HIV diagnosed between 2014 and 2024 and receiving care at a specialized HIV service in Minas Gerais, Brazil. CART and Random Forest models were developed using clinical, individual, social, and programmatic vulnerability variables from the same study population. Previously obtained performance estimates were compared descriptively, including accuracy, sensitivity, specificity, positive predictive value, F1-score, and area under the receiver operating
characteristic curve (AUC). Model interpretability and the clinical relevance of identified predictors were also examined.
Results: CART achieved an AUC of 0.820, accuracy of 78.2%, sensitivity of 68.2%, specificity of 88.8%, positive predictive value of 86.6%, and F1-score of 0.763. Random Forest achieved an AUC of 0.707, accuracy of 63.4%, sensitivity of 59.5%, specificity of 67.6%, positive predictive value of 66.7%, and F1-score of 0.629. CART demonstrated numerically higher values across all reported performance metrics. CD4-related measures and age at HIV diagnosis emerged as relevant predictors, alongside psychosocial and programmatic vulnerability factors. CART provided explicit decision rules that facilitated the interpretation of vulnerability profiles, whereas Random Forest offered a complementary assessment of variable importance.
Conclusion: CART showed more favorable reported predictive performance than Random Forest in this study population, suggesting that greater algorithmic complexity does not necessarily improve prediction of sustained viral suppression outcomes. The interpretability of CART may offer advantages for translating predictive findings into information relevant to multidisciplinary HIV care. These findings highlight the importance of balancing predictive performance and interpretability when selecting machine learning approaches, while underscoring the need for further comparative validation before clinical implementation. Keywords: Women Living with HIV; Sustained Viral Suppression; Machine Learning; CART; Random Forest.
