Using interpretable decision trees to explore TB-HIV integration under data constraints: a proof-of-concept analysis

Strengthening tuberculosis (TB) and HIV service integration remains a major priority in resource-limited settings, where fragmented care contributes to delayed diagnosis, poor treatment continuity, and suboptimal patient outcomes. Evaluating implementation conditions associated with integrated care remains challenging because health system data are often simplified, inconsistently recorded, and insufficiently structured to capture the complexity of real-world service-delivery processes. This study aimed to demonstrate the application of an interpretable decision-tree framework to explore selected implementation and service-delivery conditions associated with TB-HIV integration using a highly constrained secondary dataset.

A quantitative exploratory proof-of-concept analysis was conducted using a simplified secondary dataset comprising 10 TB-HIV integration implementation cases. Funding, healthcare worker training, patient education, medication access, and infection control were included as binary implementation indicators and entered simultaneously into the decision-tree model. The Gini index was used descriptively to assess node impurity and decision structure. Correlation network analysis was performed to examine patterns of variable alignment, with Pearson correlation coefficients (equivalent to the phi coefficient for binary variables) interpreted solely as descriptive indicators of co-occurrence within the dataset. Given the extremely small sample size and simplified binary structure, all analyses were exploratory and illustrative rather than predictive, inferential, or causal.

Funding emerged as the primary discriminating variable in the decision tree, with all cases reporting increased funding classified as successful (6/6) and all cases without increased funding classified as unsuccessful (4/4). Although all implementation variables were evaluated by the algorithm, no additional variable provided sufficient information gain to generate further branching. Descriptive correlation analysis indicated alignment between funding and TB-HIV integration outcomes (r = 0.78), healthcare worker training (r = 0.76), patient education (r = 0.62), and infection control (r = 0.58). These coefficients are presented solely as descriptive indicators of variable alignment and should not be interpreted as measures of effect size, association strength, or statistical evidence.

This proof-of-concept study demonstrates how interpretable decision-tree approaches can be used to explore implementation conditions associated with TB-HIV integration under severe data constraints. The findings do not identify determinants of integration success but rather illustrate how selected implementation conditions were structured within a constrained analytical framework. The study highlights both the utility and limitations of interpretable analytical approaches when applied to small, dichotomized health systems datasets. Future research should employ larger, more granular, process-sensitive, and contextually rich datasets to support robust evaluation of TB-HIV integration in real-world settings.