Decision Trees Reveal Funding as the Decisive Factor in TB-HIV Integration Under Severe Data Constraints

Tuberculosis remains one of the world’s deadliest infectious diseases, killing an estimated 1.23 million people in 2024 out of 10.7 million who fell ill, and people living with HIV face a substantially elevated risk of developing active TB, particularly in sub-Saharan Africa. In resource-limited settings, fragmented TB and HIV services contribute to delayed diagnosis, poor treatment continuity, and suboptimal outcomes, making the integration of TB-HIV care a major public health priority. Yet evaluating whether integration programmes actually succeed has long been hampered by a stubborn obstacle: health system data in these settings are often simplified, inconsistently recorded, and too coarse to capture the complexity of real-world service delivery. A new proof-of-concept study published in Frontiers in Tuberculosis now demonstrates how an interpretable machine-learning approach can extract transparent insights from exactly such constrained data.

The research, led by Ntandazo Dlatu of the Walter Sisulu Institute for Clinical Governance at Walter Sisulu University in Mthatha, South Africa, together with colleagues from the university’s Department of Laboratory Medicine and Pathology, applied a decision-tree framework to a secondary dataset of just ten TB-HIV integration implementation cases. Rather than attempting to build a predictive model or identify causal determinants of success, the team set out to show how interpretable analytical methods behave when applied to small, dichotomized health systems datasets—a scenario that is common in resource-constrained settings but rarely examined methodologically. The study was conceived explicitly as a methodological demonstration intended to support exploration and hypothesis generation rather than inference.

The dataset drew on a previously published study by Dlatu and colleagues that examined models of TB-HIV service integration in resource-constrained healthcare settings in South Africa’s Eastern Cape. Each of the ten cases represented a programmatic or service-delivery instance in which TB and HIV services were operationally coordinated within a specific healthcare environment, rather than an individual patient or facility. Five governance-relevant implementation conditions were included as binary indicators: funding, healthcare worker training, patient education, medication access, and infection control. Funding was interpreted as an indicator of resource availability, training as a marker of workforce preparedness, patient education as patient-facing support, medication access as service continuity, and infection control as clinical and organizational safety practice. The outcome variable classified each case as successful or unsuccessful integration according to the definitions used in the original study.

Preprocessing was deliberately minimal. Because every variable in the dataset was already binary, no scaling, normalization, or transformation was required; the team performed data cleaning, binary encoding, and consistency checks while preserving the original coding structure to maintain interpretability. All five implementation variables were then entered simultaneously into a decision-tree model built in R version 4.2.3. Entropy-based splitting identified the variables providing the greatest information gain at each node, while the Gini index—a measure of node impurity ranging from 0 for perfectly pure nodes to 0.5 for maximally mixed binary splits—was used descriptively to assess decision structure. As a limited internal consistency check, the researchers also performed leave-one-out cross-validation, iteratively training the model on nine observations and testing on the tenth, though they emphasized that this procedure could not serve as a measure of predictive performance given the dataset’s size.

The results were striking in their simplicity. Funding emerged as the primary discriminating variable, generating the greatest reduction in node impurity and becoming the sole branching node in the final tree. The root node, labelled

The striking simplicity of the resulting tree—one variable, one split, two perfectly separated groups of cases—illustrates both the appeal and the peril of interpretable machine learning under data constraints. With only ten observations and five binary predictors, the algorithm exhausts the available structure almost immediately: once funding separates the cases, no remaining variable offers enough additional information gain to justify further branching. This is not because training, patient education, medication access, or infection control are unimportant in real-world implementation, but because the dichotomized, small-sample format leaves little residual variation for the algorithm to exploit. The authors are careful to frame the tree as a descriptive map of how the variables were arranged in this particular dataset, not as a model of how integration succeeds in general.

The correlation network analysis reinforces this cautionary stance. The reported coefficients—0.78 between funding and integration outcome, 0.76 for healthcare worker training, 0.62 for patient education, and 0.58 for infection control—are, for binary variables, mathematically equivalent to phi coefficients, which describe co-occurrence patterns within the observed data. With so few cases, even visually impressive alignment can arise from chance arrangements of a handful of observations, and the authors explicitly warn against reading these values as effect sizes or statistical evidence. The gradient across the variables, however, is methodologically interesting: it suggests that in this dataset, workforce training tracked the funding signal more closely than patient-facing or safety-related conditions did, a pattern that could seed hypotheses for future, better-powered studies.

The choice of the Gini index as a descriptive lens deserves attention from readers unfamiliar with decision-tree mechanics. Node impurity quantifies how mixed the outcome classes are at any point in the tree, and splitting criteria seek the variable that most reduces this mixture. In larger datasets, competing variables often produce near-tied impurity reductions, and analysts must grapple with instability in variable selection. Here, the margin between funding and all other candidates was decisive, which is precisely what one would expect when a single binary variable happens to align perfectly with a binary outcome in a tiny sample. The study thus doubles as a teaching case: it shows what an idealized, noise-free separation looks like in tree form, and by extension how suspicious such perfection should be when sample sizes are this small.

The clinical governance framing adds conceptual depth to what might otherwise appear to be a purely technical exercise. By treating funding, training, patient education, medication access, and infection control as operational proxies for governance dimensions—accountability, workforce competence, patient engagement, service continuity, and safety—the authors position the analysis within a broader conversation about how health systems quality can be glimpsed through administrative data. Routine health information systems in resource-limited settings rarely capture implementation fidelity, service coordination, or contextual nuance directly; they record coarse indicators instead. The study’s contribution is to show that even such coarse indicators can be organized transparently, provided the analytical method itself remains legible to programme managers and policymakers rather than hidden inside opaque algorithms.

This emphasis on interpretability situates the work within a growing debate about the role of machine learning in public health evaluation. Complex models may offer predictive power but often at the cost of transparency, which is a particular liability in health systems research where findings must inform resource allocation decisions and be trusted by implementers on the ground. Decision trees occupy the opposite end of the spectrum: their logic can be inspected node by node, their assumptions are minimal, and their outputs translate naturally into actionable questions—here, essentially, whether resources increased. The trade-off, as this study candidly demonstrates, is that interpretability does not compensate for data poverty; it merely makes the limitations of the data visible rather than obscuring them behind model complexity.

The burden of HIV-associated TB in sub-Saharan Africa gives the methodological stakes real urgency. People living with HIV face substantially elevated risks of developing active TB, and diagnosing TB in immunocompromised patients remains difficult because of atypical presentations and reduced sensitivity of conventional diagnostic tools. While molecular diagnostics such as Xpert MTB/RIF Ultra have improved detection, their rollout remains uneven where infrastructure, organization, and financing fall short. Integrated service delivery—coordinating prevention, diagnosis, treatment, and follow-up across TB and HIV programmes—is widely promoted as a way to reduce missed opportunities and strengthen continuity of care, yet the evidence base for which implementation conditions matter most remains thin, partly because the data needed to build that evidence are themselves fragmented.

The authors’ call for future research points toward larger, more granular, process-sensitive, and contextually rich datasets. Such datasets would allow the binary indicators used here to be replaced or supplemented with measures capturing the degree of implementation, the sequencing of integration activities, and the organizational contexts in which services are coordinated. They would also permit genuine inferential and predictive modelling, with adequate statistical power to estimate associations, test interactions among implementation conditions, and validate findings across settings. Until then, the study stands as a carefully bounded demonstration: interpretable decision trees can illuminate the structure of constrained health systems data, but the insights they yield are hypotheses to be tested, not conclusions to be acted upon.

For practitioners and researchers working in similar resource-constrained environments, the practical lesson may be twofold. First, transparent exploratory methods can make productive use of the small, simplified datasets that already exist, extracting descriptive structure that supports planning conversations and priority-setting. Second, the very transparency of these methods exposes how much depends on the quality and richness of the underlying data, strengthening the case for investment in health information systems that can record implementation processes with the granularity that rigorous evaluation ultimately requires.

Subject of Research: Interpretable decision-tree analysis of TB-HIV service integration using a small binary secondary dataset

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

Article References: Using interpretable decision trees to explore TB-HIV integration under data constraints: a proof-of-concept analysis. (n.d.). https://doi.org/10.3389/ftubr.2026.1853414

Image Credits: AI Generated

DOI: 10.3389/ftubr.2026.1853414

Keywords: tuberculosis, HIV, health systems, machine learning, decision trees, implementation science, South Africa, integrated care, interpretable, decision, trees, explore