Explanatory Modelling of Tuberculosis Treatment Outcomes: The Role of Community Engagement and Clinical Governance

1. Introduction

Drug-resistant tuberculosis (DR-TB) remains a major global public health challenge, disproportionately affecting low- and middle-income countries [1,2,3]. In high HIV-prevalence settings, especially in rural regions of sub-Saharan Africa, treatment outcomes are shaped by a complex interaction of biological resistance, patient vulnerabilities, and health system capacity. Despite advances in molecular diagnostics and improved treatment regimens, unfavourable outcomes, including treatment failure, loss to follow-up, and death, remain common among patients with DR-TB [4,5,6,7]. While pathogen-related factors, such as resistance phenotype, strongly influence prognosis, structural and programmatic elements also significantly affect treatment trajectories [8,9,10,11]. In resource-limited rural settings, support for adherence, care continuity, patient tracing, and integration of TB–HIV services are often inconsistent. These contextual challenges underscore the importance of clinical governance and community engagement as key structural determinants of treatment success [12]. Clinical governance frameworks emphasize accountability, monitoring, quality improvement, and systematic oversight of healthcare delivery, whereas community engagement focuses on building trust, encouraging adherence, and enabling locally responsive services. Increasingly, health systems are adopting integrated community engagement–clinical governance (CE–CG) approaches to enhance tuberculosis program performance [13,14,15,16]. These often include community health worker-led tracing, structured adherence counselling, community-based follow-up, routine program monitoring via dashboards, and coordinated TB–HIV care pathways. However, assessing the impact of governance and community engagement interventions remains challenging. These mechanisms operate through complex, multidimensional processes that are not easily captured by traditional causal models, especially when using routine program data [17,18,19,20]. Predictive modeling offers an opportunity to explore how demographic, clinical, and programmatic factors interact to influence treatment outcomes. Still, most existing predictive studies focus mainly on individual-level prognostic tools rather than explanatory models that account for structural and programmatic factors. In settings undergoing governance reform, modelling approaches can serve as both predictive tools and system-oriented analytical frameworks that help interpret program performance and structural coherence within evolving health systems [21,22,23,24]. Consequently, this study aimed to model and predict DR-TB treatment outcomes within a context of improved community engagement and clinical governance, evaluating their potential impact on adherence and treatment success through explanatory predictive models. Specifically, multivariable logistic regression and tree-based models were used to examine the associations between demographic, clinical, and programmatic factors and treatment success. By combining epidemiological modelling with clinical governance perspectives, this research seeks to deepen understanding of how structural programme environments influence DR-TB outcomes in rural, high-burden settings.

2. Methods

2.1. Study Design and Setting

This study employed a retrospective cohort design using routinely collected programme data from a drug-resistant tuberculosis (DR-TB) treatment programme in a rural South African setting. The analysis aimed to model treatment outcomes within the context of evolving programme structures, particularly the implementation of community engagement and clinical governance (CE–CG) strategies intended to strengthen adherence support and programme oversight.

2.2. Study Population and Data Source

The dataset comprised 694 patient records extracted from the tuberculosis programme database. Variables included demographic characteristics (age and gender), clinical factors (drug-resistance phenotype, treatment regimen at initiation, and comorbidity status), treatment initiation date, and final treatment outcome.

Treatment outcomes were categorised as favourable (cured or treatment completed) or unfavourable (lost to follow-up, treatment failure, death, transfer, or still on treatment at the time of analysis). For modelling purposes, drug-resistance phenotype was further classified into severe resistance (pre-extensively drug-resistant tuberculosis [pre-XDR] or extensively drug-resistant tuberculosis [XDR]) versus other resistance patterns.

2.3. Operationalisation of Community Engagement and Clinical Governance (CE–CG)

Community engagement and clinical governance were operationalised as a programme-level contextual indicator. Patients initiating treatment from 2024 onwards were classified as receiving care during the CE–CG implementation period.

This period corresponded with the documented introduction of structured governance and community-oriented support mechanisms, including community health worker–led patient tracing, structured adherence counselling, community-based follow-up, integrated TB–HIV coordination, and routine programme oversight supported by monitoring dashboards. The CE–CG indicator was conceptualised as a contextual explanatory variable representing programme environment rather than a direct causal exposure at the individual patient level.

2.4. Data Cleaning and Complete-Case Analysis

Continuous variables were assessed for plausibility and converted to numeric format where necessary. Age was centred prior to regression modelling to improve interpretability and reduce potential multicollinearity.

Complete-case analysis was applied for multivariable modelling. Records with missing values for any modelling variable (age, gender, regimen, treatment outcome, CE–CG indicator, resistance phenotype, or comorbidity status) were excluded. Comparisons between included and excluded patients were conducted to assess potential selection bias.

2.5. Descriptive Analysis

Descriptive statistics were computed for the full dataset before applying complete-case exclusion criteria. Continuous variables were summarised using means and standard deviations (SD), while categorical variables were reported as frequencies and percentages.

To evaluate potential selection bias arising from complete-case exclusion, comparisons between included and excluded patients were conducted using independent samples t-tests for continuous variables (age) and chi-square tests for categorical variables (gender and treatment success). In addition to statistical significance testing, effect sizes were calculated to assess practical significance, using Cohen’s d for continuous variables and Cramér’s V for categorical variables.

2.6. Multivariable Logistic Regression

A multivariable logistic regression model was fitted using complete-case data to examine associations between demographic, clinical, and programmatic variables and treatment success. The model included centred age, gender, CE–CG period indicator, regimen at treatment initiation, severe resistance phenotype, and comorbidity status as predictors.

Adjusted odds ratios (aOR) with corresponding 95% confidence intervals (CI) were estimated to quantify independent associations. Statistical significance was determined using two-sided p-values with a significance threshold of 0.05.

2.7. Comparative Predictive Modelling

To explore whether non-parametric modelling approaches improved discriminative performance, a decision tree classifier was fitted using the same predictor variables as the logistic regression model.

Both models were evaluated using a 70/30 train–test split to assess internal model performance. Discriminative ability was quantified using receiver operating characteristic (ROC) curve analysis and the corresponding area under the curve (AUC). Consistent with the study’s explanatory predictive framework, model performance was interpreted descriptively to evaluate internal coherence and structural insight rather than to develop a deployable prognostic prediction tool.

2.8. Model Diagnostics and Goodness-of-Fit

Logistic regression model diagnostics included assessing McFadden’s pseudo-R², likelihood-ratio (LR) testing comparing the full model with the intercept-only model, and evaluating multicollinearity using the Variance Inflation Factor (VIF).

McFadden’s pseudo-R² was interpreted within the context of routinely collected clinical programme data, where modest explanatory power is commonly observed. VIF values below 5 were considered indicative of acceptable levels of multicollinearity among predictors.

2.9. Statistical Software

All statistical analyses were conducted in Python version 3.11 within a Jupyter Notebook environment to ensure reproducibility. Data management and preprocessing were performed using pandas version 2.1 and NumPy version 1.26. Inferential statistical analyses were conducted using SciPy version 1.11 and statsmodels version 0.14, while predictive modelling procedures, including decision tree analysis, were implemented using scikit-learn. Figures, including ROC curves and diagnostic plots, were generated using matplotlib version 3.8. Statistical significance was defined a priori as a two-sided p-value < 0.05.

3. Results

3.1. Study Population and Data Completeness

A total of 694 patient records were identified in the tuberculosis programme dataset. After applying complete-case criteria for all modelling variables (age, gender, regimen at initiation, treatment outcome, CE–CG period indicator, resistance phenotype, and comorbidity status), 282 patients (40.6%) were retained for multivariable analysis. In comparison, 412 patients (59.4%) were excluded due to missing data in one or more variables. Missingness was primarily driven by incomplete documentation of resistance phenotype and comorbidity status, whereas core demographic variables such as age and gender were largely complete. The complete-case dataset was subsequently used for all multivariable modelling analyses.

3.2. Baseline Characteristics

The mean age of the overall cohort (N = 694) was 40.69 years (SD = 17.38). The study population was predominantly male (58.8%), while females comprised 38.9% of the cohort.

Overall treatment success, defined as cure or treatment completion, was observed in 59.9% of patients. Most patients (88.8%) were initiated on a short treatment regimen, whereas 6.8% received a long regimen. Severe resistance phenotypes, defined as pre-extensively drug-resistant (pre-XDR) or extensively drug-resistant tuberculosis (XDR-TB), were relatively uncommon (1.7%) but clinically significant. A detailed summary of baseline characteristics is presented in Table 1.

3.3. Comparison of Included and Excluded Patients

To assess potential selection bias introduced by the complete-case analysis, patients included in the modelling dataset (n = 282) were compared with those excluded due to missing data (n = 412). No statistically significant differences were observed between the two groups in terms of mean age (t = −1.04, p = 0.299), gender distribution (χ² = 1.962, p = 0.161), or treatment success proportions (χ² = 3.269, p = 0.071). Effect size estimates indicated negligible differences between groups. Cohen’s d for age was −0.081, while Cramér’s V values were 0.054 for gender and 0.069 for treatment success, suggesting minimal practical imbalance between included and excluded patients. Although 59.4% of records were excluded due to incomplete modelling variables, these findings indicate that the analytic sample remained broadly comparable to the full cohort across key demographic characteristics and treatment outcomes. The observed missingness appeared primarily related to incomplete documentation of specific clinical variables rather than systematic differences in core demographic or outcome measures. However, the possibility of residual bias due to unmeasured clinical or programmatic factors cannot be entirely ruled out.

3.4. Multivariable Explanatory Predictive Modelling

A multivariable logistic regression analysis was conducted using a complete-case dataset (n = 282) to explore the associations between demographic, clinical, and programmatic factors and treatment success. The predictor variables included centred age, gender, the treatment regimen at initiation, severe resistance phenotype (pre-XDR/XDR), comorbidity status, and an indicator for the community engagement-clinical governance (CE-CG) implementation period.

After adjusting for covariates, the analysis found that treatment initiation during the CE-CG implementation period (≥2024) was significantly associated with lower odds of treatment success (aOR = 0.443, 95% CI: 0.240–0.818, p = 0.009). The presence of severe resistance phenotypes (pre-XDR/XDR) was also strongly negatively associated with treatment success, approaching statistical significance (aOR = 0.303, 95% CI: 0.089–1.029, p = 0.056). However, other covariates—such as age, gender, treatment regimen at initiation, and comorbidity status—were not significantly associated with treatment outcomes in the adjusted model.

Table 2. Adjusted multivariable logistic regression model for treatment success.

Predictor	Adjusted OR	95% CI	p-value
Constant	1.293	0.395–4.236	0.671
Age (centred)	0.999	0.983–1.015	0.871
Gender	1.334	0.777–2.291	0.296
CE–CG period	0.443	0.240–0.818	0.009
Regimen at initiation	1.220	0.515–2.890	0.652
Severe resistance (Pre-XDR/XDR)	0.303	0.089–1.029	0.056
Any comorbidity	0.935	0.542–1.612	0.808

Table 2. Adjusted multivariable logistic regression model for treatment success.

Predictor	Adjusted OR	95% CI	p-value
Constant	1.293	0.395–4.236	0.671
Age (centred)	0.999	0.983–1.015	0.871
Gender	1.334	0.777–2.291	0.296
CE–CG period	0.443	0.240–0.818	0.009
Regimen at initiation	1.220	0.515–2.890	0.652
Severe resistance (Pre-XDR/XDR)	0.303	0.089–1.029	0.056
Any comorbidity	0.935	0.542–1.612	0.808

The constant represents the model's baseline reference level. Age was centred (age_c) by subtracting the sample mean to improve interpretability and reduce potential multicollinearity.

3.5. Comparative Model Performance

To evaluate whether non-parametric modelling approaches improved predictive discrimination, a decision tree classifier was developed alongside the logistic regression model using the same set of predictor variables. Both models were evaluated using a 70/30 train–test split to assess internal predictive performance.

The logistic regression model demonstrated limited discriminative ability, with an area under the receiver operating characteristic curve (AUC) ranging from 0.523 to 0.548, indicating performance only marginally better than chance. In contrast, the decision tree classifier showed modestly improved discrimination (AUC = 0.626), suggesting greater capacity to capture nonlinear relationships and hierarchical interactions among predictors. Despite this relative improvement, overall predictive performance remained modest across both modelling approaches, indicating that the available variables have limited explanatory capacity for predicting treatment success in the current dataset (Figure 1).

3.6. Comparative Model Performance: Logistic Regression vs Tree-Based Models

To evaluate whether non-parametric modelling improved discriminative performance, a decision tree classifier was fitted alongside the multivariable logistic regression model using identical predictor variables. Both models were assessed using a 70/30 train–test split and receiver operating characteristic (ROC) curve analysis. The logistic regression model demonstrated limited discrimination (AUC = 0.523–0.548), indicating modest ability to distinguish between favorable and unfavorable treatment outcomes. The decision tree model showed improved discrimination (AUC = 0.626), suggesting that hierarchical partitioning may better capture nonlinear relationships and interaction structures within the dataset. Nevertheless, overall performance remained moderate across both approaches.

These findings reflect the complementary strengths of the modelling strategies. Logistic regression provides interpretable effect estimates (adjusted odds ratios with confidence intervals), supporting inferential clarity and governance-relevant interpretation. In contrast, tree-based models identify structural decision pathways and potential threshold effects without imposing linear assumptions. However, neither approach achieved sufficient discrimination to support individual-level prognostic application. The modest AUC values likely reflect limitations of the available predictor set, including the absence of longitudinal adherence markers, socioeconomic instability measures, and quantified governance-process indicators. Consistent with the study’s explanatory predictive framework, model performance is interpreted descriptively to assess internal coherence and structural insight rather than to develop externally generalisable prediction tools.

Figure 2. Receiver Operating Characteristic Curve for the Adjusted Explanatory Predictive Model of Treatment Success.

Figure 2. Receiver Operating Characteristic Curve for the Adjusted Explanatory Predictive Model of Treatment Success.

3.7. Model Diagnostics and Goodness-of-Fit

The multivariable logistic regression model, fitted on the complete-case dataset (n = 282), included centred age, gender, CE–CG implementation period, regimen at treatment initiation, severe resistance phenotype, and comorbidity status as predictors of treatment success. Model diagnostics showed modest explanatory power with acceptable internal consistency. McFadden’s pseudo-R² was 0.029, indicating a limited explanatory contribution of the predictors compared to the intercept-only model, which is common when analysing routinely collected clinical program data. Likelihood-ratio testing comparing the full model with the intercept-only model gave χ²(6) = 10.70, p = 0.098, showing a small, non-significant improvement in model fit. Assessment of multicollinearity revealed Variance Inflation Factor (VIF) values near 1.0 for all predictors, suggesting no problematic collinearity among the independent variables. Overall, these diagnostics support the model's structural validity while confirming its limited explanatory power. In line with the study’s explanatory modelling framework, the regression model is primarily used as a descriptive tool for examining programmatic and structural factors affecting treatment outcomes rather than as a predictor for individual clinical decisions. A summary of model diagnostics is shown in Table 3.

4. Discussion

This study used an explanatory predictive modelling approach to explore how community engagement and clinical governance (CE–CG), implemented through a program-level indicator during the implementation period, operate within the broader clinical and epidemiological context of drug-resistant tuberculosis (DR-TB) management. Instead of creating a patient-level prognostic tool, the modelling strategy was purposely designed as a governance-sensitive analytical framework to investigate structural consistency and generate insights into program performance. Explanatory modelling approaches are increasingly used to examine how interactions between health system processes, governance reforms, and clinical pathways influence programme outcomes, especially in resource-limited settings [22,24]. By adopting this systems-focused perspective, modelling can reveal not only statistical relationships but also how governance structures interact with demographic and clinical factors to influence treatment trajectories, aligning with emerging trends in health systems strengthening and implementation research [21,23]. Three main findings emerged from this analysis. First, despite excluding a large portion of records due to missing variables, the remaining sample was demographically like the full cohort, suggesting minimal selection bias. Second, severe resistance phenotypes showed a strong negative relationship with treatment success, highlighting the significant biological role of pathogen-level factors in DR-TB outcomes. Similar links between advanced resistance profiles and poor treatment results have been reported in other DR-TB cohorts, including studies in South Africa and India [25,26]. Third, the CE–CG implementation period indicator remained independently associated with treatment outcomes after adjusting for demographic and clinical factors. However, interpreting this connection requires careful consideration of the context, given the complexity of concurrent programmatic changes and possible confounding factors. Variations in treatment outcomes during periods of health system reform, care decentralization, and evolving patient management strategies have been documented in prior studies assessing TB program performance [27,28].

The observed association between the CE–CG implementation period and lower treatment success should therefore not be interpreted as evidence of a negative governance effect. In this study, CE–CG was operationalised as a contextual programme indicator rather than a direct causal exposure. The implementation period likely coincided with broader operational shifts, including increased clinical complexity, evolving case-mix patterns, and a greater prevalence of advanced resistance phenotypes. Similar dynamics have been observed during periods of programme transformation in DR-TB care, where governance strengthening occurs alongside expanding access to treatment for more complex or previously underserved patient populations [10,11,12,13]. Evidence from health systems research suggests that improvements in governance and accountability often occur simultaneously with expanded service coverage, during which outcome indicators may temporarily appear less favourable as programmes begin identifying and managing more complex cases [29,30]. Within TB programmes specifically, intensified surveillance, decentralised treatment expansion, and integrated TB–HIV care can increase detection of high-risk cases and improve classification of unfavourable outcomes, potentially creating the appearance of reduced treatment success despite underlying improvements in programme responsiveness and care quality [31,32]. The modest discriminatory performance of both logistic regression and tree-based models provides an additional theoretical insight relevant to CE–clinical governance frameworks. Governance interventions typically operate through complex, relational, and process-driven mechanisms that are difficult to capture using routinely collected administrative data. Activities such as community tracing, adherence counselling, integrated TB–HIV service coordination, and dashboard-driven programme oversight function as organisational and relational processes that influence care pathways but are rarely measured directly at the individual patient level [33,34,35,36]. In high-burden TB settings, governance mechanisms often influence outcomes by improving accountability, coordination of care, and patient–provider relationships dimensions that are difficult to quantify using conventional programme indicators [33,34,37]. Consequently, predictive discrimination may appear modest not because governance interventions lack importance, but because their relational and organisational mechanisms remain poorly represented within routine health information systems [35,36]. Within this explanatory modelling framework, the results thus emphasize the structural relationships within the treatment pathway rather than acting solely as predictive tools. The analysis confirms the primary impact of biological resistance on treatment outcomes and highlights the limitations of routine program datasets in capturing governance quality and process-level care mechanisms within TB systems [38,39,40,41]. Improving routine data systems by adding measurable governance indicators such as adherence counselling frequency, clinician review intervals, contact tracing activities, decentralized follow-up coverage, and TB–HIV service integration metrics could significantly enhance the depth of explanation and predictive ability of future programme evaluations [38,41].