Introduction

Tuberculosis (TB), caused by the Mycobacterium tuberculosis complex (MTBC), remains a major public health challenge worldwide, with a disproportionate burden in emerging economies1,2. In densely populated urban settings, TB cases often exhibit spatial clustering rather than random dispersion, driven by localized transmission dynamics and the influence of social and environmental risk factors3. Understanding these spatial patterns is critical for identifying transmission hotspots and informing targeted public health interventions. In Accra, Ghana, where TB prevalence remains high, spatial clustering is likely amplified by rapid urbanization, high population density, migration, and persistent socio-economic disparities4. According to the WHO Global Tuberculosis Report 2024, Ghana’s estimated TB incidence in 2023 was 134 cases per 100,000 population (95% UI: 125–145), reflecting a modest decline over the past decade but slower progress compared to global averages1. Case notification rates have improved with expanded diagnostic capacity, yet a substantial gap remains between estimated incidence and reported cases. Neighboring countries such as Nigeria (219 per 100,000) and Côte d’Ivoire (128 per 100,000) report comparable or higher burdens1, underscoring the persistent regional challenge of TB control in West Africa. Recent studies have demonstrated the value of geospatial analysis in elucidating TB transmission dynamics, revealing that high-risk clusters frequently arise in areas marked by poverty, overcrowding, and environmental stressors5. Moreover, spatial proximity between TB cases has been linked to shared, modifiable risk factors, such as poor living conditions and limited access to healthcare services6.

MTBC has evolved with remarkable success, demonstrating a high degree of adaptation to the human host that enables it to persist, replicate, and transmit efficiently within human populations7,8. This evolutionary adaptation contributes to the selective transmission of particular bacterial strains, shaping the inter-host distribution of MTBC genotypes and broader epidemiological patterns7,8. Crucially, these transmission dynamics are closely linked to spatial proximity and patterns of human interaction, with the likelihood of transmission increasing with physical closeness and frequency of contact between individuals9,10,11. Conversely, as the spatial distance between potential hosts increases, the probability of transmission declines, underscoring the central role of geographic and social connectivity in the spread of TB9,10,11.

Studies of TB spatial epidemiology have provided valuable insights into the identification of high-risk clusters and their association with population-level factors such as poverty, HIV prevalence, substance use, and occupation12,13,14,15,16. However, most of these studies have focused primarily on area-level or ecological risk factors, often neglecting the individual-level social behaviors, exposures, and susceptibilities that shape transmission and clinical progression. While population-level analyses are important for public health planning, they may not fully capture the nuances of interpersonal interactions and shared risk factors that drive local transmission within urban communities. Individual-level risk factors offer a more granular understanding of TB dynamics by accounting for each person’s unique circumstances, including social behaviors, living conditions, comorbidities, and immunocompromised status17,18. Considering these individual-level factors can identify those at highest risk for infection and disease and can inform prevention, screening, and treatment strategies that are more precisely targeted to individual needs. Moreover, such analyses may uncover underlying sociocultural, economic, and environmental determinants of TB risk that are obscured when relying solely on population-level data19.

In this study, we leveraged individual-level data from a TB treatment cohort enrolled at a single referral health facility in the Greater Accra Region, Ghana, to examine whether shared TB risk factors were associated with closer residential proximity among TB cases. We hypothesized that cases sharing clinical and socio-behavioral risk factors, such as HIV co-infection and substance use history, would live closer to one another compared to cases without shared risk factors. The public health relevance of this study lies in its contribution to a better understanding of how individual-level risk factors relate to residential proximity among TB cases in an urban, high TB/HIV burden setting. By identifying patterns of shared risk factors and spatial proximity, this study can help generate hypotheses for future research and contribute to efforts aimed at refining TB control strategies in similar urban environments.

Materials and methods

Study location

The study included participants residing in the Greater Accra region and was conducted at the Korle Bu Teaching Hospital (KBTH) Chest Clinic, the largest public tertiary hospital in the southern part of the country. KBTH is in the Accra Metropolitan Authority (AMA) administrative district. According to the 2021 population and housing census, the AMA has an estimated population of 284,124 and covers a total land area of 173 square kilometers20. The KBTH Chest Clinic sees approximately 70 TB cases per month and draws patients nationwide21. With its central location in a high burden setting and its role as a major TB referral center, the diversity of TB patients presenting at the KBTH Chest Clinic enables passive surveillance of TB risk factors in the population.

Study population

Persons aged 15 years or older with TB, including those with new TB diagnosis and those with recurrent or previously treated TB disease, were eligible for study participation. Individuals with multidrug-resistant TB (MDR-TB) were eligible and were included if they received care at the KBTH Chest Clinic. Patients diagnosed with extensively drug-resistant TB (XDR-TB) were not enrolled, as such cases are referred to specialized treatment centers. The study participants were screened and diagnosed with TB according to the Ghanaian National TB Control Program (NTP) guidelines, using a same-day spot sputum sample analyzed by GeneXpert MTB/RIF Ultra for initial diagnosis and sputum smear microscopy for acid-fast bacilli for treatment follow-up. Treatment was also initiated based on the physician’s clinical diagnosis of TB if the GeneXpert MTB/RIF Ultra test was either negative or when an appropriate specimen for testing was not available. We included individuals with TB in the study if they had a strong clinical suspicion of TB based on abnormal chest X-ray findings and/or symptoms of TB disease (cough > 1 week, hemoptysis, fever, weight loss, night sweats, and severe fatigue) with known HIV status or willingness to test. We excluded from study participation individuals with TB who were under the age of 15 years at the time of screening, were too sick to provide informed consent, or had a history of relapse after second-line anti-TB treatment.

Methods

Patients answered a questionnaire on demographic, social, and behavioral risk factors for TB. Residential GIS locations were collected using handheld GPS devices during post-diagnosis home visits for contact investigations by trained public health nurses and disease investigators. These residential coordinates were used as a proxy for potential exposure environments, recognizing that individuals may be exposed to MTBC both within and outside the home setting.

Spatial clustering and statistical analysis

Spatial clustering analysis

Administrative boundary polygons for Ghana were obtained from the Database of Global Administrative Areas (GADM), version 4.122 and used to map TB cases based on residential coordinates. Because district-level analysis was too coarse to guide local TB control, the study area was divided into a grid of 1.5 km2 cells using QGIS23, with TB cases aggregated within each cell. This resolution was selected to balance analytic granularity with case density, ensuring sufficient case counts per unit for reliable cluster detection while approximating neighborhood-scale transmission dynamics24,25. Population estimates were derived by applying a zonal statistics routine in QGIS to rasterized 2020 UN-adjusted population data from the WorldPop project26, with a spatial resolution of 100 m227. The total population in each 1.5 km2 cell was estimated by summing the population counts of overlapping raster cells. Crude TB incidence per 100,000 people was calculated for each grid cell and mapped alongside case counts to visualize the spatial distribution of TB (Fig. 1).

Fig. 1
figure 1

Spatial Distribution of TB Cases and Incidence per 100,000 Population, Greater Accra Region. Panel A: Point distribution of enrolled TB cases. Panel B: Number of TB cases aggregated by 1.5 km2 cell. Panel C: Crude TB incidence per 100,000 population within 1.5 km2 grid cells, based on 2020 population estimates. For privacy reasons, the figure does not include a pop-out locator map. The study setting (Accra Metropolitan Area, Korle Bu Teaching Hospital) is described in the text.

Full size image

Local Indicators of Spatial Association (LISA) analysis was conducted using GeoDa28 to identify clusters of TB incidence at the grid-cell level. LISA identifies local clusters by evaluating whether a grid cell has a value significantly higher or lower than would be expected under spatial randomness, considering neighboring cells. High-high clusters represented grid cells with elevated incidence surrounded by similarly high-incidence cells, whereas low-low clusters indicated low-incidence cells surrounded by low-incidence neighbors. High-low and low–high outliers were also detected. A first-order queen contiguity matrix defined spatial neighbors, assigning a weight of 1 if cells shared an edge or vertex, and 0 otherwise. Statistical significance was assessed using pseudo-p-values generated from 999 random permutations (p < 0.05)28. The methodology for LISA analysis has been described previously29.

Spatial proximity definition

To quantify residential proximity, we first transformed the geographic coordinates of each case’s home residence into projected Universal Transverse Mercator (UTM) coordinates (Zone 30N, EPSG:32,630) using the sf package in R30. Next, we generated all unique dyadic pairs of cases and computed Euclidean distances between the residences within each pair. To enhance interpretability and ensure all values were positive, we centered the distance variable at the minimum observed value (1.29 m) and then shifted it upward by 150 m. This transformation was made to anchor model intercept at a meaningful spatial threshold of 150 m beyond the closest observed pair. The resulting variable was then rescaled by dividing by 100, such that one unit in the model corresponded to a 100-m increase in residential separation from the 150 m baseline. This transformed distance variable was used as the outcome in all subsequent modeling.

Statistical analysis

Descriptive statistics for demographic, clinical, and socio-behavioral risk factors were summarized (Table 1). To assess the association between shared risk factors and spatial proximity, we fit a Bayesian generalized linear mixed model (GLMM) using the brms package in R31. The outcome was the scaled and shifted Euclidean distance, modeled with a Gamma distribution and a log link, appropriate for continuous, positive, and right-skewed outcomes. The model included fixed effects for shared risk factors (Supplementary Table 1), entered as binary or categorical variables, and crossed random intercepts for each individual in a dyad to account for the non-independence of observations. We specified weakly informative priors: Normal (0, 5) for fixed effects, Normal (0, 10) for the intercept, and Exponential (2) for the standard deviations of the random effects. The model was estimated using Markov Chain Monte Carlo (MCMC) with two chains, 5,000 iterations per chain, and convergence was assessed via trace plots and R-hat statistics. Posterior distributions were approximately normal, and posterior predictive checks indicated good model fit, with observed and simulated distances closely aligned (Supplementary Fig. 2). Model coefficients were interpreted as the multiplicative change in residential distance per 100-m increase from the 150-m baseline. Coefficients less than 1 indicated that dyads sharing a given characteristic tended to live closer together. Associations were considered meaningful when the 95% credible interval of the posterior distribution excluded 1. All analyses were conducted in R version 4.3.132.

Table 1 Demographics, clinical, social, and behavioral characteristics of 150 new and previously treated tuberculosis cases diagnosed from June 2022—June 2023 at a referral health facility in the Greater Accra Region, Ghana.
Full size table

Sample size and study design

This analysis is a secondary study nested within a prospective treatment cohort (N = 150). We did not conduct a separate a priori sample size calculation for the spatial/dyadic regression; instead, we analyzed all available participants from the parent study. The parent cohort size (N = 150) was determined by a power analysis for a different primary aim of detecting a 35% prevalence of within‑host genetic diversity among single‑strain infections and a 15–25 percentage‑point difference between sputum sampling strategies (spot vs. early morning) assuming 30% discordance, providing ≥ 80% power for that aim. For the current work, the dyadic design yields 11,175 unique residential pairs (150 × 149/2). Because dyads are not independent, we modeled crossed random intercepts for each individual to account for dependence; thus, the effective information is greater than N = 150 but less than the dyad count. We therefore interpret effect sizes with attention to their uncertainty and view the findings as hypothesis‑generating, to be validated in larger, multi‑site cohorts.

Human Ethics approval and consent to participate

This study was approved by the Institutional Review Boards (IRBs) of the University of Florida (IRB202003042) and the University of Ghana Korle-Bu Teaching Hospital (IRB/000,135/2020). All methods were carried out in accordance with relevant guidelines and regulations, including the Declaration of Helsinki. Written informed consent was obtained from all participants using forms reviewed and approved by the respective IRBs.

Results

Study participant characteristics

We enrolled 150 new (n = 131, 87.3%) and previously treated (n = 19, 12.7%) individuals with pulmonary TB. This included two participants with multidrug resistant TB (MDR-TB): one newly diagnosed and one return after default. No extensively drug-resistant TB (XDR-TB) cases were enrolled. The diagnosis of TB was confirmed by GeneXpert MTB/RIF Ultra in 66% of participants and clinical diagnosis in 34% of participants. The treatment-naive and previously treated groups did not differ significantly in terms of demographic, social, or behavioral risk factors (Table 1). Overall, the cohort was predominantly male (68.0%) and of working age (80.0% aged 25–64 years), with half of them (51.3%) engaged in unskilled work. Participants reporting known TB exposure included those with household contact with TB (12.0%) and those with social contact with TB (24.0%) in the past two months. Approximately one-third had a history of alcohol use (36.0%), and over a quarter had a history of recreational drug use (26.7%) in the past year. Fifteen percent were HIV-positive, of whom over 80% had lived with HIV less than one year before TB diagnosis.

Spatial patterns and local clustering of TB incidence

The spatial distribution of TB cases within the study area is shown in Fig. 1. Most patients resided in the southern portion of the study area, near the KBTH, with up to 12 patients per 1.5 square kilometers grid cell (Fig. 1B). Crude TB incidence was not biased by the absolute number of patients per grid cell (Fig. 1C), as areas with higher case counts also corresponded to areas of higher population density (Supplementary Fig. 1). Adjusting for population density provided a more accurate representation of disease burden, ensuring that the identification of TB hotspots reflected true variations in crude incidence rather than differences in population distribution. This approach increased confidence in the detection of localized areas of elevated TB risk through subsequence spatial autocorrelation analysis.

Local Moran’s I (LISA) analysis revealed significant spatial clustering of TB incidence across the study area (Fig. 2). Two high-high clusters were identified in the southern region surrounding KBTH, along with a smaller high-high cluster in the western area and another in the eastern area. High-high clusters indicate grid cells with higher-than-expected TB incidence surrounded by similarly high-incidence neighbors, suggesting potential hotspots for TB transmission. In addition, several spatial outliers were detected, including high-incidence cells surrounded by low-incidence neighbors and vice versa. However, the interpretation of spatial outliers should be made cautiously, as the observed patterns may be influenced by the study’s facility-based sampling design rather than representing true underlying transmission dynamics across the broader population.

Fig. 2
figure 2

Local Spatial Clusters of TB Incidence Identified by LISA Analysis, Greater Accra Region. Each 1.5 km2 grip cell was classified based on Local Moran’s I analysis of crude TB incidence. High-high clusters indicate cells with higher-than-expected incidence surrounded by similarly high-incidence neighbors. Low-Low clusters indicate cells with lower-than-expected incidence surrounded by low-incidence neighbors. High-Low outliers represent high-incidence cells surrounded by low-incidence neighbors. Low–High outliers represent low-incidence cells surrounded by high-incidence neighbors. Clusters were defined using a first-order queen contiguity spatial weights matrix. Statistical significance was determined at p < 0.05 based on 999 random permutations. Non-significant cells are shown in white.

Full size image

Associations between case characteristics and residential proximity

The estimated baseline residential distance between tuberculosis cases with reference characteristics (i.e., when risk factors were not shared) was 6118 m (95% credible interval [CrI]: 4540–8218 m), accounting for centering at the minimum observed pairwise distance, a 150-m shift, and rescaling to 100-m units. All effect estimates represent multiplicative changes in residential distance per 100-m increase from the 150-m baseline (Fig. 3). Sharing the same religious affiliation was associated with a 6% reduction in residential distance compared to mixed-religion pairs (multiplicative effect: 0.94, 95% CrI: 0.91–0.97). Case pairs who reported recent exposure to non-household individuals with cough exhibited a 5% reduction in distance (multiplicative effect: 0.95, 95% CrI: 0.90–1.00), and case pairs who shared exposure to household members with cough showed a 10% reduction (multiplicative effect: 0.90, 95% CrI: 0.82–0.98). Shared presence of hemoptysis was associated with a 2% reduction in residential proximity (multiplicative effect: 0.98, 95% CrI: 0.96–1.00). Treatment history demonstrated strong spatial effects: pairs who were both newly diagnosed lived 50% farther apart (multiplicative effect: 1.50, 95% CrI: 1.24–1.82), whereas pairs who were both previously treated lived 55% closer together (multiplicative effect: 0.45, 95% CrI: 0.37–0.55), indicating tighter residential clustering among previously treated individuals. This pattern may signal sustained exposure to high transmission environments but could also reflect barriers to treatment completion or care access, given that eight of the 19 previously treated individuals in our cohort experienced treatment default or failure.

Fig. 3
figure 3

Associations between shared risk factors and residential proximity among tuberculosis cases in the Greater Accra Region, Ghana. Forest Plot of Significant Risk Factors Associated with Residential Proximity among TB cases. Cases. Estimated multiplicative effects and 95% credible intervals from a Bayesian cross-random effects gamma regression model examining residential distance between tuberculosis cases in Accra, Ghana. Euclidean residential distance was centered at the minimum observed pairwise distance, shifted by 150 m, and rescaled so that one unit corresponds to 100 m. Multiplicative effects less than 1.0 (blue) indicate greater residential clustering (closer proximity), while effects greater than 1.0 (red) indicate greater separation between cases. Non-significant estimates are not shown.

Full size image

In a direct comparison of predicted residential distances between pairs of individuals with previously treated TB and those with newly diagnosed TB (Fig. 4), the posterior probability that individuals with previously treated TB lived farther apart than those with new TB diagnosis was 0. This provides strong evidence that individuals with previously treated TB are more tightly clustered in residential space.

Fig. 4
figure 4

Predicted residential distance between tuberculosis case pairs according to shared treatment history. The figure shows the posterior mean predicted residential distance (in meters) for case pairs where both individuals were formerly treated for TB (bothFormer) compared to case pairs where both were newly diagnosed (bothNew). Euclidean residential distance was centered at the minimum observed pairwise distance, shifted by 150 m, and rescaled so that one unit corresponds to 100 m. Predictions were based on a Bayesian gamma regression model adjusting for demographic, clinical, and exposure characteristics. The probability that individuals with previously treated TB live farther apart than new TB diagnosis was 0, indicating strong evidence that previously treated individuals cluster more tightly in residential space.

Full size image

Discussion

Our spatial epidemiologic study of TB cases in Accra, Ghana, strengthens available evidence that shared socio-behavioral factors may drive the geospatial clustering of urban TB33. While previous studies have documented the spatial heterogeneity of TB cases in different settings34,35,36,37, we demonstrated that individuals with TB in our treatment cohort were spatially clustered in areas of high population density and among individuals with shared risk factors, including shared religious affiliation, recent non-household exposure to individuals with cough, and clinical presentation with hemoptysis. Although these associations reached statistical significance, the estimated effects were models and in some cases the 95% credible intervals were close to the null. This suggests that, while consistent with established knowledge of TB exposure risk in social and community settings, the magnitude of these associations should be interpreted with caution. Their clinical or public health significance remains uncertain, and larger studies will be needed to determine whether such modest spatial effects translate into meaningful opportunities for intervention. The significance of these shared risk factors aligns with established knowledge of TB exposure risk in social settings and among close contacts38,39,40. Shared religious affiliation may reflect participation in community activities, such as communal worship, that facilitate opportunities for TB transmission41. Similarly, recent household and non-household exposure to individuals with cough and shared hemoptysis symptoms are consistent with active transmission dynamics within close-contact social networks. Close contact with individuals exhibiting persistent coughs, a common symptom of TB, can contribute to transmission. This is particularly relevant within social networks where friends and relatives share similar living conditions, socioeconomic status, and behaviors42,43. Delayed recognition of TB symptoms, such as persistent cough, within these groups may delay diagnosis and initiation of appropriate treatment, consequently allowing for continued transmission within social circles. Our findings also revealed that individuals with previously treated TB exhibited strong spatial clustering, which may indicate ongoing reinfection risk through sustained residence in high-transmission environments44. However, an alternative explanation is that this clustering reflects relapse associated with incomplete or inadequate treatment. Studies from diverse settings have shown that poor treatment outcomes, including failure and default, are themselves spatially clustered and shaped by barriers to care such as distance, cost, and stigma45,46. Furthermore, even after treatment completion, individuals remain at elevated risk for recurrent TB through both relapse and reinfection, underscoring the vulnerability of previously treated populations47. Taken together, these findings suggest that the clustering of previously treated individuals likely reflects a combination of biological (e.g., reinfection risk) and structural (e.g., treatment access and adherence) factors. Targeted TB control efforts must therefore address not only transmission dynamics in high-risk communities but also health system barriers that hinder treatment completion and long-term recovery.

This study has several limitations. Movement patterns, workplace exposures, and social contacts outside the home were not captured in this study and may also contribute to transmission risk. Additionally, we could not account for other factors that may influence case geographic distribution, including MTBC strain types10,48 and migration flows44,49. Data on potential socio-behavioral risk factors, such as household size, income, education, and healthcare access, were not available and may have influenced the observed spatial clustering patterns. We were also unable to capture employment informality which has been identified as an important predictor of TB vulnerability50. Our analysis was also based on cases reported by a single hospital, which may not fully represent the broader TB burden in the study area, as some cases could have been diagnosed elsewhere, undiagnosed, or misclassified. Given the modest sample size, our study may have been underpowered to detect small associations with high precision. Although the dyadic design yields a large number of residential pairs, the effective sample size is limited because of non-independence among dyads. We accounted for this using crossed random intercepts in the Bayesian regression model, but uncertainty in effect estimates remains. Consequently, the results should be interpreted as hypothesis-generating, and validation in larger, multi-site cohorts will be necessary to confirm the robustness and generalizability of these findings.

Despite these limitations, the study also has important strengths. We leveraged individual-level data, rather than ecological proxies, to examine whether shared social and clinical factors are associated with spatial proximity. The integration of geocoded home addresses with Bayesian dyadic regression provides a novel approach to exploring TB clustering in a high-burden urban setting. Furthermore, by explicitly modeling residential proximity as a continuous outcome, our study moves beyond binary cluster definitions and offers a more nuanced understanding of potential transmission dynamics. These strengths enhance the contribution of the study to TB spatial epidemiology in West Africa.

In conclusion, our study revealed significant spatial clustering of TB cases in areas of high population density. Shared socio-behavioral factors, including shared religious affiliation, recent non-household exposure to individuals with cough, and clinical presentation with hemoptysis, were associated with closer residential proximity among TB cases. Leveraging these findings offers strategic opportunities for targeted interventions aimed at interrupting TB transmission within social and community networks. Further studies across Ghana are necessary to corroborate these findings and to better elucidate the mechanisms driving the spatial clustering of TB cases with shared socio-behavioral risk factors. Such investigations would provide a more comprehensive understanding of TB transmission dynamics and inform the development of more effective, context-specific control measures.