Urbanization, geographic accessibility, and delayed entry into tuberculosis-designated care in southwestern China: a machine-learning prediction study

Technical Summary: Urbanization, Geographic Accessibility, and Delayed TB Care in Southwestern China

Problem Statement
Tuberculosis (TB) patient delay—the interval between symptom onset and entry into effective care—remains a critical barrier to early diagnosis, treatment, and transmission control. While previous research has identified demographic, socioeconomic, and clinical determinants of delay, studies in mountainous, rapidly urbanizing regions often rely on coarse geographic proxies (e.g., rural residence or administrative boundaries) rather than fine-scale road-network accessibility. In settings like Guangxi and Guizhou, China, where terrain is complex and settlements are dispersed, these coarse measures may fail to capture the true physical burden of reaching TB-designated facilities. Furthermore, the dynamic nature of urbanization and infrastructure investment continuously reshapes spatial access, yet few studies examine how these processes alter the pathways to designated care. This study addresses the gap in understanding how fine-scale geographic accessibility contributes to delayed entry into TB-designated care within a multi-ethnic, mountainous context.

Methodology
The authors conducted a retrospective prediction-modelling study using a large, routine surveillance dataset of 60,778 newly registered pulmonary TB patients from Guangxi and Guizhou provinces.

• Data Sources: The analysis integrated TB surveillance records, clinical and public health management variables, health insurance information, hospitalization diagnosis codes, and high-resolution geographic accessibility data.
• Outcome Definition: The primary outcome was "delayed entry into TB-designated care," defined as an interval of at least 14 days between symptom onset and the first recorded visit to a TB-designated facility. A 7-day threshold was used for sensitivity analysis.
• Feature Engineering: Predictors included demographics, migration status, insurance type, comorbidities (HIV/AIDS, diabetes), temporal features, symptom characteristics, and detailed geographic accessibility metrics (nearest and average road-network travel time, road-network distance, and straight-line distance). Missing data were handled via informative categories, mode imputation, or model-specific estimation.
• Model Development: The dataset was split into a 70% training set and a 30% held-out test set using outcome stratification. Five-fold stratified cross-validation was employed for hyperparameter tuning. Candidate models included LightGBM, XGBoost, Random Forest, Elastic Net logistic regression, and a mean-probability ensemble.
• Evaluation Metrics: Performance was assessed via discrimination (ROC-AUC, PR-AUC), calibration, decision curve analysis (net benefit), and variable importance (SHAP values). Thresholds were selected to maximize Youden's index, with additional reporting for sensitivity-favored thresholds.

Key Results

• Prevalence: 63.5% (38,590/60,778) of patients experienced delayed entry into TB-designated care. Delayed patients were generally older, more likely to be farmers, and faced less favorable geographic accessibility (e.g., median nearest road-network travel time was 2126 seconds vs. 2035.5 seconds for non-delayed patients).
• Model Performance: LightGBM achieved the best performance on the held-out test set, with a ROC-AUC of 0.703 and a PR-AUC of 0.796. At the Youden-index threshold, it demonstrated an accuracy of 0.651, sensitivity of 0.678, and specificity of 0.603. XGBoost performed nearly identically, while Random Forest and Elastic Net showed lower discrimination.
• Predictor Importance: Geographic accessibility indicators (road-network distance, travel time, and straight-line distance) emerged as the most informative predictor domains, ranking higher than conventional demographic, clinical, and service-use variables. SHAP analysis confirmed that these spatial variables drove the model's predictive power.
• Robustness: Sensitivity analysis using a 7-day delay threshold yielded a higher ROC-AUC (0.727) and PR-AUC (0.874), consistent with the higher event rate. Subgroup analyses across age, sex, and residence status showed broadly consistent discrimination (ROC-AUC 0.688–0.706).

Significance and Claims
The paper claims that fine-scale geographic accessibility provides critical predictive information that is often obscured by coarse administrative proxies. The study repositions the understanding of TB delay in mountainous, multi-ethnic settings:

1. Spatial Dynamics: Geographic barriers are not static topographic features but are dynamically reshaped by urbanization, infrastructure investment, and health-system decentralization.
2. Predictive Hierarchy: When direct measures of road-network access are included, they often supersede traditional predictors like patient source, insurance status, or occupation in identifying at-risk populations. This suggests that previous associations between these factors and delay may partly reflect unmeasured spatial access issues.
3. Policy Implications: The authors argue that reducing TB delay requires integrating transport planning, urban development strategy, and health-service redesign. Policies should not treat geographic barriers as immutable background conditions. Instead, routine TB surveillance linked with spatial accessibility measures can support targeted outreach, referral coordination, and seasonal service planning (e.g., adjusting for agricultural cycles or rainy seasons).
4. Methodological Contribution: The study demonstrates the utility of machine learning (specifically ensemble tree-based methods) in integrating complex, non-linear interactions between demographic, clinical, and spatial data for public health risk stratification.

The authors maintain a modest tone regarding the model's utility, characterizing it as a "surveillance-based planning tool" rather than a deterministic decision rule for individual patients. They emphasize that external and prospective validation is necessary before operational deployment and acknowledge limitations such as the lack of data on behavioral factors (e.g., health literacy, stigma) and the inability to observe care-seeking prior to the first TB-designated visit.