An Integrated Machine Learning Framework for Predictive Tuberculosis Surveillance Using Patient-Level Classification, Spatial Risk Clustering, and Temporal Forecasting in Aceh, Indonesia

Tuberculosis (TB), a contagious infectious disease caused by Mycobacterium TB, remains a major public health challenge in Indonesia, particularly in high-burden provinces such as Aceh [1]. Despite comprehensive efforts through national TB control programs, including active case finding, treatment supervision, and community awareness campaigns, TB continues to contribute significantly to morbidity and mortality. These persistent challenges reflect ongoing gaps in healthcare accessibility, patient adherence, early detection, and resource allocation across districts [2]. Moreover, the complex epidemiology of TB, shaped by socio-economic disparities, population mobility, environmental conditions, and variations in healthcare infrastructure, further complicates disease surveillance and intervention planning [3].

A thorough understanding of TB trends and associated risk factors is essential for designing effective public health strategies. Analyses encompassing patient-level treatment adherence, district-level clustering of TB incidence, and temporal patterns of disease transmission offer critical insights into the distribution and propagation of TB within communities. Such insights allow policymakers to prioritize high-risk areas, allocate resources more efficiently, and implement targeted interventions to reduce disease burden [4]. However, conventional epidemiological and statistical approaches often fail to capture the multidimensional interactions among patient behaviors, spatial heterogeneity, and temporal variations, limiting their capacity to support responsive and adaptive TB surveillance systems [5]. In recent years, machine learning has emerged as a powerful analytical tool capable of processing large, complex datasets to identify hidden patterns and generate predictive insights for classification, clustering, and forecasting in infectious disease surveillance [6]. Although machine learning has been applied to TB for various analytical tasks, most existing studies focus on a single dimension, leaving a gap in integrated frameworks that holistically address patient, spatial, and temporal characteristics.

To address these limitations, this study proposes a hybrid machine learning model for TB surveillance, an integrated framework designed to optimize smart TB surveillance in Aceh. The integrated model consists of three complementary modules: (i) patient-level treatment adherence classification using Support Vector Machine (SVM), (ii) district-level TB risk clustering using K-Medoids enhanced with Grey Relational Analysis (GRA) for optimized medoid initialization, and (iii) temporal prediction of TB incidence using polynomial regression. By combining these analytical components, the hybrid model simultaneously captures patient, spatial, and temporal dimensions of TB, enabling more accurate identification of high-risk patients, vulnerable districts, and future case trajectories. This hybrid approach overcomes the limitations of single-dimensional analyses and offers a scalable, data-driven model applicable to other high-burden regions.

The main contributions of this study are as follows:

•Hybrid machine learning approach: Integrating classification, prediction, and clustering into a unified TB surveillance framework.

•Enhanced patient-level classification: Utilizing SVM with multiple kernel functions to identify key predictors of treatment adherence.

•Robust temporal forecasting: Applying regression-based modeling to project future TB incidence trends.

•Optimized spatial clustering: Employing GRA-based medoid selection within the K-Medoids algorithm to more accurately identify high-risk districts.

The remainder of this paper is structured as follows: Section 2 reviews existing literature on TB surveillance and machine learning applications. Section 3 describes the dataset, preprocessing procedures, and methodological framework. Section 4 presents the experimental results, including model performance and clustering analyses. Section 5 discusses the findings, their implications for TB control programs, and directions for future research. Through this integrated model, the study seeks to enhance smart TB surveillance and strengthen evidence-based decision-making in Aceh.

3.1 Proposed method—Hybrid machine learning model

The proposed hybrid machine learning model integrates three complementary analytical modules designed to address patient-level, regional, and temporal dimensions of TB surveillance in Aceh, Indonesia. The overall workflow, combining classification, clustering, and prediction processes, is illustrated in Figure 1.

1.png

Figure 1. Workflow of the proposed method

The first module focuses on treatment completion classification using SVM, which categorizes patients into treatment completed and treatment not completed. This classification provides essential insights into individual treatment outcomes that influence recovery rates and overall disease control. The second module performs regional risk clustering at the district level. The K-Medoids algorithm is used to group districts with similar incidence patterns, while GRA is applied to optimize medoid initialization. This hybrid clustering approach enhances stability and accuracy, enabling more reliable identification of high-risk and low-risk districts. The third module is responsible for predicting TB spread using Polynomial Regression to capture non-linear temporal trends from historical case data. This predictive component supports the forecasting of transmission dynamics and assists in formulating proactive intervention strategies.

3.2 Data preprocessing

Before model development, several preprocessing steps were applied to improve data quality, consistency, and reproducibility. Missing values in critical attributes were removed, while limited missing numerical values were imputed using mean substitution. Outliers were identified using the Interquartile Range (IQR) method. However, extreme values reflecting actual epidemiological conditions were retained unless confirmed as data-entry errors.

Numerical features were normalized using Min–Max scaling to reduce variations in feature magnitude and improve computational stability. For the classification module, the dataset was divided into training and testing sets using a stratified 80:20 split to preserve the original class distribution. All randomized processes, including train-test splitting and medoid initialization, were conducted using a fixed random seed of 42 to ensure reproducibility.

Class imbalance was also evaluated, as the “Completed” category dominated the dataset compared to the “Not Started” and “Incomplete” classes. Although the original distribution was maintained to reflect real surveillance conditions, imbalance-aware techniques such as Synthetic Minority Oversampling Technique (SMOTE), class-weight adjustment, and ensemble learning are recommended for future improvement. In addition, clustering experiments were repeated multiple times, and the reported Calinski–Harabasz Index (CH Index) and Davies–Bouldin Index (DBI) values represent average results to reduce sensitivity to random initialization.

3.3 Patient treatment completion using Support Vector Machine

This module uses the SVM algorithm [21-23] to classify TB patients into three treatment categories: Not Started, Incomplete, and Completed. Each patient is represented by a multidimensional feature vector of demographic, clinical, and treatment-related attributes. The multi-class problem is handled via the one-versus-rest scheme, where a binary classifier is trained for each class, and the category with the highest decision score is assigned. Performance is evaluated using cross-validation [24] and class-wise metrics, supporting data-driven monitoring of treatment adherence and early identification of patients at risk of non-completion. The complete procedure of the proposed patient treatment completion classification module is summarized in Algorithm 1.

The decision boundary is defined by a hyperplane, as shown in Eq. (1) [25].

$f(x)=w^T x+b$       (1)

The SVM optimization seeks to maximize the margin while allowing certain misclassifications, formulated as in Eq. (2) [26]:

$\min _{w, b} \frac{1}{2}\|w\|^2+C \sum_{i=1}^N \xi_i$       (2)

subject to the constraint expressed in Eq. (3) [27]:

$\mathcal{Y}_i\left(w^T x_i+b\right) \geq 1-\xi_i, \forall_i$        (3)

where, C is the regularization parameter, and ξ_i represents slack variables.

The classification performance is evaluated using standard metrics derived from the confusion matrix. Accuracy, precision, recall, and F1-score are calculated respectively as shown in Eqs. (4)–(7) [28]:

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$       (4)

Precision $=\frac{T P}{T P+F P}$      (5)

Recall $=\frac{T P}{T P+F N}$    (6)

$F 1=2 \times \frac{\text { Precision × Recall }}{\text { Precision }+ \text { Recall }}$      (7)

The patient-level data for classification of TB treatment outcomes in Aceh, obtained from the Provincial Health Office for the period 2019–2024, are presented in Table 1. Each record contains a set of treatment indicators (HRZE1–HRZE8 and HR1–HR16), represented as binary values, indicating the presence or absence of specific treatment responses. The final column denotes the treatment status, categorized as Not Started, Incomplete, or Completed.

Table 1. Dataset information for tuberculosis (TB) treatment outcome classification

The low F1-scores observed in the “Not Started” and “Incomplete” categories indicate substantial class imbalance within the surveillance dataset. Misclassification of these minority groups may delay intervention for patients at risk of treatment discontinuation. Future improvements should incorporate imbalance-aware approaches such as SMOTE, class-weighted SVM optimization, and ensemble learning strategies to improve minority-class detection reliability.

3.4 Regional risk clustering (K-Medoids and Grey Relational Analysis + K-Medoids)

To identify TB risk patterns across districts, two clustering approaches are implemented. The conventional K-Medoids algorithm [29-31] groups districts based on similarities in incidence rates, while the GRA + K-Medoids method improves cluster stability by optimizing medoid initialization. Cluster quality is assessed using the CH Index [32] and DBI [33-35]. The overall workflow of the proposed regional risk clustering approach is outlined in Algorithm 2.

Given a dataset {x₁, x₂, …, x_n} and the desired number of clusters k, the dissimilarity between two objects is computed using the Manhattan distance, as shown in Eq. (8):

$d\left(x_i, x_j\right)=\sum_{m=1}^p\left|x_{i m}-x_{j m}\right|$           (8)

The clustering cost is calculated as the sum of distances between each object and its assigned medoid, as expressed in Eq. (9):

Cost $=\sum_{m=1}^n\left(x_i, m\left(x_i\right)\right)$       (9)

The K-Medoids algorithm proceeds as follows [36]:

1) Initialization:

Select k objects as the initial medoids (either randomly or using GRA optimization).

2) Assignment step:

Assign each object to the nearest medoid based on the Manhattan distance.

3) Update step:

For each cluster, compute the total cost if the current medoid is replaced by another object within the cluster. The object that minimizes the clustering cost becomes the new medoid.

4) Iteration:

Repeat the assignment and update steps until medoids do not change or the total cost converges to a minimum.

To improve the robustness of medoid initialization, GRA is employed. The normalized value of each indicator is obtained as in Eq. (10):

$x_i^{\prime}(k)=\frac{x_i(k)-\min (x(k))}{\max (x(k))-\min (x(k))}$         (10)

The Grey Relational Coefficient (GRC) is then computed as shown in Eq. (11):

$\xi_i(k)=\frac{\Delta_{\min }+\rho \Delta_{\max }}{\Delta_i(k)+\rho \Delta_{\max }}$   (11)

where, $\Delta_i(k)=\left|x_0(k)-x_i(k)\right|$ and $\rho \in[0,1]$ is the distinguishing coefficient, typically set to 0.5. The overall Grey Relational Grade (GRG) is then calculated as in Eq. (12):

$\gamma_i=\frac{i}{n} \sum_{k=1}^n \xi_i(k)$      (12)

Table 2. District-level tuberculosis (TB) incidence data in Aceh

Districts with the highest GRG values are selected as initial medoids for clustering. Cluster quality is assessed using the CH and the DBI. The CH Index is defined as in Eq. (13):

$C H=\frac{T_r\left(B_k\right) /(k-1)}{T_r\left(W_k\right) /(n-k)}$       (13)

while the DBI is calculated as in Eq. (14):

$D B I=\frac{1}{k} \sum_{i=1}^k \max _{j \neq 1} \frac{s_i+s_j}{d\left(c_i, c_j\right)}$       (14)

where, s_i is the average intra-cluster distance and d(c_i, c_j) is the distance between centroids. A high CH index and a low DBI indicate optimal clustering quality. Table 2 presents the district-level TB incidence data, including population, detected cases, and corresponding incidence percentages, which serve as input for the clustering analysis.

3.5 Tuberculosis spread prediction (Polynomial Regression)

Historical TB case data are analyzed using Polynomial Regression to capture non-linear transmission trends. The model forecasts potential future case counts at the district level. Prediction accuracy is evaluated using the Mean Squared Error (MSE) metric. Let the dataset be denoted as (x_i, y_i), where x_i corresponds to the temporal variable (e.g., year) and y_i represents the number of recorded TB cases. The complete procedure of the proposed TB spread prediction module is summarized in Algorithm 3.

The polynomial regression model of degree d is formulated as shown in Eq. (15) [37]:

$\hat{\mathcal{Y}}=\beta_0+\beta_1 x+\beta_2 x^2+\ldots+\beta_d x^d$       (15)

In matrix representation, the model can be expressed compactly as in Eq. (16) [38]:

$\hat{\mathcal{Y}}=X \beta$     (16)

where, X denotes the design matrix constructed from the polynomial terms of the predictor variable, and β is the coefficient vector to be estimated. The parameter estimation is performed using the Ordinary Least Squares (OLS) method, which minimizes the sum of squared residuals, as expressed in Eq. (17):

$\hat{\beta}=\left(X^T X\right)^{-1} X^T y$        (17)

The predictive accuracy of the model is evaluated using the MSE, which measures the average squared difference between the observed values and the predicted values, as defined in Eq. (18) [39]:

$\mathrm{MSE}=\frac{1}{n} \sum_{i=1}^n\left(\mathcal{Y}_i-\hat{\mathcal{Y}}_i\right)^2$        (18)

A lower MSE value indicates that the polynomial regression provides a more reliable fit to the observed TB case trends [40].

The historical TB case data, organized by district and year as presented in Table 3, serve as input for polynomial regression to forecast future case counts.

Table 3. Historical tuberculosis (TB) cases for prediction

Although polynomial regression effectively captured short-term nonlinear trends, occasional negative prediction values may occur during extrapolation and are not epidemiologically realistic. In the current implementation, negative outputs were constrained to zero during interpretation. Future work should investigate constrained regression techniques, Poisson-based approaches, and hybrid forecasting models to improve reliability for count-based TB surveillance data.

This study introduced an integrated Hybrid Machine Learning Framework for TB surveillance in Aceh Province, combining patient-level classification, district-level clustering, and temporal forecasting. Together, these modules form a decision-support pipeline that provides complementary insights for more effective intervention planning and resource allocation. The classification module employed SVM with systematic kernel and parameter tuning under 10-fold cross-validation. The linear kernel achieved the best performance (mean accuracy = 98.4%, optimal C = 10 with test accuracy 92.5%), while the RBF kernel performed moderately well, and polynomial and sigmoid kernels yielded considerably weaker results. The classifier demonstrated near-perfect accuracy for the majority “Completed” class but struggled with minority categories due to class imbalance, highlighting the need for imbalance-aware approaches if detection of minority classes is operationally important. District-level clustering was performed using K-Medoids with GRA for medoid initialization. Compared with random initialization, the GRA-based method achieved higher clustering quality (CH = 128.0 vs. 122.6; DBI = 0.904 vs. 0.95) and required fewer iterations. The final stratification grouped districts into three risk categories, with six high-risk districts (including Pidie, Aceh Utara, Aceh Timur, Banda Aceh, Bireuen, and Aceh Besar) bearing the heaviest TB burdens. This spatial risk mapping provides a stable and interpretable basis for geographically prioritized interventions.

Polynomial regression was applied to temporal case trends and revealed a general upward trajectory for 2025–2030, with the steepest increases in Pidie, Aceh Utara, and Aceh Timur. While the forecasts provided useful directional signals for short- to medium-term planning, occasional implausible negative values underscored the limitations of polynomial extrapolation, emphasizing the need for epidemiological validation before operational use. The integrated framework carries clear implications for policy and practice. At the clinical level, SVM classification can flag non-adherent patients for follow-up. At the programmatic level, GRA-based clustering enables health authorities to prioritize high-burden districts for surveillance and treatment resources. At the strategic level, temporal forecasting supports procurement, staffing, and budget planning. When combined, these modules allow for a more coordinated and efficient approach to TB control. Future work should enhance both granularity and robustness by incorporating socioeconomic, healthcare access, mobility, and environmental data into the clustering and forecasting modules; testing ensemble and deep-learning models for temporal prediction; and applying imbalance-aware classification strategies to improve detection of minority adherence classes. Real-time data integration would further strengthen operational responsiveness, while extending the framework to other provinces would help test its generalizability and inform national-level TB control strategies.