Comparative Analysis of Clustering Algorithms for Health Risk Profiling Based on Dietary and Physical Activity Patterns

Non-Communicable Diseases (NCDs) are a group of diseases that are not caused by infection but are influenced by lifestyle, environmental, and genetic factors [1]. According to the World Health Organization (WHO), the prevalence of NCDs is increasing among young adults, a stage of life that should ideally be free from serious health problems. The Global Burden of Disease Study 2019 revealed that NCDs such as hypertension, diabetes, and cardiovascular disease have become the leading causes of morbidity and mortality worldwide, including in Indonesia [2]. Lifestyle factors, particularly food consumption patterns, adherence to a healthy diet, and physical activity, play a crucial role in shaping NCD risk. Previous research has also shown that the prevalence of risk factors such as obesity and hypertension is rising among young age groups [3-5]. A systematic analysis of the Global Burden of Disease Study 2010 emphasized that NCDs are now among the leading causes of death globally, with significant increases among young adults [6, 7]. This highlights the urgency of profiling food consumption and lifestyle behaviors as a preventive approach to reduce long-term health burdens.

Food consumption patterns and adherence to healthy diets are essential in preventing obesity, diabetes, and cardiovascular diseases. However, these behaviors are often shaped by demographic characteristics, physical activity, and health awareness, making them complex to analyze. In Indonesia, mixed findings have emerged: one study reported no significant association between fast-food consumption and obesity in adolescents [8, 9], while others found that irregular diets and unhealthy food intake increased the risks of gastritis [10] and hypertension [11]. These studies suggest that poor diet quality and unbalanced nutrition remain important risk factors for NCDs. In addition, health literacy, particularly media health literacy, has been shown to influence behaviors such as diet, smoking, and physical activity, making it a potential preventive strategy among adolescents and young adults [12].

Beyond behavioral insights, data-driven approaches have been increasingly adopted. For example, Alosaimi et al. [11] showed that a combination of poor diet, low physical activity, and sedentary lifestyles significantly increases the risks of obesity and cardiovascular disease [13]. Recent analyses using European adolescent cohorts further reinforce this concern. Findings indicate that adolescents who adhere more closely to healthy dietary patterns such as the EAT Lancet recommendations are significantly more likely to achieve ideal cardiovascular health profiles [14]. Complementary evidence shows that eating behaviors driven by emotional stress, food interest, and motivational factors are strongly associated with higher consumption of energy-dense foods [15]. Nevertheless, current health systems remain more focused on disease detection after onset rather than proactive, data-driven prevention based on food consumption and lifestyle [11].

To address this issue, clustering methods have been widely applied to segment food consumption and lifestyle patterns. K-Means is popular due to its scalability and computational efficiency. López-Gil and Martínez-López [13] applied K-Means to children’s macronutrient intake and identified three major dietary profiles [10]. Hierarchical Clustering provides the advantage of exploring data structures at multiple levels, as seen in workflows that combine Hierarchical Clustering and K-Means with principal component analysis for dietary pattern derivation in adult women [16]. Meanwhile, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) offers the ability to identify clusters of arbitrary shapes and isolate noise [17]. Using the sampling-based clustering enhancement introduced by de Moura Ventorim et al. [18], the method was applied to map obesity risks based on eating habits and physical activity, enabling the detection of vulnerable outlier populations. However, most existing studies rely on only one clustering method, limiting the comprehensiveness of segmentation results.

Although numerous studies have examined dietary and lifestyle factors associated with NCDs, few have systematically compared clustering algorithms for health-related profiling. Existing approaches tend to rely on a single method and thus fail to capture the variability of individual behaviors in food consumption and lifestyle. Yet each algorithm has distinct strengths and weaknesses that may influence segmentation quality.

This study seeks to fill that gap by developing a segmentation model using three clustering algorithms: K-Means, Hierarchical Clustering, and DBSCAN to determine which method produces the most cohesive and well-separated clusters for health behavior profiling. The novelty of this research lies in its systematic, machine learning based comparison of clustering techniques in the context of NCD prevention. By leveraging these methods, this study aims to identify distinct patterns of food consumption and lifestyle, thereby providing a foundation for more targeted promotive and preventive interventions, particularly for young adults at risk.

3.1 K-Means clustering

The optimal number of clusters was determined using the elbow method. As shown in Figure 1, the distortion score decreased consistently as the number of clusters (k) increased. The inflection point was observed at k = 8, with a distortion score of approximately 6960.193. This point represents the optimal number of clusters, where the rate of decrease in distortion score begins to flatten significantly. Selecting this elbow point ensures a balance between clustering accuracy and model simplicity, thereby preventing overfitting and ensuring effective segmentation of data [19-21].

Figure 1. The elbow method graph for determining the optimal

The blue curve in the graph represents the distortion score, which consistently decreases as the number of clusters increases, indicating improved compactness within clusters. Meanwhile, the green dashed line indicates the computation fit time required for each clustering process, showing the computational cost as k increases. The vertical dashed line highlights the identified elbow point at k = 8, which is selected as the optimal cluster number for further analysis.

Although the elbow method identified k = 8 as the optimal point based on the distortion score, subsequent clustering analysis consolidated the eight clusters into three clinically and practically meaningful main groups based on health profiles. These three groups are: (1) healthy group comprising individuals with optimal health values, (2) moderate-risk group showing health indicators requiring attention, and (3) high-risk group with critical health conditions. This consolidation approach maintains the granularity of k = 8 in detailed analysis while providing a practical interpretation applicable to the public health context.

3.2 Hierarchical Clustering and DBSCAN

For the Hierarchical Clustering method, the agglomerative approach with the Ward linkage criterion was applied [22]. The dendrogram in Figure 2 illustrates the merging process of clusters. The vertical axis represents the linkage distance, which reflects the level of dissimilarity between merged clusters. By applying a cutoff threshold (red dashed line), the optimal number of clusters was determined to be four to five. These clusters highlight groups of individuals with similar food consumption patterns and lifestyle behaviors.

For DBSCAN, the density-based approach was implemented with parameters eps = 0.5 and min_samples = 5 (tuned experimentally to balance compactness and noise detection). Unlike K-Means and Hierarchical Clustering, DBSCAN does not require pre-specifying the number of clusters [23]. Instead, it identifies dense regions in the data and isolates outliers. This property is particularly suitable for food consumption data, which often exhibit irregular shapes and heterogeneous distributions.

Figure 2. Dendrogram results of Hierarchical Clustering using the Ward method

3.3 Performance evaluation of clustering algorithms

To evaluate the clustering results, three internal validation metrics were used: Silhouette Score, DBI, and Dunn Index. These metrics assess cohesion and separation within clusters. The consolidated comparison is presented in Table 2, ensuring consistency across methods.

Table 2. Comparison of clustering algorithm performance

The evaluation results demonstrate that DBSCAN outperforms both K-Means and Hierarchical Clustering, consistent with previous findings showing DBSCAN’s robustness in handling complex datasets [24, 25]. With a Silhouette Score of 0.414, DBSCAN shows well-separated clusters, while its DBI of 1.062 indicates compact groupings with clear separation. The Dunn Index of 1.136 further supports the robustness of DBSCAN clusters.

In contrast, Hierarchical Clustering indicates limited performance. Its Silhouette Score (0.046) and Dunn Index (0.239) indicate weak cluster separation and a considerable degree of overlap between clusters, although its performance is marginally better than that of K-Means. Meanwhile, K-Means exhibits the poorest clustering quality, with a very low Silhouette Score (0.045) and the highest Davies–Bouldin Index (2.936), suggesting unclear and overlapping cluster structures. In comparison, DBSCAN clearly outperforms the other algorithms, achieving a substantially higher Silhouette Score (0.416), a low Davies–Bouldin Index (0.448), and the highest Dunn Index (1.430), reflecting well-separated and compact clusters.

From a methodological perspective, this result aligns with computational studies indicating that datasets with high variability require clustering approaches capable of capturing sub-spatial structures and irregular cluster shapes, for which density-based methods such as DBSCAN are particularly effective [26].

Overall, the comparison across the three algorithms highlights the strength of DBSCAN in capturing the complexity of food consumption and lifestyle patterns, particularly due to its ability to detect outliers and adapt to clusters of arbitrary shapes. These findings are consistent with previous studies, such as Alosaimi et al. [11], and are further supported by technical reviews showing DBSCAN’s superior adaptability in handling irregular behavioral data [27], which reported that clustering methods can effectively identify combined patterns of diet and physical activity [28]. Similarly, the HELENA Study confirmed that unhealthy dietary habits are often linked with low physical activity levels [11]. Furthermore, evidence from the PeNSE 2015 Survey demonstrated that adolescents with poor diets and insufficient physical activity face an elevated risk of obesity and hypertension from an early age. Together, these results suggest that machine learning-based clustering, especially DBSCAN, provides a more accurate foundation for designing targeted and data-driven health interventions aimed at reducing the risk of NCDs among young adults.

The eight clusters obtained from K-Means were analyzed based on their mean scores across key health indicators, including dietary diversity, physical activity, BMI, smoking habits, sanitation, and healthcare access. Clusters with similar health characteristics were merged into three meaningful groups: healthy, moderate risk, and high risk to enhance interpretability. The same grouping logic was consistently applied to the Hierarchical and DBSCAN outputs to ensure comparable health profile segmentation across algorithms.