Clustering Analysis of Ankle–Brachial Index Related Metabolic and Body Composition Profiles Using K-Means Approach

Abstract

Purpose

The Ankle–brachial Index (ABI) is a non-invasive diagnostic tool used to assess peripheral perfusion disorders. However, the increasing number of clinical parameters involved in ABI assessment may complicate clinical decision-making. This study aimed to explore clustering of metabolic and body composition profiles in relation to ankle–brachial index (ABI) values using a K-means approach to identify patterns associated with peripheral perfusion.

Patients and Methods

A cross-sectional study was conducted using primary and secondary data from 1093 participants collected from community health centres, hospitals, and academic institutions in the Malang region between 2023 and 2025. Thirteen clinical and metabolic parameters were analysed, including age, blood pressure, blood glucose, lipid profile, uric acid, and body composition indices. ABI was measured using vascular Doppler ultrasound, and body composition was assessed using bioelectrical impedance analysis. K-means clustering was performed using the K-means algorithm, both with and without feature selection based on the gain ratio method. Cluster quality was evaluated using the Elbow Method and Silhouette Score.

Results

The optimal clustering solution was identified at k = 2 for both approaches. Feature selection reduced the number of parameters from 13 to 7 and improved cluster quality, with the Silhouette Score increasing from 0.503 to 0.559. Visceral fat, resting metabolism, body mass index, body fat, skeletal muscle, subcutaneous fat, and cell age were the most influential parameters in distinguishing patient clusters. The resulting clusters represented distinct metabolic risk profiles, ranging from healthier individuals to those with elevated cardiometabolic risk.

Conclusion

K-means clustering identifies distinct metabolic and body composition profiles associated with ankle–brachial index (ABI) values. These clusters represent data-driven patterns and should be interpreted as descriptive associations rather than diagnostic or predictive categories. The findings provide exploratory insights into metabolic characteristics related to peripheral perfusion and suggest that metabolic and body composition parameters may help group ABI into distinct profiles. This approach may have potential relevance in nursing practice by supporting clinical consideration; however, further validation is required.

Introduction

Peripheral perfusion disorders are a common manifestation of vascular disease and are associated with increased cardiovascular morbidity and mortality.¹ The Ankle–Brachial Index (ABI) is widely recognized as a simple, non-invasive method for detecting peripheral arterial disease and for estimating overall cardiovascular risk.^2,3 ABI measurement has therefore been recommended in both hospital and community settings as part of vascular risk assessment. In nursing practice, impaired peripheral perfusion is an important clinical condition that supports clinical judgment and care planning.^4–6

Data mining techniques are increasingly used to explore patterns in complex health data. These approaches can assist in identifying relationships among variables and improving the understanding of heterogeneous datasets. In vascular research, data-driven methods may provide insights into how metabolic and body composition parameters are associated with vascular status, including ABI.^7–9 However, these approaches should be understood as exploratory tools rather than definitive clinical decision-making methods.

Among data mining techniques, clustering is an unsupervised learning method that groups individuals based on similarity without predefined outcome labels.¹⁰ In contrast to classification methods that predict predefined ABI categories (ischemic, normal, or high), clustering identifies natural groupings within the data. In this study, the K-means algorithm was used as a simple and widely applied clustering technique to explore patterns in metabolic and body composition variables.

It is important to emphasize that clustering is a descriptive method and does not establish diagnostic categories or causal relationships. The identified clusters represent groups of individuals with similar characteristics and may reflect underlying metabolic profiles. Therefore, any observed relationship between these clusters and ABI values should be interpreted cautiously as an association rather than a diagnostic pattern. Additionally, clustering methods have inherent limitations in clinical applications, as they depend on the selected variables and do not incorporate clinical outcomes directly.

Evidence on the application of clustering approaches in the context of ABI assessment remains limited. Therefore, this study aimed to explore the grouping of individuals based on metabolic and body composition parameters using the K-means approach and to examine their association with ABI values. The findings are expected to provide a descriptive understanding of metabolic profiles related to peripheral perfusion, while acknowledging the exploratory nature and limitations of the method.

Materials and Methods

Study Design and Participants

This study employed a cross-sectional analytical design using an K-means clustering approach. Data were obtained from both primary and secondary sources. Primary data were collected from community health centres in the Malang region, Indonesia, in 2025, while secondary data were obtained from community health centres, hospitals, and an academic database between 2023 and 2024. Participants were recruited using a consecutive sampling approach, in which all eligible individuals attending the selected facilities during the study period were invited to participate. Inclusion criteria included adults aged ≥18 years who underwent ABI and body composition measurements, while individuals with incomplete data were excluded from the analysis. Participation was voluntary, and all participants provided written informed consent prior to data collection. A total of 1093 participants were included in the analysis. This study was conducted in accordance with the principles of the Declaration of Helsinki. Ethical approval was obtained from the Health Research Ethics Committee of the Faculty of Health Sciences, Universitas Brawijaya (Approval No. 162/UN10.F17.10.4/TU/2025).

Study Variables and Measurements

Thirteen clinical, metabolic, and body composition parameters were analysed, including age, systolic blood pressure, diastolic blood pressure, blood glucose, cholesterol, uric acid, body mass index (BMI), skeletal muscle percentage, body fat percentage, visceral fat, resting metabolism, cell age, and subcutaneous fat. The Ankle–Brachial Index (ABI) was measured using a vascular Doppler device with an 8-MHz probe (Hi-Dop BT200V, Bistos Co., Ltd., Seongnam, Korea). Body composition parameters were assessed using a bioelectrical impedance analyser (Omron Karada Scan HBF-375, Omron Healthcare, Kyoto, Japan).

Data Preprocessing and Exploratory Analysis

Prior to clustering analysis, exploratory data analysis (EDA) was performed to examine data distribution, identify correlations among variables, and detect potential outliers. Visual exploration included scatter plot matrices and correlation heatmaps.¹¹ Data preprocessing was conducted through normalisation and standardisation to ensure comparability among variables with different units and value ranges.¹²

Clustering Analysis

Initial clustering was performed using the K-means algorithm with all parameters included. The optimal number of clusters was determined using the Elbow Method and Silhouette Score, which evaluate within-cluster inertia and cluster separation.¹³ Based on these evaluations, k = 2 was identified as the optimal clustering solution, and subsequent analyses focused on this configuration. To improve clarity and avoid redundancy, clustering results for higher k values (k = 3 and k = 4) were not emphasised in the final presentation (Figures 1 and 2).

Figure 1.

Comparison of Clustering Performance Without FS based on Inertia Values & Silhouette Score.

Figure 2.

Comparison of Clustering Performance With FS based on Inertia Values & Silhouette Score.

Feature Selection and Dimensionality Reduction

To reduce dimensional complexity, a variance-based and correlation-based feature evaluation approach was applied.¹⁴ Highly correlated variables were examined, and redundant features were considered for exclusion to improve clustering interpretability prior to K-means analysis. In addition, ankle–brachial index (ABI) values were used as a reference to support the identification and prioritisation of metabolic and body composition variables with potential relevance to vascular status. This approach was selected instead of supervised feature selection methods to maintain consistency with the unsupervised nature of K-means clustering.

Cluster Characterisation

For each optimal clustering solution, cluster characteristics were analysed by calculating mean values for each parameter, examining data distribution using box plots, and identifying distinguishing variables based on variation analysis across clusters. Additional visualisation using box plots was performed to support interpretation, while prioritising tabular presentation for clarity.

Data Privacy and Ethical Considerations

All data were anonymised prior to analysis to ensure participant confidentiality. No personally identifiable information was included in the dataset. Although data processing was conducted using Google Colaboratory, only de-identified data were used, and data handling complied with institutional data protection and ethical standards.

Statistical Software and Tools

All analyses were conducted using Python programming language. K-means clustering was implemented using the scikit-learn library, while data manipulation and visualisation were performed using pandas, NumPy, and matplotlib.¹⁵ The use of external platforms was limited to computational purposes, with strict adherence to data privacy and security protocols.

Results

Clustering Performance with and without Feature Reduction

Comparison of clustering performance demonstrated that both approaches achieved optimal solutions at k = 2 based on the Silhouette Score. Clustering without feature reduction yielded a Silhouette Score of 0.503, whereas clustering after feature reduction to seven parameters improved the Silhouette Score to 0.559, representing an 11.1% increase (Figures 1 and 2). Detailed inertia values and Silhouette Scores for each k value are presented in Table 1.

Table 1.

Comparison of Clustering Performance with FS Based on Inertia Value & Silhouette Score

Feature Selection	k value	Inertia	Silhouette
Without FS	2	27,816,919.89	0.503
Without FS	3	17,356,636.67	0.454
Without FS	4	12,357,055.48	0.410
Without FS	5	9,555,971.69	0.383
Without FS	6	8,129,123.61	0.354
Without FS	7	7,151,362.26	0.331
Without FS	8	6,400,419.24	0.312
Gain Ratio	2	25,886,345.09	0.559
Gain Ratio	3	13,531,052.62	0.506
Gain Ratio	4	8,691,586.39	0.502
Gain Ratio	5	5,810,366.53	0.503
Gain Ratio	6	4,298,037.85	0.499
Gain Ratio	7	3,460,777.87	0.492
Gain Ratio	8	2,786,660.97	0.499

The application of K-means clustering enabled the exploration of relationships between metabolic and body composition parameters in relation to ankle–brachial index (ABI) values. The resulting clusters provided a data-driven grouping of participants, which facilitated the identification of parameters that contributed most to variation across clusters. This approach supports the examination of peripheral perfusion–related profiles based on metabolic and body composition characteristics.

A Silhouette Score in the range of 0.503–0.559 indicates moderate cluster separation, suggesting that while clusters are distinguishable, some degree of overlap between groups may still be present. Therefore, the clustering results should be interpreted with caution. The observed clustering pattern suggests that the selected parameters, subcutaneous fat, skeletal muscle, body fat, cell age, visceral fat, body mass index, and resting metabolism, may contribute to differentiating metabolic profiles within the dataset. In contrast, systolic and diastolic blood pressure, blood glucose, cholesterol, and uric acid were not retained during feature selection, indicating that these variables showed less contribution to cluster differentiation in this analysis.

The inertia curve also demonstrated a marked decrease at k = 2, supporting the stability and interpretability of a two-cluster solution compared with higher k values. Although clustering at k = 4 with feature selection showed a comparable Silhouette Score to k = 2 without feature selection, both solutions provided distinct data-driven groupings.

Cluster Characteristics at K = 2 with Feature Selection

At k = 2, two clusters with relatively balanced distributions were identified. Cluster 0 comprised 560 participants (51.2%), while Cluster 1 included 533 participants (48.8%). Parameters related to subcutaneous fat, skeletal muscle, and total body fat showed similar distributions across clusters and remained within normal ranges (Figure 3).

Figure 3.

Visualization of parameter value distribution based on the two formed clusters with Feature Selection.

The most prominent distinguishing parameters were visceral fat (variation 50.25%), resting metabolism (23.11%), and cell age (19.40%). Cluster 0 exhibited a higher metabolic risk profile, characterised by higher mean resting metabolism (1718.8), elevated visceral fat levels (13.8), higher body mass index (28.5), and older cell age (49.6 years). In contrast, Cluster 1 demonstrated a healthier metabolic profile, with lower resting metabolism (1321.6), lower visceral fat (6.9), lower body mass index (23.7), and younger cell age (39.9 years).

Cluster Analysis at K = 3 and K = 4

Additional clustering analyses were conducted to explore alternative solutions. At k = 3, with a Silhouette Score of 0.506, three distinct clusters were identified (Figure 4). Cluster 0 included 454 participants (41.5%), Cluster 1 included 337 participants (30.8%), and Cluster 2 included 302 participants (27.6%). Visceral fat remained the strongest discriminative parameter (variation 64.21%), followed by resting metabolism (31.73%) and cell age (25.72%). One cluster represented a healthier profile with lower visceral fat and younger cell age, while another cluster demonstrated a higher-risk profile with elevated visceral fat and older cell age.

Figure 4.

Visual Cluster Diagram K=3 with Feature Selection.

At k = 4, cluster separation became more specific (Figure 5). The healthy profile cluster decreased in size to 238 participants, whereas the highest-risk cluster comprised 93 participants and demonstrated markedly elevated values across metabolic parameters, including a mean body mass index of 34.4 and a cell age of 59.1 years. The remaining participants were distributed into two intermediate-risk clusters. Across all clustering solutions, visceral fat consistently emerged as the most influential parameter, with a variation of 76.09% at k = 4.

Figure 5.

Visual Cluster Diagram K=4 and Top 3 Parameters with Feature Selection.

Effect of Feature Reduction on Clustering Performance

Feature reduction based on correlation and variance analysis identified seven parameters that yielded equal or improved clustering performance compared with the full parameter set. Performance improvement was most pronounced at higher k values, reaching a 224.0% increase at k = 8 based on the clustering performance improvement analysis (Figure 6). These findings indicate that reducing dimensionality by excluding less informative variables enhances cluster quality and stability.

Figure 6.

Clustering Performance After Feature Selection.

Although the clustering results demonstrate moderate separation and internal consistency, external validation using independent datasets is required to confirm the robustness and generalisability of the identified clusters.

Discussion

This study demonstrates that K-means clustering with an optimal solution of K = 2 provides the most stable grouping of metabolic and body composition profiles in relation to the ankle–brachial index (ABI), both with and without feature selection. However, the K = 4 solution without feature selection achieved comparable cluster quality while offering additional subgroup differentiation. These findings indicate that clustering performance can be achieved without relying on a large number of variables, although such groupings should be interpreted as data-driven patterns rather than diagnostic classifications.¹⁶ Feature reduction further revealed that only 7 out of 13 body composition and metabolic parameters contributed to cluster formation, highlighting the relevance of selected physiological indicators rather than exhaustive data inclusion.¹⁷

The ability of the reduced-parameter model to approximate clustering outcomes obtained from the full dataset suggests potential utility in simplifying data representation. From a nursing perspective, this approach may assist in understanding patterns of peripheral perfusion-related characteristics using fewer but informative parameters, rather than establishing definitive diagnostic criteria. Time efficiency remains an important consideration in clinical assessment; however, the present findings should not be interpreted as directly improving clinical decision-making without further validation.¹⁸ Furthermore, the application of K-means clustering methods aligns with developments in nursing informatics and digital health, particularly in data exploration, documentation support, and hypothesis generation, rather than direct clinical decision support.^8,17

The seven parameters identified as most influential-resting metabolism, visceral fat, body mass index (BMI), subcutaneous fat, skeletal muscle mass, body fat percentage, and cellular age reflect physiological characteristics associated with peripheral perfusion status. Among these, visceral fat emerged as a prominent distinguishing variable within the identified clusters. This finding is consistent with previous studies reporting an association between ABI values and central obesity indicators.^18,19 Visceral adiposity is known to increase the risk of vascular complications through inflammatory and metabolic pathways, particularly in individuals with obesity, insulin resistance, and type 2 diabetes mellitus.^20,21 However, within the context of this clustering analysis, these relationships should be interpreted as associations rather than causal effects. The present findings suggest that clustering of ABI-related profiles may contribute to exploratory identification of vascular risk patterns, but not to definitive early diagnosis.²²

From a pathophysiological standpoint, reduced ABI values in individuals with increased visceral fat are closely linked to endothelial dysfunction and chronic inflammation. Visceral adipose tissue secretes pro-inflammatory mediators, such as C-reactive protein and adipocyte fatty acid–binding protein (A-FABP), which contribute to arterial stiffness, luminal narrowing, and impaired peripheral perfusion.^21,23 Nevertheless, the current study design does not allow inference of causal relationships between these mechanisms and the observed clustering patterns. Several studies have demonstrated that central obesity indices including waist-to-height ratio, conicity index, and a-body shape index exhibit stronger correlations with ABI reduction than BMI, as they more accurately reflect visceral fat distribution.^19,24 These findings support the use of alternative adiposity indicators in conjunction with ABI to enhance PAD risk stratification.

Resting metabolism also played a significant role in ABI clustering, reflecting the close interaction between metabolic activity and vascular function. Alterations in metabolic profiles, including lipid and amino acid metabolism, have been shown to correlate with ABI values and PAD severity.²⁵ However, in this study, resting metabolism should be interpreted as a variable contributing to cluster formation rather than a direct determinant of ABI changes. Moreover, pre-examination conditions, such as resting duration before ABI measurement, may influence sympathetic tone and vascular resistance, affecting systolic pressure readings and ABI accuracy.²⁶ This challenge is particularly evident in patients with diabetes mellitus, who often present with autonomic dysfunction and impaired endothelial regulation.²⁷ Interventions aimed at improving metabolic balance, such as structured physical exercise, have been reported to enhance ABI values and peripheral vascular function.²⁸ However, such clinical implications cannot be directly derived from the clustering results of this study.

In contrast to visceral fat, subcutaneous fat showed minimal influence on ABI values, supporting evidence that subcutaneous adiposity does not significantly impair peripheral vascular function.²⁹ This observation reflects patterns within the dataset rather than definitive clinical conclusions. Instead, vascular impairment is more strongly associated with central fat distribution and related inflammatory mechanisms.^20,30 Emerging molecular evidence further suggests that adipokines, such as asprosin, may mediate endothelial dysfunction and contribute to PAD development.³¹

Skeletal muscle mass was also identified as a relevant parameter in ABI clustering. Decreased ABI values have been associated with muscle atrophy, myosteatosis, and reduced functional capacity, which collectively increase cardiovascular risk and impair mobility.^32,33 Reduced peripheral blood flow limits oxygen delivery and muscle regeneration, accelerating sarcopenia and functional decline.³⁴ However, within this study, skeletal muscle should be interpreted as part of a broader metabolic profile rather than an independent predictor of vascular status. These findings support the use of ABI not only as a vascular assessment tool but also as an indirect indicator of musculoskeletal health, although causal relationships cannot be established within the current analytical framework.³⁵

Cellular aging emerged as another important factor influencing ABI-based diagnostics. Vascular aging, characterised by endothelial senescence, reduced nitric oxide bioavailability, and increased arterial stiffness, contributes directly to ABI decline.³⁶ This process is accelerated in older adults and individuals with diabetes mellitus.³⁷ Regular ABI monitoring, combined with exercise-based nursing interventions such as Buerger Allen Exercise and diabetic foot exercises, has been shown to improve peripheral perfusion and ABI values.^38,39 However, the present study does not evaluate intervention effects and therefore cannot support causal or clinical recommendations.

Several limitations should be acknowledged. First, clustering is an unsupervised analytical approach that identifies patterns based on data similarity and does not establish diagnostic categories or predictive models. Second, the moderate Silhouette Scores observed in this study indicate that cluster separation is not strong, and some overlap between groups may exist. Third, this study utilised both primary and secondary data sources, which may introduce variability related to differences in data collection procedures. Fourth, external validation using independent datasets is required to confirm the robustness and generalisability of the identified clusters. Despite these limitations, the findings provide meaningful implications for exploratory understanding of metabolic and body composition profiles associated with ABI values. The results may contribute to hypothesis generation in nursing and vascular research, rather than direct clinical application. Therefore, clustering-based approaches should be considered as complementary analytical tools for pattern identification, not as standalone methods for diagnosis or clinical decision-making.

Conclusion

This study demonstrates that K-means clustering applied to metabolic and body composition parameters can identify distinct participant profiles in relation to ankle–brachial index (ABI) values. The findings highlight the relevance of variables such as visceral fat, resting metabolism, and cell age in differentiating these profiles. These clusters represent data-driven metabolic patterns and should be interpreted as descriptive associations rather than diagnostic or predictive categories. The results provide exploratory insights into metabolic characteristics related to peripheral perfusion and may contribute to hypothesis generation in future research. Further studies with external validation are required to confirm the robustness and generalisability of these findings.

Funding Statement

This research is funded by the Indonesian Endowment Fund for Education (LPDP) on behalf of the Indonesian Ministry of Higher Education, Science and Technology and managed under the EQUITY Program (Contract No.4305/B3/DT.03.08/2025 and No.03042/UN10/B/KS/2025).

Abbreviations

ABI, ankle–brachial index; BMI, body mass index; PAD, peripheral artery disease; EDA, exploratory data analysis; FS, feature selection.

Disclosure

The authors report no conflicts of interest in this work.