A Study on Risk Factors Associated with Gestational Diabetes Mellitus

Abstract

Background/Objectives: Gestational Diabetes Mellitus (GDM) is a global health issue with immediate and long-term maternal–fetal complications. Current diagnostic approaches, such as the Oral Glucose Tolerance Test (OGTT), have limitations in accessibility, sensitivity, and timing. This study aimed to identify key nodes and structural interactions associated with GDM using graph theory and network analysis to improve early predictive strategies. Methods: A literature review inspired by PRISMA guidelines (2004–2025) identified 44 clinically relevant factors. A directed graph was constructed using Python (version 3.10.12), and centrality metrics (closeness, betweenness, eigenvector), k-core decomposition, and a Minimum Dominating Set (MDS) were computed. The MDS, derived using an integer linear programming model, was used to determine the smallest subset of nodes with systemic dominance across the network. Results: The MDS included 20 nodes, with seven showing a high out-degree (≥4), notably Apo A1, vitamin D, vitamin D deficiency, and sedentary lifestyle. Vitamin D exhibited 15 outgoing edges, connecting directly to protective factors like HDL and inversely to risk factors such as smoking and obesity. Sedentary behavior also showed high structural influence. Closeness centrality highlighted triglycerides, insulin resistance, uric acid, fasting plasma glucose, and HDL as nodes with strong predictive potential, based on their high closeness and multiple incoming connections. Conclusions: Vitamin D and sedentary behavior emerged as structurally dominant nodes in the GDM network. Alongside metabolically relevant nodes with high closeness centrality, these findings support the utility of graph-based network analysis for early detection and targeted clinical interventions in maternal health.

1. Introduction

Health and well-being are part of the United Nations Sustainable Development Goals (SDGs) and Global Goals [1]. Gestational Diabetes Mellitus (GDM) has been recognized by the American Diabetes Association (ADA) as a distinct type of diabetes, with its own pathophysiological mechanisms that require specific approaches for diagnosis and treatment [2]. This condition, defined as glucose intolerance first detected during pregnancy, is associated with both short- and long-term complications for the mother and her offspring. These include an increased risk of developing Type 2 Diabetes Mellitus (T2DM), cardiovascular disease, childhood obesity, metabolic disorders, and reproductive dysfunction [3,4,5].

Globally, GDM is one of the most common pregnancy complications, with prevalence rates ranging from 5% to 25%, depending on the diagnostic criteria used and the characteristics of the population studied [6,7,8]. This variability underscores the need to develop more efficient and context-specific preventive and predictive strategies.

Several factors have been associated with the development of GDM, including obesity, advanced maternal age, family history, sedentary lifestyle, and excessive gestational weight gain [9,10], as well as biomarkers such as triglycerides [11,12,13], vitamin D [14,15] and iron metabolism parameters [13,16,17,18,19].

Currently, the Oral Glucose Tolerance Test (OGTT) is the primary diagnostic method, although it presents important limitations: it is administered late in pregnancy, is costly, has limited accessibility in certain regions, and may cause discomfort to the patient [20]. Moreover, its discriminative capacity, namely, its ability to distinguish between women with and without gestational diabetes, varies across populations, with Area Under the Curve (AUC) values ranging from 0.63 to 0.83 in different studies [21,22]. These limitations underscore the need for complementary approaches that allow the identification of key risk factors at earlier stages. In this context, network-based models offer an innovative alternative for exploring complex relationships among multiple clinical variables [23,24,25]. These models allow biomedical data to be represented as directed or weighted networks and enable the application of topological metrics, such as different forms of centrality (closeness, betweenness, and eigenvector), to identify key nodes within the system [26]. Unlike traditional statistical models or molecular approaches, network models allow the identification of interaction patterns that emerge from the relational structure of the data, beyond their spatial representation.

Although topological indices have been applied in ecology [27], social networks [28] and the analysis of complex systems [29], their use in the study of GDM remains incipient. This approach may contribute to a deeper understanding of the factors involved in the development of the disease and serve as a structural basis to complement existing tools, such as the OGTT or HOMA-IR (Homeostasis Model Assessment of Insulin Resistance), an index calculated from fasting glucose and insulin levels that estimates insulin resistance. Likewise, the findings may provide valuable input for the development of future clinical prediction models.

In this context, the aim of the present study is to identify key risk factors associated with the development of Gestational Diabetes Mellitus through the analysis of a directed network constructed using a network-based model and evaluated with topological metrics such as centrality, k-core, and the Minimum Dominating Set (MDS). This approach seeks to provide a structural basis for future strategies in clinical risk stratification. The article is organized as follows: first, the methodology used for network construction and analysis is described; next, the results of the topological analysis and the clinical interpretation of the most relevant nodes are presented; finally, the implications of the findings and their potential applications in clinical and predictive contexts are discussed.

2. Materials and Methods

Although the present study does not constitute a systematic review in the strict sense, a structured search strategy inspired by the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines were adopted to ensure a transparent and reproducible selection of scientific literature. The search was conducted in the PubMed-MEDLINE, Scopus, Web of Science, ClinicalTrials.gov, and Google Scholar databases, covering the period from 2004 to March 2025. Search terms included “gestational diabetes mellitus,” “type 2 diabetes mellitus,” “correlation,” and “association,” combined with the clinical factors considered in this study. Each factor was assigned an alphabetical label (a, b, c…) used consistently throughout the network construction, figures, and topological analyses to facilitate cross-reference and node identification. Throughout the manuscript, node labels are set in italics (e.g., u, o, c, u_1) to avoid ambiguity, whereas variable names (e.g., vitamin D, sedentary behavior, Apo A1) are written in roman text. In the tables, the factors are displayed according to their category: genetic and environmental (

Table 1. Genetic and environmental factors included in the network model and their correlation with Gestational Diabetes Mellitus (GDM).

Label	Study Variable	Correlation with GDM	Correlation with Other Variables	Reference
a	History of GDM	1	x	[30,31,32,33,34,35,36,37,38,39]
r	Smoking	1	x	[40,41,42,43]
o	Sedentary Lifestyle	1	x	[44,45,46]
b	Physical Activity during Pregnancy	2	x	[47,48,49,50,51,52]
m	African Women	1		[53,54,55]
m_2	Asian Women	1	x	[54,55,56,57,58,59]
m_3	Latin American Women	1		[53,54]
t	Hepatitis B Virus (HBV)	1	x	[60,61,62,63,64,65]

Note. The “Correlation with GDM” column indicates the type of significant correlation observed: a value of 1 corresponds to a direct correlation and 2 to an inverse correlation. The “x” mark in the “Correlation with other variables” column indicates that the variable also showed significant correlations with at least one other variable in the model, suggesting its potential involvement in relevant systemic interactions. Abbreviations: GDM, Gestational Diabetes Mellitus.

), metabolic (

Table 2. Metabolic factors included in the network model and their correlation with Gestational Diabetes Mellitus (GDM).

Label	Study Variable	Correlation with GDM	Correlation with Other Variables	Reference
e	Uric Acid (UA)	0	x	[66,67,68,69,70,71]
e_3	Uric Acid—Quartile 3	1		[67,70,71,72,73]
e_4	Uric Acid—Quartile 4	1	x	[70,71,72,73,74]
d	Apolipoprotein B (Apo B)	1	x	[12,75,76]
c	Apolipoprotein A1 (Apo A1)	1	x	[12,30,34,77,78,79,80,81,82,83]
f	Maternal Age	1		[34,64,78,80,83,84,85,86,87,88,89,90]
f_1	Maternal Age < 35 years	1	x	[54,83,85,86]
f_2	Maternal Age ≥ 35 years	1	x	[30,64,78,79,80,82,84,86,88,89,91]
g	Ferritin	1	x	[13,16,18,82,92,93,94,95,96,97,98,99,100]
g_1	Ferritin—Quartile 1	0	x	[16,97,101]
g_2	Ferritin—Quartile 2	1	x	[13,16,18,82,101,102,103,104,105]
g_3	Ferritin—Quartile 3	1	x	[13,16,18,82,92,95,101,104,106,107]
g_4	Ferritin—Quartile 4	1	x	[13,18,92]
h	Fasting Plasma Glucose	1	x	[34,98,108,109,110,111,112,113]
i	Gestational Weight Gain (GWG)	1	x	[34,80,81,85,88,114,115,116,117,118]
i_1	GWG < IOM recommendations	0	x	[88,117]
i_2	GWG > IOM recommendations	1	x	[80,88,114,115,118]
k	Sleep Hours	2	x	[119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136]
l	Body Mass Index (BMI)	1	x	[53,81,82,83,85,87,107,115,116,135]
l_1	BMI—Underweight			[53,116]
l_2	BMI—Normal			[116,136,137]
l_3	BMI—Overweight	1	x	[14,53,78,81,82,85,107,116,136,137,138,139]
l_4	BMI—Obesity	1	x	[14,53,64,81,88,108,115,116,135,136,137,139,140,141,142,143,144,145,146,147]
j	High-Density Lipoprotein (HDL)	2	x	[12,69,77,148,149,150,151,152]
n	Insulin Resistance	1	x	[15,153,154,155,156]
s	Triglycerides (Tg)	1	x	[11,12,13,34,76,77,81,84,87,95,107,109,148,151,157,158,159,160,161,162,163,164,165,166]
s_1	Tg—Range 1			[148]
s_2	Tg—Range 2			[95,148,159]
s_3	Tg—Range 3	1	x	[76,84,95,107,148,159,161,163]
s_4	Tg—Range 4	1	x	[148,159,161,162]
s_5	Tg—Range 5	1	x	[148,159,167]
u	Vitamin D (Vit D)	2	x	[14,15,87,91,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183,184,185,186,187,188,189,190,191,192,193,194,195]
u_1	Vitamin D Deficiency	1	x	[14,173,177,185,196]
u_2	Vitamin D Insufficiency	1	x	[176,184,185]
u_3	Vitamin D Sufficiency	2	x	[184,185]

Note. The “Correlation with GDM” column indicates the type of significant correlation observed: 1 = direct correlation, 2 = inverse correlation, 0 = no significant correlation. An “x” in the “Correlation with other variables” column indicates that the variable showed significant correlations with at least one other variable included in the model, suggesting its potential involvement in relevant systemic interactions. Abbreviations: GDM, Gestational Diabetes Mellitus.

), and fetal sex (

Table 3. Fetal sex and its role in the network model of gestational diabetes.

Label	Study Variable	Correlation with GDM	Correlation with Other Variables	Reference
q	Male Fetal Sex (SF m)	1	x	[116,197,198]
p	Female Fetal Sex (SF f)	0	x	[116,198]

Note. The “Correlation with GDM” column indicates the type of significant correlation observed: a value of 1 corresponds to a direct correlation, and 0 indicates no significant correlation with GDM. An “x” in the “Correlation with other variables” column indicates that the variable showed significant correlations with at least one other variable included in the model, suggesting its potential involvement in relevant systemic interactions. Abbreviations: GDM, Gestational Diabetes Mellitus.

).

The retrieved articles were reviewed comprehensively, with priority given to studies conducted in human populations, particularly in women of reproductive age or mixed populations. Included studies comprised observational designs (cross-sectional and cohort), randomized controlled clinical trials, systematic reviews, meta-analyses, and narrative reviews deemed relevant. The selection process is summarized in Figure 1 using a flow diagram adapted to the PRISMA style.

Please note that the 277 records refer to unique publications, not to individual patients. The network analysis was constructed from study-level correlations rather than pooled patient data. The selection of factors was based on risks recognized by the American Diabetes Association (ADA) and the World Health Organization (WHO), as well as on biochemical-nutritional markers commonly used in clinical practice. In addition, emerging components with potential predictive value, identified in recent scientific literature, were incorporated.

Initially, 66 clinical or risk factors were identified, from which 44 variables were selected based on previously established criteria for data completeness and theoretical or statistical relevance. Additionally, only those variables that showed at least one direct correlation with Gestational Diabetes Mellitus and, simultaneously, at least one correlation with another component of the set were included, to ensure their functional integration within the system. These variables were grouped according to their etiological origin and mechanism of action, including genetic and environmental factors, metabolic factors, and fetal sex characteristics, to avoid redundancies and optimize analytical interpretation.

Each of the 44 selected variables was classified according to the evidence of statistical correlation with the diagnosis of GDM reported in the reviewed studies. For coding purposes, a value of 1 was assigned to significant direct correlations, a value of 2 to significant inverse correlations, and a value of 0 when no significant association was identified. Additionally, variables showing significant correlations with other members of the set were marked with an “x”. This coding is summarized in Table 1, Table 2 and Table 3, which also define the nodes used in the network model.

The constructed network corresponds to an unweighted directed network, in which each node represents one of the variables included in the system. Directed edges between nodes were established when a statistically significant correlation (p < 0.05) was observed between two variables, following a functional influence logic. The direction of each edge reflects a potential functional relationship, based on the reviewed literature and theoretical principles regarding the pathophysiology of GDM. This relationship is represented by a directed connection from the source node (factor with a potential explanatory role) to the target node (influenced factor), without implying a direct causal relationship.

This structure allowed the analysis of node connectivity and topological relevance using different metrics, as detailed in the following section.

In the second stage, the GDM network was constructed as an unweighted, directed network comprising 44 nodes, with oriented connections determined by statistically significant correlations reported in the literature. The network was derived from a binary correlation matrix that included only associations with a p-value < 0.05 and a functionally supported direction based on theoretical evidence.

In this model, each node represents a clinical, biochemical, or behavioral component, while each connection indicates a functionally oriented relationship between two nodes, directed from the source node (component with a potential explanatory role) to the target node (influenced component). The network was subsequently represented and implemented using computational algorithms in the Python programming language, enabling its visualization and topological analysis.

From the constructed directed network model (GDM network), four complementary structural analysis approaches were applied:

• Minimum Dominating Set (MDS): A minimal subset of nodes was identified with the structural capacity to reach the rest of the network through directed paths. This set allows the inference of which clinical components may play a key role in regulating or activating functional pathways within the system.
• Topological centralities: Three centrality metrics were calculated to assess the structural relevance of each node, following the definitions proposed by Jordán and Scheuring [26]:

  ○

  Closeness, which evaluates how close a node is to the rest of the network, reflecting its potential for rapid and direct influence.

  ○

  Betweenness, which estimates a node’s capacity to act as a bridge between pairs, potentially indicating modulatory components within the system.

  ○

  Eigenvector, which considers both the number of connections and the structural importance of the connected nodes, allowing the identification of nodes with global influence.

Based on the values obtained, a tertile classification was applied instead of quartiles, categorizing the results as high (>66.6%), medium (33.3–66.6%), and low (<33.3%). This methodological decision was made in response to the high concentration of low or null values, a common characteristic in directed networks, which limits the discriminative capacity of the metrics when quartiles are used.

• 7-core subnetwork: We applied a k-core analysis to detect the most densely interconnected region of the GDM network. In this approach, nodes with fewer than k internal connections are progressively removed until only nodes with degree ≥ k remain, forming the k-core. In our analysis, the highest cohesive level identified was k = 7; when higher thresholds (k ≥ 8) were applied, the network fragmented, and no cohesive subgraph persisted. In this 7-core subnetwork, each node is directly connected to at least seven other members of the subnetwork. This structure highlights the most densely interconnected region of the network, representing a functional “core” that offers insight into the most tightly integrated risk factors for GDM.
• Topological overlaps: Finally, nodes with a prominent presence in more than one metric (e.g., high centrality plus membership in the MDS and/or the 7-core subnetwork) were identified, with the aim of recognizing key components within the system that could play a structurally relevant role. Detailed statistical coefficients and node-level classification values derived from these analyses are provided in Appendix A (

  Table A1. Table of Statistical Coefficients.

  Statistical Coefficient	Intervals	References
  Odds Ratio (OR)	Values >1: increase in the probability of the event.	[243]
  	Values <1: decrease in the probability
  Pearson Correlation Coefficient (r)	Values from −1 to 1:	[244]
  	1 indicates a perfect positive correlation, −1 indicates a perfect negative correlation, 0 indicates no correlation.
  Spearman Correlation Coefficient (ρ)	Values from −1 to 1:	[245]
  	Same as Pearson, but for non-parametric variables.
  Rate Ratio (RR)	Values >1: higher rate of occurrence in the exposed group.	[246]
  	Values <1: lower rate of occurrence in the exposed group.
  Beta Coefficient (β)	Its value depends on the magnitude and direction of the relationship between the variables. A negative β value indicates an inverse relationship, while a positive value indicates a direct relationship.	[247]

  and

  Table A2. Outdegree and indegree according to closeness centrality.

  Nodes *	Closeness Centrality	Outdegree	Indegree	Classification
  e	0.55536272	1	16	High
  n	0.54719562	4	19	High
  s	0.52407468	5	16	High
  h	0.49612403	5	16	High
  j	0.47704234	3	11	High
  t	0.41808205	5	5	High
  g	0.41343669	9	5	High
  l_4	0.41343669	8	8	High
  u	0.37209302	15	4	High
  l_3	0.37209302	3	4	High
  d	0.36479708	2	2	High
  f	0.35103115	1	5	High
  s_5	0.34136975	1	5	High
  q	0.32928586	1	2	High
  i	0.30251465	3	5	High
  p	0.29298663	1	1	Medium
  k	0.2884442	11	4	Medium
  l	0.2862254	2	3	Medium
  r	0.27359781	2	1	Medium
  f_2	0.26963263	2	1	Medium
  l_2	0.2443314	0	2	Medium
  m	0.22688599	2	2	Medium
  s_1	0.0372093	0	1	Medium
  s_2	0.0372093	0	1	Medium
  s_3	0.0372093	4	1	Medium
  s_4	0.0372093	2	1	Medium
  c	0.03488372	4	1	Medium
  b	0.03100775	8	1	Medium
  u_1	0.02325581	7	1	Low
  o	0.02325581	10	1	Low
  a	0	1	0	Low
  e_3	0	1	0	Low
  e_4	0	2	0	Low
  f_1	0	1	0	Low
  g_1	0	1	0	Low
  g_2	0	1	0	Low
  g_3	0	4	0	Low
  g_4	0	3	0	Low
  i_1	0	1	0	Low
  i_2	0	3	0	Low
  m_2	0	2	0	Low
  m_3	0	2	0	Low
  u_2	0	1	0	Low
  u_3	0	1	0	Low

  * The nodes are ordered from highest to lowest.

  ).

Formally, let $D=(V,E)$ be a directed network, where $V$ is the set of vertices (nodes) and $E$ the set of directed edges. For a subset $S\subseteq V$, the set $N^{+}\left(S\right)$, called the set of external neighbors (or out-neighborhood) of $S$, is defined as the set of all vertices outside $S$ that receive a directed edge from some vertex within $S$, that is,

$$N^{+}\left(S\right)=\{v\in V\setminus S\mid \exists u\in S such that \left(u\to v\right)\in E\}$$

A set $S$ of vertices in a directed network $D=(V,E)$ is a dominating set if, for each vertex $v \in V\setminus S$, there exists a vertex $u \in S$ such that $u \to v \in E$; that is, every vertex in $V\setminus S$ has an incoming connection from a vertex in $S$. In other words, $S$ is a dominating set of $D$ if $V\setminus S \subseteq N^{+}\left(S\right)$. The cardinality (i.e., the number of elements) of the smallest dominating set of a directed network $D$ is called the domination number and is denoted by $\gamma \left(D\right)$.

An integer linear programming (ILP) model was used to identify the MDS within the GDM network, thereby ensuring the structural coverage of all nodes from the smallest possible number of vertices. The following describes the mathematical formulation used to solve this optimization problem.

Given a directed network $D=(V,E)$ of order $n$, the adjacency matrix of $D$, denoted by $A = {\left(a_{ij}\right)}_{n\times n}$, is defined as follows: $a_{ij}=1 if \left(i\to j\right)\in E$, and $a_{ij}=0$ otherwise. Given a dominating set $S$ of cardinality $\gamma \left(D\right)$, and for each $i\in V\left(D\right)$, the following decision variables are defined:

$$x_i=\left\{\begin{array}{cc}1,& if \mathrm{i}\in \mathrm{S}\\ 0,& \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\end{array}\right.$$

(1)

The dominating set problem can be formulated as the following integer optimization problem

$${\sum }_{i=1}^n{x}_i=\mathsf{\gamma }\left(D\right)$$

(2)

Subject to

$${\sum }_{i=1}^n{a}_{ij}\ast x_i\ge 1,\forall j\in V\left(D\right)$$

(3)

The objective function (2) minimizes the size of the dominating set. The $n$ constraints of type (3) ensure that, for each vertex $j\in V\setminus S$, at least one of its neighbors belongs to the set $S$; in other words, all vertices not in $S$ are dominated by some vertex in the selected set.

The minimum dominating set (MDS) is defined as the smallest subset of nodes that can directly reach all other elements in the network. This property enables a structural representation of system coverage and has been previously applied in the analysis of complex networks. In the present study, the MDS was used as a criterion to select a representative set of nodes based on their topological position within the network.

The complete adjacency matrix, as well as the Python scripts used for constructing the network and calculating topological metrics, are publicly available in the Mendeley Data repository: https://doi.org/10.17632/s92w2sb5ng.1, accessed on 14 October 2025. There are no restrictions on their access or reuse. The development and verification of the codes were carried out through assisted programming and thorough review by the research team. Algorithms were implemented for calculating centrality metrics (closeness, betweenness, eigenvector) [199,200] and for obtaining the MDS through integer linear programming [201], adjusted according to criteria reported in the literature. All analyses were automated under the direct supervision of the responsible researchers.

3. Results

3.1. Network Model Construction

The network is defined as follows: each node represents one of the 44 selected factors, grouped according to their genetic, environmental, metabolic, or fetal sex-related origin. These factors were operationalized as clinical, biochemical, or behavioral indicators, all of them associated with the diagnosis of GDM (see Table 1, Table 2 and Table 3). Directed edges between nodes were established when a statistically significant correlation (p-value < 0.05) was observed between two variables, following a functional influence rationale. The direction of the edge indicates which variable acts as a potential explanatory factor (source) and which is the influenced variable (target), according to its role in pathophysiology. Additionally, the type of correlation was color-coded: green for positive (direct) correlations and blue for negative (inverse) correlations. This coding allows visualization not only of the existence of relationships but also of their potential functional implications within the system.

3.1.1. Minimum Dominating Set

Figure 2 presents an MDS of the GDM network. This set is composed of 20 nodes (in red), from which the remaining nodes (in light blue) can be reached through direct correlations (green) or inverse correlations (blue).

Given that the above MDS satisfies the dominance property, it ensures complete structural coverage of the network. In this way, MDS analysis allows the identification of strategic clinical components with the capacity for systemic influence, representing key points from which it is possible to access or indirectly modulate multiple other functional elements of the system. This approach is particularly useful in complex systems such as the pathophysiology of GDM, where risk factors interact in a multifactorial and non-linear manner.

From an applied standpoint, the nodes within the MDS constitute priority candidates for clinical evaluation and for the development of simplified predictive models, as they preserve the system’s explanatory power while relying on a reduced set of key clinical factors. This property not only improves diagnostic efficiency but may also lower operational and logistical costs, particularly in resource-constrained settings.

Within the MDS, the nodes with the highest out-degree, c (Apo A1), u (vitamin D), and o (sedentary lifestyle), were identified. Their high connectivity suggests a cross-cutting regulatory role within the network, exerting structural influence over multiple clinical and biochemical factors.

3.1.2. Closeness Centrality

Figure 3 shows the GDM network coded according to closeness centrality values. This metric reflects the structural proximity of each node to the rest of the network, indicating its potential capacity to efficiently influence or receive information.

The nodes with the highest closeness centrality were e (uric acid), n (insulin resistance), and s (triglycerides), followed by 12 additional nodes classified as high (

Table 4. Classification of nodes in the GDM network by closeness centrality.

Centrality	Nodes	Variable	Out-Degree	In-Degree
High	e	Uric Acid	1	16
	n	Insulin Resistance	4	19
	s	Triglycerides	5	16
	h	Fasting Plasma Glucose	5	16
	j	HDL	3	11
	t	HBV	5	5
	g	Ferritin	9	5
	l_4	BMI—Obesity	8	8
	u	Vitamin D	15	19
	l_3	BMI—Overweight	3	4
	d	Apo B	2	2
	f	Maternal Age	1	5
	s_5	Tg—Range 5	1	5
	q	Male Fetal Sex	1	2
	i	Gestational Weight Gain	3	5

Note. Nodes with high closeness centrality are shown, together with their corresponding variable, number of outgoing connections (out-degree), and number of incoming connections (in-degree) in the directed network. Node codes correspond to the variables described in Table 1, Table 2 and Table 3. Nodes with medium or low closeness centrality are provided in Appendix A.

). These findings indicate that, on average, these nodes are positioned at the shortest possible distance from the rest of the system, making them potential entry or diffusion points from which influence can propagate more rapidly and extensively.

In this context, early detection of changes in the clinical components represented by these nodes could anticipate functional alterations in other regions of the network, reinforcing their value as highly relevant clinical indicators. Among the high-closeness nodes, u (vitamin D), g (ferritin), and l_4 (BMI obesity) were notable for having the greatest number of outgoing connections (15, 9, and 8, respectively), suggesting a strong capacity to affect and influence other nodes.

Conversely, nodes n (insulin resistance), e (uric acid), s (triglycerides), h (fasting glucose), and j (HDL) exhibited the highest numbers of incoming connections (19, 16, 16, 16, and 11, respectively), indicating high receptivity to influence within the system. This property positions them as nodes particularly sensitive to network dynamics, making them priority candidates for predicting the clinical risk associated with GDM development.

Nodes classified with medium or low closeness centrality are listed in Appendix A, together with their respective in-degree and out-degree values.

3.1.3. Betweenness Centrality

Figure 4 presents the GDM network coded by betweenness centrality values, a metric that quantifies how often a node serves as a transit point along the network’s shortest paths. Nodes with higher values have a greater ability to mediate or modulate communication between other components of the system.

In this visualization, the scale is restricted (0 to 0.16), likely due to the network’s densely connected structure. This configuration reduces the need for intermediary paths, as multiple direct connections exist between nodes. Consequently, no node was classified as high in this metric, with only medium and low levels being observed (Appendix A).

In this analysis, the node with the highest betweenness centrality was u (vitamin D), followed by 13 other nodes classified at the medium level (

Table 5. Classification of nodes within the GDM network according to their betweenness centrality.

Centrality	Nodes	Variable	Out-Degree	In-Degree
Moderate	u	Vitamin D	15	4
	n	Insulin Resistance	4	19
	s	Triglycerides	5	16
	h	Fasting Plasma Glucose	5	16
	l_4	BMI—Obesity	8	8
	j	HDL	3	11
	t	HBV	5	5
	g	Ferritin	9	5
	k	Sleep Hours	11	4
	i	GWG	3	5
	d	Apo B	2	2
	e	Uric Acid	1	16
	u_1	Vitamin D Deficiency	7	1
	o	Sedentary Lifestyle	10	1
	s_3	Tg—Range 3	4	1

Note. Nodes with medium betweenness centrality are shown, together with their corresponding variable, number of outgoing connections (out-degree), and number of incoming connections (in-degree) in the directed network. Node codes correspond to the variables described in Table 1, Table 2 and Table 3. Nodes with low betweenness centrality are provided in Appendix A.

). The structural position of vitamin D suggests a role as a functional intermediary, facilitating connections between regions of the network that would otherwise be less connected. Additionally, vitamin D exhibited the highest out-degree (15 connections), whereas n (insulin resistance), also with medium betweenness, showed the highest in-degree (19 connections). This indicates that while u can influence multiple other clinical components and n can be affected by various factors, both nodes operate with high direct activity. Nevertheless, their value as topological articulators of the overall network flow is limited. Consequently, although they are functionally relevant, they do not behave as key intermediary nodes within the global structure of the system.

3.1.4. Eigenvector Centrality

Figure 5 shows the GDM network coded according to eigenvector centrality values, represented using a red color scale. This metric identifies the most influential nodes within the network by considering not only the number of direct connections each node has, but also the structural relevance of its neighbors. In other words, a node with high eigenvector centrality is connected to other nodes that are themselves highly connected, positioning it within functional clusters with strong topological integration.

In this visualization, 15 nodes were classified as having a high level of eigenvector centrality (

Table 6. Classification of nodes in the GDM network by eigenvector centrality.

Centrality	Nodes	Variable	Out-Degree	In-Degree
High	e	Uric Acid	1	16
	j	HDL	3	11
	n	Insulin Resistance	4	19
	s	Triglycerides	5	16
	h	Fasting Plasma Glucose	5	16
	t	HBV	5	5
	g	Ferritin	9	5
	l_4	BMI—Obesity	8	8
	u	Vitamin D	15	4
	d	Apo B	2	2
	f	Maternal Age	1	5
	l_3	BMI—Overweight	3	4
	s_5	Tg—Range 5	1	5
	q	Male Fetal Sex	1	2
	i	Gestational Weight Gain	3	5

Note. Nodes with high eigenvector centrality are shown, together with their corresponding variable, number of outgoing connections (out-degree), and number of incoming connections (in-degree) in the directed network. Node codes correspond to the variables described in Table 1, Table 2 and Table 3. Nodes with low eigenvector centrality are provided in Appendix A.

). Among them, e (uric acid), j (HDL), n (insulin resistance), and s (triglycerides) stand out for exhibiting the most intense shades, reflecting particularly high structural influence within the system. These nodes are in densely connected regions of the network, acting as points with high potential for amplification and systemic impact.

An additional relevant finding was that the nodes s (triglycerides), n (insulin resistance), and h (fasting glucose), in addition to having high eigenvector centrality, exhibited the highest in-degree values. This indicates that they receive direct influence from a large number of other nodes in the system, reinforcing their role as key receivers in the flow of biomedical information.

In contrast, the node corresponding to vitamin D (u) stood out for having the highest out-degree among all nodes with high eigenvector centrality. This characteristic suggests that vitamin D exerts direct influence over multiple relevant clinical components, positioning it as a high-hierarchy emitter node within the network.

The complete classification of nodes with medium or low eigenvector centrality, together with their in-degree and out-degree values, is provided in Appendix A.

3.1.5. Structural Core According to 7-Core GDM Network

Figure 6 shows a 7-core GDM network, comprising 10 nodes that, by definition, have at least seven internal connections with other members of the same group. This type of analysis makes it possible to identify the most densely interconnected core of the network, suggesting the existence of a functional cluster of clinical components that participate in multiple interactions potentially relevant to the pathophysiology of GDM.

k = 7 was selected as the highest value that preserves a cohesive and functional structure. Higher values, such as k = 8 or k = 9, substantially reduce the number of nodes, thereby limiting the detection of relevant patterns. Thus, this 7-core network represents an optimal balance between connection density and the inclusion of representative nodes, allowing the identification of functional groupings with high potential for clinical and pathophysiological interaction.

Of these 10 nodes, only two overlapped with the MDS: u (vitamin D) and o (sedentary lifestyle). This overlap is significant, as it indicates that both nodes exhibit dual topological relevance: on one hand, they are part of the minimum set from which the entire network can be reached (structural dominance), and on the other, they are integrated within the most cohesive core of relationships (high functional density).

From a clinical perspective, this dual presence underscores the strategic role of vitamin D and sedentary lifestyle within the analyzed system. Vitamin D, in addition to directly influencing multiple metabolic components, actively participates in the functional core of the network, reinforcing its value as a modulatory marker and a potential intervention point in processes such as insulin resistance, lipid metabolism, and inflammation.Sedentary lifestyle, in turn, not only has the capacity to exert a transversal impact on other factors within the system but is also closely linked to components with high clinical weight, such as obesity, dyslipidemia, and glucose metabolism. Its dual role in the network positions it as a highly modifiable risk factor with systemic impact, and therefore, a priority target for preventive actions.

On the other hand, the node g (ferritin, continuous value) was located within the 7-core network, although it was not identified as dominant in the MDS. In contrast, the latter included the nodes g_3 and g_4 (ferritin in higher quartiles), suggesting that elevated ferritin values have greater structural reach, whereas the continuous value participates in areas of high interconnectivity. This distinction highlights the importance of considering both absolute levels and clinically critical ranges in the analysis of biomarkers.

Taken together, the 7-core analysis makes it possible to identify nodes that, in addition to being integrated into the functional core of the system, represent clinically relevant factors due to their systemic influence and intervention potential. The partial overlap with the MDS highlights that both approaches provide complementary information: while the MDS identifies an efficient set for minimal coverage, the 7-core subnetwork reveals dense nuclei of metabolic interaction that may be key to understanding multifactorial processes such as gestational diabetes.

Finally,

Table 7. Topological convergence of 7-core nodes with centrality measures and structural dominance.

Nodes	Closeness ↑	Betweenness ↑	Eigenvector ↑	MDS	7-Core
Triglycerides (s)	✓	–	✓	x	✓
Uric Acid (e)	✓	–	✓	x	✓
BMI—Obesity (l4)	✓	–	✓	x	✓
HDL (j)	✓	–	✓	x	✓
Insulin Resistance (n)	✓	–	✓	x	✓
Fasting Glucose (h)	✓	–	✓	x	✓
Vitamin D (u)	✓	–	✓	✓	✓
Sedentary Lifestyle (o)	x	–	x	✓	✓
Sleep Hours (k)	x	–	x	x	✓
Ferritin (g)	✓	–	✓	x	✓

Note. The table shows the nodes belonging to the 7-core subnetwork and their overlap with other topological metrics: high centralities (closeness, betweenness, and eigenvector) and membership in the Minimum Dominating Set (MDS). The arrows (↑) indicate nodes with high values in the corresponding centrality metric (closeness, betweenness, or eigenvector). The ✓ symbol indicates a notable presence in each category, x indicates absence, and – is used when the metric was not applicable or did not reach a representative value. Inclusion in the 7-core subnetwork reflects a high degree of internal interconnectivity among nodes, whereas the MDS identifies those with the structural capacity to reach the rest of the network. Abbreviations: MDS, Minimum Dominating Set.

presents the convergence of key nodes identified through different topological approaches: high centralities, the MDS, and the 7-core subnetwork. For this comparison, the subset of nodes belonging to the 7-core subnetwork was taken as a reference and contrasted with those previously classified as having high topological centrality (betweenness, closeness, and eigenvector), as well as with the dominant nodes of the MDS.

This analysis makes it possible to identify nodes that perform multiple roles within the network, for example, those that are not only densely connected within the 7-core subnetwork but also structurally dominate the network or exhibit high levels of centrality. The convergence of these different approaches reinforces their systemic relevance and suggests that they may constitute critical points of metabolic regulation, with potential diagnostic or therapeutic value.

4. Discussion

The network analysis identified a Minimum Dominating Set (MDS) comprising 20 nodes, from which it is structurally possible to reach all other nodes representing clinical and biochemical components. This property of dominance positions these nodes as strategic control points within the GDM-related system.

Seven nodes within the MDS exhibited more than four outgoing connections, suggesting a high capacity to influence other system components. These emitter nodes were c (Apo A1), u (vitamin D), g_4 (ferritin quartile 4), u_1 (vitamin D deficiency), g_3 (ferritin quartile 3), o (sedentary lifestyle), and i_2 (GWG > IOM). Beyond their structural relevance, these variables are supported by the literature for their involvement in pathophysiological mechanisms associated with GDM and other clinically significant outcomes.

• Apo A1 (c), the main component of high-density lipoproteins, has been shown to improve pancreatic β-cell function and increase insulin sensitivity. It has also been associated with benefits in cardiovascular diseases, neurological disorders, thrombotic processes, and oncogenesis [199].
• Vitamin D (u) acts across multiple systems, cardiovascular, musculoskeletal, immune, endocrine, and neurological, and is involved in fundamental genetic and epigenetic mechanisms essential for metabolic homeostasis [200,201]. Its deficiency (u_1) has been associated with an increased risk of chronic diseases linked to oxidative stress, such as insulin resistance, osteoporosis, cognitive decline, and osteomalacia [200].
• Elevated ferritin levels (g_3 and g_4) during pregnancy have been linked to an increased risk of GDM. This association is attributed to iron overload, which promotes systemic inflammation, oxidative stress, and β-cell dysfunction [202]. Additionally [202], elevated ferritin levels have been implicated in neurodegenerative diseases, cellular aging, and oncogenic transformation processes [19,203].
• Sedentary lifestyle (o) is an independent risk factor for GDM and other chronic diseases. Studies show that sitting for more than 5 h per day is associated with an increased risk of cardiovascular disease (OR 1.90), respiratory disease (OR 1.61), and multimorbidity (OR 2.80) [204]. Even shorter periods (>3 h/day) increase the risk, with physical activity failing to attenuate these associations [205].
• Excessive gestational weight gain (i_2) is associated with a higher risk of developing GDM, particularly when it occurs during the first or second trimester of pregnancy [114], and with adverse effects on the newborn’s health [206].

Notably, node u (vitamin D) exhibited the highest number of outgoing edges (15), followed by node o (sedentary lifestyle) with 10. This configuration suggests that both nodes function as structural emitters, from which trajectories propagate toward other components of the system. In the case of vitamin D, 11 edges were inverse, targeting variables such as smoking [207], obesity [183], ferritin [179], glucose [91], maternal age [182], uric acid [174], insulin resistance [188], and triglycerides [185], indicating negative associations with metabolic and cardiovascular risk factors. The four direct edges were directed toward HDL, gestational weight gain (GWG), hours of sleep, and male fetal sex, reflecting associations with protective or modulatory variables.

Several studies have reported that higher maternal vitamin D levels are correlated with increased HDL concentrations [186,190] and influence gestational weight gain (GWG), both directly and through mediation by other clinical factors associated with metabolism and nutritional status [191], as well as improving sleep quality [208]. Moreover, vitamin D may reduce the risk of sleep disorders by up to 22% for each unit increase [209]. On the other hand, its metabolism and transport may vary according to fetal sex, which could explain its differential impact on maternal health [169] and fetal immune function [170].

Regarding sedentary behavior, seven direct edges linked it to triglycerides, glucose, uric acid, BMI, and maternal age, while three inverse edges connected it to physical activity during pregnancy, sleep duration, and ferritin. These patterns suggest positive associations with cardiometabolic risk factors [210,211] and inflammation [212], and negative associations with protective factors [213,214].

From a mathematical perspective, both vitamin D and sedentary behavior occupied hierarchical positions in the network as emitter nodes and were part of the 7-core subnetwork, comprising the 10 most interconnected nodes. This overlap suggests that both actively participate in the functional core of the system.

Their strategic positioning indicates that Apo A1, vitamin D, and sedentary behavior are not only individually associated with GDM risk but may also contribute to the systemic modulation of functional axes such as lipid metabolism, nutritional status, and lifestyle behaviors. This observation highlights the importance of incorporating both biochemical and behavioral factors into predictive models, particularly those that, like sedentary behavior, are highly modifiable. From an intervention standpoint, these nodes represent critical targets for clinical action, with the potential to trigger cascading effects on other risk factors, thereby reinforcing their value in early evaluation and in the design of more effective preventive strategies.

In addition to the MDS and 7-core analyses, closeness centrality is considered particularly useful in predictive models [215]. This metric identifies nodes with high structural accessibility, those capable of influencing multiple system components through short and efficient paths. Such a property facilitates the anticipation of systemic effects from individual clinical variables, which is critical in clinical contexts requiring early detection or timely intervention.

On the other hand, eigenvector centrality also provides valuable insights by highlighting nodes connected to other highly influential ones. However, as it focuses on the structural hierarchy of the network, its usefulness lies more in analyzing the relational positioning within the system than in directly predicting functional propagation processes [216,217]. This type of application has been reported, for example, in the study by Binnewijzend et al., where voxel-to-voxel eigenvector centrality maps were used to identify relevant brain regions in patients with Alzheimer’s disease, linking these patterns to biomarkers and cognitive decline.

In this analysis, 15 nodes exhibited high closeness and eigenvector centrality, including e (uric acid), n (insulin resistance), s (triglycerides), and j (HDL). Integrating the clinical components represented by these nodes into predictive models could enhance diagnostic sensitivity from the early stages.

In particular, insulin resistance is characterized by a diminished response of target tissues to the action of insulin. However, its clinical definition still presents limitations, as no universally accepted diagnostic test exists [218]. The choice of the index used to assess this condition significantly influences both the diagnostic strategy and the construction of predictive models for GDM. Although the hyperinsulinemic-euglycemic clamp is considered the gold standard, its technical complexity has favored the use of more accessible tools, such as HOMA-IR, HOMA2, QUICKI, and the triglyceride/glucose and triglyceride/HDL ratios.

For example, Ilayu et al. reported that the triglyceride/glucose ratio showed remarkable discriminatory ability for detecting insulin resistance (AUC: 0.92; sensitivity: 0.90; specificity: 0.79) [219]. Complementarily, Paracha et al. compared different indices against the OGTT, considering the clinical standard [153]. In their study, both HOMA-IR and QUICKI demonstrated good discriminative power, measured by the AUC, a metric that reflects a test’s ability to correctly distinguish between healthy and diseased individuals. An AUC of 1.0 indicates perfect accuracy, while a value of 0.5 reflects no discriminative capacity, similar to chance.

The HOMA-IR index, with a cut-off point ≤ 2, achieved a sensitivity of 94.5% and an AUC of 0.913, positioning it as a useful tool for screening or early detection by minimizing the risk of false negatives. Conversely, the QUICKI index, with a cut-off point of 0.34, showed a better balance between sensitivity (86.4%) and specificity (83.3%), with an AUC of 0.905, making it suitable for confirmatory use or clinical follow-up.

Several studies have reported the applicability of HOMA-IR during pregnancy, noting that its increase in early stages may represent a risk factor for the development of GDM [220,221]. In the Mexican population, trimester-specific reference values of HOMA-IR have even been proposed to define insulin resistance: first trimester (≥1.6), second trimester (≥2.9), and third trimester (≥2.6) [222]. In this context, its performance could be optimized by integrating it with other central variables within the network, such as uric acid, triglycerides, fasting glucose, and HDL, all of which were identified with high closeness centrality in the present analysis.

For example, Hou et al. reported that triglyceride levels during the first and second trimesters exhibited outstanding predictive capacity for the development of GDM (AUC: 0.96), regardless of maternal BMI or ethnicity [223]. This finding reinforces the clinical value of triglycerides as an early biomarker, not only for GDM but also as an independent risk factor in the progression toward type 2 diabetes mellitus [109,148,158,167,224,225]. Complementarily, Zhu et al. found that a fasting plasma glucose (FPG) level of 92 mg/dL was correlated with the diagnosis of GDM between weeks 24 and 28, although it should not be considered sufficient on its own to confirm the diagnosis [226]. Similarly, Ozgu-Erdinc et al. reported that cut-off points of 87.5, 92, and 99.5 mg/dL offered varying combinations of sensitivity and specificity (70.0/66.1%, 52.4/81.1%, and 23.7/94.9%, respectively) [109]. These findings support a linear relationship between higher fasting glucose levels in the first trimester and an increased risk of developing GDM in later weeks [227,228,229,230]. Furthermore, values below 78 mg/dL have been proposed as a potential threshold to rule out GDM and avoid unnecessary testing [231].

The combined consideration of these metabolic factors, also identified as strategic nodes within the network structure, could provide a solid foundation for the development of more sensitive and robust predictive models, as well as guide earlier and more effective clinical interventions.

On the other hand, vitamin D also acted as a structural intermediary in the network according to betweenness centrality, reinforcing its role as a functional bridge between metabolic, inflammatory, and behavioral processes. For instance, the association between elevated triglyceride levels and hs-CRP (“high-sensitivity C-reactive protein,” a marker of low-grade systemic inflammation) is only significant in pregnant women with vitamin D deficiency, but not in those with adequate levels [185]. This suggests that sufficient vitamin D concentrations may exert a modulatory effect, inhibiting hyperlipidemia-induced inflammation during pregnancy. Beyond glucose metabolism, vitamin D has also been linked to additional benefits: its supplementation has been associated with slower epigenetic aging [182], and it has been proposed to attenuate maternal inflammation that could negatively affect the implantation or survival of male fetuses [195]. Furthermore, low vitamin D levels have been associated with greater severity of liver fibrosis and higher positivity for hepatitis B [113].

Vitamin D showed a consistent presence across the three main metrics—closeness, eigenvector, and betweenness. This cross-metric convergence positions it as the only node with structural relevance confirmed from multiple topological perspectives. Furthermore, it was the node with the highest number of outgoing edges in the entire network, reinforcing its role as a key emitter with the capacity to exert a direct impact on multiple clinical components represented in the network. Taken together, these findings underscore the value of directed network analysis as a tool for identifying priority nodes in complex clinical systems.

Despite the central structural role of vitamin D in our network, the World Health Organization does not recommend routine supplementation during pregnancy, including gestational diabetes, due to insufficient evidence of consistent benefits on maternal and neonatal outcomes. Supplementation is only advised in cases of documented deficiency, with a reference intake of about 200 IU/day [232]. In contrast, evidence syntheses such as the Cochrane Review (2019, updated 2024) [233], report that vitamin D supplementation during pregnancy probably reduces the risk of gestational diabetes, although without specifying an optimal dosage and often using higher daily amounts than those proposed by the WHO. Furthermore, recent evidence highlights that benefits may not be limited to correcting deficiency: maternal levels of 25 (OH)D above 40 ng/mL have been associated with preventive effects on infectious and autoimmune conditions and with improved metabolic outcomes, including reduced risk of gestational diabetes [234]. These findings suggest that supplementation during pregnancy may be better guided not only by deficiency status but also by the metabolic risk profile and by achieving functional serum vitamin D thresholds rather than adhering to minimal intake recommendations. Further high-quality longitudinal and interventional studies are needed to clarify optimal target levels and to support the update of current international guidelines, particularly regarding metabolic profiles such as gestational diabetes. Moreover, integrating vitamin D into predictive models, alongside other clinically relevant nodes, could support the development of earlier, more accurate, and personalized prevention strategies for GDM.

Nevertheless, this study has certain limitations. The network was constructed from cross-sectional analyses, which prevents establishing causal or temporal relationships. Therefore, validation in longitudinal studies is needed to assess its predictive value during pregnancy. Additionally, GDM subtypes were not differentiated, even though they may involve distinct underlying mechanisms.

Moreover, the optimal functional levels of certain biomarkers may not align with conventional clinical reference ranges. For instance, while fasting plasma glucose (FPG) values below 100 mg/dL are traditionally considered normal, studies such as that by Zhu et al. have reported that levels as low as 92 mg/dL are already correlated with GDM diagnosis between gestational weeks 24 and 28 [226]. Other authors have proposed even lower cut-off points, such as 87.5 mg/dL [109].

Similarly, the optimal thresholds for vitamin D in the general population remain unclear, reflecting a lack of consensus among international institutions. For example, the National Academy of Medicine (NAM) recommends levels between 20 and 50 ng/mL, the U.S. Endocrine Society suggests a range of 30 to 149 ng/mL, and the Spanish Society of Endocrinology and Nutrition advises values between 30 and 99 ng/mL. This variability complicates the definition of functional thresholds in specific metabolic contexts. In particular, the relationship between vitamin D levels and insulin resistance, a central mechanism in GDM pathophysiology, remains poorly defined [235,236,237,238]. Another factor to consider is that current studies rarely integrate magnesium status alongside vitamin D, despite their well-documented interdependence. The activation, transport, and cellular action of vitamin D depend on magnesium [239], while vitamin D promotes intestinal absorption of this mineral, establishing a bidirectional relationship. Dysregulation of this dyad has been associated with an increased risk of insulin resistance [240,241,242]. Therefore, jointly assessing vitamin D and magnesium levels could improve the accuracy of predictive models. In particular, the use of markers such as erythrocyte magnesium could provide a more precise estimation of this interaction, enabling the development of a functional vitamin D/magnesium ratio tailored to individual metabolic status. Incorporating this perspective into network-based studies could yield more comprehensive and clinically relevant insights for gestational diabetes.

5. Conclusions

The findings presented here highlight the need to define clinically relevant thresholds based on predictive evidence rather than on traditional diagnostic criteria. In this regard, network-based models provide a valuable tool not only for identifying structurally relevant nodes representing key clinical factors but also for detecting concentration levels most closely associated with risk, as observed for triglycerides, ferritin, and vitamin D. Within this framework, vitamin D and sedentary behavior emerged as structurally dominant and clinically actionable nodes, underscoring the importance of integrating both biochemical and lifestyle factors into predictive models. This integrative approach offers a more personalized and physiologically coherent framework for the clinical management of women at metabolic risk, while also supporting the design of more sensitive and effective diagnostic and preventive strategies for GDM.