Health Misinformation Vulnerability Centrality (MVC)

• Health Misinformation Vulnerability Centrality (MVC) is a hybrid metric that integrates network exposure with intrinsic vulnerability to quantify users' propensity to absorb and retransmit false health claims.
• MVC employs an iterative update over a static network, multiplying in-degree by initial vulnerability over 5–10 rounds to simulate repeated misinformation exposure.
• Empirical findings indicate MVC uncovers additional influencers missed by traditional metrics, enhancing the effectiveness of targeted misinformation interventions by up to 25%.

Health Misinformation Vulnerability Centrality (MVC) is a centrality metric introduced to identify nodes in online social networks that are simultaneously structurally well connected and highly susceptible to health misinformation. In the formulation reported by Sikosana et al., MVC weights a user’s connectivity, interpreted as exposure, with an empirically derived vulnerability factor, with the aim of surfacing users who may not be absolute network hubs but who, once exposed, are especially likely to absorb and retransmit false health claims (Sikosana et al., 11 Jul 2025). Within the broader framework of advanced centrality metrics proposed alongside dynamic influence centrality (DIC) and propagation centrality (PC), MVC is positioned as a behaviourally grounded complement to traditional structural measures for analysing health misinformation diffusion during crises such as the COVID-19 pandemic (Sikosana et al., 11 Jul 2025).

1. Conceptual basis and problem setting

MVC is defined on a static directed graph $G=(V,E)$ of $n$ users, with adjacency matrix $A \in \mathbb{R}^{n \times n}$, where $A_{ij}>0$ if user $v_j$ retweets or mentions $v_i$; otherwise it is $0$. The in-degree of node $v_i$ is given by

$\mathrm{indegree}(v_i)=\sum_j A_{ij},$

and denotes how many distinct sources can reach $v_i$ (Sikosana et al., 11 Jul 2025).

The metric is designed around two coupled properties. First, users differ in baseline susceptibility to health misinformation. Second, exposure multiplies susceptibility: the more incoming links a user has, the greater the amplification of any predisposition to believe false claims. MVC therefore combines structure, through in-degree, with behaviour, through an initial vulnerability score, in order to identify what the paper describes as “weak links” through which misinformation can percolate widely even if those nodes are not global hubs (Sikosana et al., 11 Jul 2025).

This framing is explicitly contrasted with traditional centralities. Degree and eigenvector centrality ignore individual gullibility, while betweenness and closeness ignore user-level differences in credulity. MVC is presented as filling this gap by directly modeling how susceptible individuals amplify false claims (Sikosana et al., 11 Jul 2025). A plausible implication is that MVC is not intended to replace purely structural measures, but to supplement them when misinformation susceptibility is heterogeneous across accounts.

2. Mathematical formulation

Each node $n$0 is assigned an intrinsic susceptibility score

$n$1

based on features such as historical retweeting of unverified content, use of emotionally charged language, or external credibility ratings (Sikosana et al., 11 Jul 2025).

For $n$2, with $n$3–$n$4, vulnerability scores evolve according to

$n$5

Here, $n$6 is the vulnerability at step $n$7, and $n$8 remains fixed for the static snapshot (Sikosana et al., 11 Jul 2025).

After $n$9 iterations, the resulting scores are min–max normalized to $A \in \mathbb{R}^{n \times n}$0:

$A \in \mathbb{R}^{n \times n}$1

The variables and parameters are specified as follows:

Quantity	Meaning
$A \in \mathbb{R}^{n \times n}$2	initial vulnerability, user-defined or estimated
$A \in \mathbb{R}^{n \times n}$3	in-degree of $A \in \mathbb{R}^{n \times n}$4 in $A \in \mathbb{R}^{n \times n}$5
$A \in \mathbb{R}^{n \times n}$6	number of iterations; recommended 5–10
$A \in \mathbb{R}^{n \times n}$7	final vulnerability centrality of $A \in \mathbb{R}^{n \times n}$8

The underlying intuition is that temporal dynamics are approximated by iterating the build-up of vulnerability over $A \in \mathbb{R}^{n \times n}$9 discrete rounds of exposure, mimicking repeated encounters with misinformation (Sikosana et al., 11 Jul 2025). This suggests that the “temporal” component in MVC is not derived from a time-stamped evolving graph, but from repeated updates over a static snapshot.

3. Algorithmic procedure and computational properties

The computational methodology is given as a five-step algorithm. The inputs are $A_{ij}>0$0, the adjacency matrix $A_{ij}>0$1, the initial vulnerability vector $A_{ij}>0$2, and $A_{ij}>0$3 such as $A_{ij}>0$4. The procedure first precomputes $A_{ij}>0$5 for all $A_{ij}>0$6. It then iterates from $A_{ij}>0$7 to $A_{ij}>0$8, updating each node by

$A_{ij}>0$9

After $v_j$0 iterations, min–max normalization is applied, and the output is $v_j$1 for all nodes (Sikosana et al., 11 Jul 2025).

The stated time complexity is linear in network size for fixed $v_j$2: precomputing in-degree requires $v_j$3, each iteration costs $v_j$4, and the total cost is

$v_j$5

Several implementation choices are specified. The number of iterations is chosen to be small, $v_j$6–$v_j$7, to ensure rapid convergence; beyond $v_j$8, score changes are minimal. Storing in-degree in a vector avoids repeated scans of $v_j$9. If true user-vulnerability features are unavailable, one can sample $v_i$0 from $v_i$1 with a fixed random seed for reproducibility (Sikosana et al., 11 Jul 2025).

The recommended hyperparameters are correspondingly limited: $v_i$2 as the number of iterations, an initial vulnerability distribution based preferably on real features or alternatively simulated from $v_i$3, and min–max normalization to $v_i$4 (Sikosana et al., 11 Jul 2025). Because the update depends only on precomputed in-degree and the current vulnerability vector, MVC is computationally simple relative to centralities that require repeated shortest-path or eigensystem computations. That interpretation follows directly from the reported complexity, although the paper does not formalize a direct asymptotic comparison to specific traditional metrics.

4. Empirical findings on the FibVID dataset

In the FibVID evaluation, Table 3 reports how many new influencers each novel metric contributed relative to the union of four traditional measures: degree, eigenvector, betweenness, and closeness. The traditional metrics identified $v_i$5 distinct influencers. Adding MVC contributed three new user IDs, $v_i$6, described as a $v_i$7 increase in influencer coverage. Across all three novel metrics, PC, MVC, and DIC, the total influencer set increased from $v_i$8 to $v_i$9 nodes, a $0$0 increase (Sikosana et al., 11 Jul 2025).

The paper also reports simulated intervention results. Removing the top $0$1 traditional influencers in a diffusion model cut misinformation volume by $0$2. Removing the expanded set of $0$3, including MVC-driven nodes, yielded a $0$4 reduction, corresponding to a $0$5 relative improvement over the traditional baseline (Sikosana et al., 11 Jul 2025).

MVC-specific findings are emphasized. The three MVC-exclusive nodes were absent from all traditional top-10 lists. In addition, proxy ground-truth validation in Table 2 showed that high-MVC nodes coincide with elevated counts of emotionally charged language and unverified retweets, which the paper interprets as confirming alignment between MVC and observable user behaviour (Sikosana et al., 11 Jul 2025). This supports the intended interpretation of MVC as a mechanism for surfacing susceptible amplifiers rather than merely high-degree accounts.

The summary in the same source further states that MVC alone added $0$6 new influencers to the PC+traditional set, and $0$7 more relative to traditional alone (Sikosana et al., 11 Jul 2025). Taken together, these results indicate that MVC changed the composition of the influencer set, not only its size.

5. Generalisability beyond COVID-19 discourse

To assess cross-domain robustness, the framework was validated on the Monant Medical Misinformation dataset, which covers broader medical misinformation topics including vaccines, alternative treatments, and pharmaceutical skepticism. In this setting, all metrics were recomputed rather than transferred without recalibration (Sikosana et al., 11 Jul 2025).

The key findings are reported from Figure 1. Among the “top influencers,” only $0$8 user IDs were common to both traditional and advanced metrics; $0$9 were unique to traditional metrics and $v_i$0 were unique to advanced metrics. In the advanced-metric top-10 shown in Figure 2C, several accounts, including IDs $v_i$1 and $v_i$2, emerged solely because of high MVC scores despite moderate degree (Sikosana et al., 11 Jul 2025).

The paper interprets these results as confirming that MVC reliably flags susceptible amplifiers across topics, from COVID-19 to vaccine hesitancy and beyond (Sikosana et al., 11 Jul 2025). A plausible implication is that the metric is sensitive to patterns of susceptibility that are not reducible to topic-specific interaction volume. However, the evidence presented is empirical rather than theoretical: the generalisability claim rests on recomputation and comparative results across the two datasets.

6. Applications, interpretation, and reporting practice

Several practical use cases are identified. For targeted fact-checking and moderation, health platforms can compute MVC in near real time to rank users by misinformation susceptibility. The paper gives an example in which accounts with $v_i$3 might be prioritized for media-literacy nudges, inline fact-check labels, or reduced recommendation weighting (Sikosana et al., 11 Jul 2025). For public health campaign design, MVC identifies “weak-link amplifiers” whose personal networks are small but unusually credulous, allowing corrective messages such as trusted-source infographics to be delivered directly into these users’ feeds to pre-empt rumor uptake (Sikosana et al., 11 Jul 2025).

For monitoring dashboards, the proposed use is explicitly layered. PC spots “super-spreaders” with cascade reach, MVC spots “super-vulnerables” likely to absorb false claims, and DIC spots “long-tail” resurgers. On that basis, dashboards can alert moderators when any high-MVC node begins interacting with unverified claims, enabling rapid response (Sikosana et al., 11 Jul 2025). This suggests a division of analytical labour among the three advanced metrics rather than a single-metric pipeline.

The best practices stated in the source are also specific. Initial $v_i$4 should be grounded in real-world observables such as prior fact-check shares or sentiment analysis when possible. The iteration count $v_i$5 should be calibrated via hold-out evaluation, with $v_i$6–$v_i$7 iterations reported as sufficient in networks up to $v_i$8 nodes. MVC should be reported alongside degree and PageRank to capture a fuller picture of influence (Sikosana et al., 11 Jul 2025).

The summary characterization of MVC is that it embeds user-level susceptibility into network structure, revealing “hidden” conduits of health misinformation and providing a practical, scalable, and behaviourally grounded tool for platforms and public-health agencies seeking to pre-empt and counteract harmful health misinformation (Sikosana et al., 11 Jul 2025). In encyclopedic terms, MVC is therefore best understood as a hybrid centrality: its distinctive feature is not a new notion of connectivity alone, but the explicit incorporation of vulnerability into centrality scoring.