Patient-Zero Set in Network Epidemics

• Patient-Zero Set is the collection of nodes or hyperedges that could plausibly be the initial source of an outbreak, capturing uncertainty from noisy data and network dynamics.
• Methodologies such as Bayesian inference, contact tracing, belief propagation, and graph neural networks are used to compute posterior probabilities or identify zero-indegree nodes as candidate origins.
• Applications span epidemiology, malware detection, and information diffusion, while challenges include ambiguity in dense networks and computational scalability in simulations.

A patient-zero set is the collection of nodes in a network that are plausible candidates for the origin (or "index case") of a contagion or information diffusion process given observed data, model assumptions, and inference procedures. This concept generalizes the point estimate of “patient zero” to a set-valued or probabilistic description, capturing both uncertainty from noisy/incomplete observations and the inherent non-identifiability arising from the network structure and dynamics. The patient-zero set is formalized differently across algorithmic paradigms—Bayesian inference, contact-tracing logic, probabilistic graphical models, and message-passing algorithms—but generally represents nodes or hyperedges that, conditional on observations, could have acted as initial sources of the outbreak consistent with all available evidence.

1. Formal Definitions and Notational Variants

Let $G=(V,E)$ be a (possibly hyper-)graph representing contact structure, and $X^T$ an observed snapshot of nodal states (e.g., susceptible, exposed, infectious, recovered) at time $T$. The patient-zero set $S^*$ is typically defined, for a candidate node $i\in V$, via a posterior probability or feasibility test:

• Bayesian Formulation: Compute $P_i^0 := P(x_i^0 = I | X^T)$, the marginal posterior probability that node $i$ was infected at $t=0$. Then for a threshold $\theta$,

$S^*(\theta) = \{i \in V : P_i^0 \ge \theta \}$

or by selecting the top-$K$ nodes with highest $P_i^0$ (Altarelli et al., 2014).

• Feasibility-Based (Contact Tracing) Approach: Define

$$\mathcal S = \left\{ s \in V : \exists\text{ infection tree compatible with data in which } s \text{ is patient-zero} \right\}$$

where data comprises test and contact queries, possibly including onset times and household information (Ódor et al., 2021).

• Transmission DAGs (Visual Analytics): After reconstructing a directed acyclic event-graph of inferred transmission, the patient-zero set is those nodes of indegree zero:

$Z = \{\,p \in V_T : \mathrm{indegree}_T(p) = 0 \}$

representing all sources with no upstream infector in the reconstructed cascade (Baumgartl et al., 2020).

• Hypertree/Group Models: In SI dynamics on hypertrees (e.g., social bubbles), the patient-zero set may take the form of a single hyperedge or all hyperedges consistent with observed infection subtrees, with estimation based on maximum likelihood or overlap-weighted path lengths (Spencer et al., 2020).

These definitions encompass both hard (set-valued) and soft (ranked or probabilistic) versions, depending on application context and available computational resources.

2. Information-Theoretic and Structural Limits

The identifiability of the patient-zero set is fundamentally constrained by both the epidemic process and network topology.

• Ticking Clock Limit: On an Erdős–Rényi graph under SIR/SEIR dynamics, the maximum time $t_\text{max}$ after which patient-zero is no longer detectable by any algorithm is given by

$t_\text{max} \simeq \frac{\log N}{\gamma(R_0 - 1)}$

where $N$ is the graph size, $\gamma$ the recovery rate, $R_0$ the basic reproduction number (Shah et al., 2020). Beyond $t_\text{max}$, the infection subgraph becomes sufficiently dense that multiple origins are consistent with the data, and any patient-zero set necessarily grows.

• Cycle-Induced Ambiguity: Even before $t_\text{max}$, cycles in the infection subgraph generate inescapable ambiguity, reflected in top-1 accuracy bounds:

$$P_\text{max} \leq \frac{1}{3} + \frac{2}{3}(1-p)^{\binom{p|G_I|}{2}}$$

for edge probability $p$ and infection subgraph $G_I$ (Shah et al., 2020).

• Detection Probability: In SIR on general networks, the correct-source detection probability decays exponentially with source separation and depends critically on the separation exponent

$$\alpha(p,q) = \ln\frac{p+q}{q} - \ln(\langle k\rangle - 1)$$

where $p$ is infection, $q$ is recovery probability, and $\langle k\rangle$ is mean degree (Antulov-Fantulin et al., 2014). For $\alpha<0$, source detection is practically impossible.

A direct implication is that, outside of tree-like and sparse regimes, the credible patient-zero set is almost always non-singleton.

3. Algorithmic and Probabilistic Methods for Patient-Zero Set Estimation

Message Passing and Probabilistic Inference

• Belief Propagation (BP): The BP approach models the full posterior over sources and parameters, with Bethe free-energy minimization yielding node-wise marginals $P_i^0$. The patient-zero set is then either a thresholded set $S^*(\theta)$ or a top-$K$ list (Altarelli et al., 2014).
• Monte Carlo and Soft-Margin Simulation: Exhaustive simulation of SIR/SEIR processes started from each candidate source, possibly with kernel-weighted “soft margin” scoring, is used to estimate $P(Θ=i|r_*)$ (Antulov-Fantulin et al., 2014). Early pruning and similarity kernels improve tractability.
• Graph Neural Networks (GNNs/GCNs): Parameter-agnostic deep message-passing architectures (e.g., L-layer GCN with residuals and normalization) can be trained discriminatively to predict $p_i$, the estimated posterior probability that node $i$ was the source, with cross-entropy loss to one-hot labels. These models can rank plausible patient-zero candidates and rapidly focus epidemiological resources (Shah et al., 2020).

Contact-Tracing and Logical Inference

• Source Detection via Contact Tracing Framework (SDCTF): The patient-zero set $\mathcal S$ is the set of nodes for which there exists some infection pathway and parameter assignment compatible with all observed queries, including the presence of asymptomatics and noisy or incomplete contact knowledge. The objective is to reduce $\mathcal S$ to a singleton through adaptive queries (Ódor et al., 2021).
• DAG-based Reconstruction (Hospital Outbreak Analytics): Time-resolved contact DAGs reconstructed from transfer and test logs yield a transmission graph $T$, with patient-zero set $Z$ as the root nodes. Integration with genomic data can further filter $Z$ (Baumgartl et al., 2020).

Group and High-Order Models

• Hypertree/Group Gathering Estimation: For SI processes on hypertrees with group contacts (e.g., social bubbles), the patient-zero set is defined over hyperedges. Here, a closed-form maximum likelihood estimator identifies the hyperedge whose weighted overlap distances match observed infection patterns. This algorithm runs in $O(n)$ time over the hyperedge set (Spencer et al., 2020).

4. Empirical Performance and Practical Implications

• Accuracy vs. Network Topology: On tree-like synthetic graphs, GCNs and DMP yield $\sim$64–80% top-1 accuracy, while on dense ER and RGG graphs, GCNs achieve 2–3$\times$ higher accuracy than classic message-passing (Shah et al., 2020).
• Robustness to Noise and Partial Observation: BP methods localize patient zero to the top-2% of candidates under moderate noise, and can reliably infer S/R classification (AUC $\gtrsim$0.9) even under state confusion (Altarelli et al., 2014).
• Time Constraints: All methods rapidly lose accuracy as the outbreak ages, consistent with the $t_\text{max}$ theoretical bound.
• Query Complexity: In settings with minimal contact knowledge and asymptomatics, local search (LS+) identifies patient-zero with $O(L\Delta)$ queries (path length $L$, max degree $\Delta$), sublinear in network size and robust to high asymptomatic rates (Ódor et al., 2021).
• Visual Analytics Pipelines: Integrated event- and contact-DAG systems can reduce manual investigation times by orders of magnitude, surface multiple plausible patient-zero candidates, and allow expert curation. Genomic data may reduce the set to a unique origin or small clonal cluster (Baumgartl et al., 2020).

Algorithm/Approach	Defines Patient-Zero Set As	Computation/Inference
Belief Propagation	Nodes with high $P_i^0$	Variational inference (BP, Bethe free energy)
GCN / Neural Network	Nodes with high predicted $p_i$	Trained on simulation, cross-entropy loss
Contact Tracing	Consistency set $\mathcal S$	Adaptive logical/empirical tracing
Visual Analytics / DAG	Zero-indegree in inferred pathway	Data-driven, combined with genomics
Hypertree MLE	Hyperedges matching infection pattern	Closed-form via weighted path-length

5. Limitations, Robustness, and Extensions

• Non-identifiability: Multiple source nodes or hyperedges may be equally consistent with observations in dense or highly cyclic subgraphs, rendering the patient-zero set inherently non-singleton beyond certain outbreak ages or densities (Shah et al., 2020, Spencer et al., 2020).
• Model Limitations: Many algorithms assume full knowledge of the underlying contact network, SIR/SI parameters, or noiseless observations. Approaches such as BP and SDCTF seek to relax these assumptions via marginalization, gradient ascent on likelihoods, or adaptive querying (Altarelli et al., 2014, Ódor et al., 2021).
• Scalability: Simulation-based approaches (e.g., soft-margin Monte Carlo) are computationally intensive for large or high-density networks, while message-passing and GCNs offer significant speedups, sometimes $>100\times$ over classic algorithms (Shah et al., 2020).
• Extensions: Further research directions include (i) patient-zero set estimation on general hypergraphs with cycles, (ii) joint inference of epidemic parameters and sources, (iii) marginalization over observation times and partial/incomplete test data, and (iv) integration with multi-modal (e.g., genomic) evidence (Spencer et al., 2020, Altarelli et al., 2014).

6. Applications and Significance

The concept of the patient-zero set underpins rapid outbreak response, enabling prioritization of high-likelihood index cases for containment or backward/forward tracing. Tools such as GCNs and visual analytics pipelines allow real-time ranking of plausible sources from abstract network snapshots, while probabilistic and contact-tracing approaches retain rigorous uncertainty quantification. Asymptomatic transmission, parametric uncertainty, and partial network observability fundamentally shape both the methods applied and the size/credibility of the inferred patient-zero set. Applications extend beyond classical epidemiology to malware/rumor source tracking, information cascades, and group-based contagion processes (Shah et al., 2020, Spencer et al., 2020, Ódor et al., 2021).