Children who had COVID-19 accumulate more diagnosis codes in the months that follow. They see doctors more often, and every visit adds codes. The open question is whether this extra burden is organized or diffuse. An organized burden would look like a genuine multi-organ syndrome, in which breathlessness, palpitations and fatigue reliably travel together across organ systems. A diffuse burden would be a scatter of codes that comes from more contact with the health system. Current pediatric studies cannot tell these apart. They estimate one risk ratio per outcome, or they scan the administrative ICD-10 code tree for over-represented branches. Neither approach asks how codes co-occur within a child once you account for how many codes that child received.
We propose to answer this with a tool from network science, assortative mixing: the tendency of connected nodes to share a trait, measured as a correlation over network links. We build a network of diagnosis codes linked by how often they co-occur in the same child. Each code is labeled with its organ system. We then ask whether post-infection children show more links that cross organ systems than matched uninfected children. The comparison is made against a reference that keeps each child's total code count fixed and each code's overall frequency fixed. A pediatric long COVID syndrome should show up as a COVID-specific drop in organ-system assortativity that survives this conditioning. Utilization noise should not.
Core Claim
Raw co-occurrence networks built from electronic health records (EHRs) mostly reflect who gets seen most. A child with forty codes links every pair of them. That child therefore wires frequent codes to one another and produces apparent hubs and an apparent dense core even if the codes are clinically unrelated. Evtushenko and Kleinberg (The paradox of second-order homophily; Author Correction: The paradox of second-order homophily in networks) show the general version of this danger: in Facebook100, node degree and first-order homophily are correlated, with values generally between −0.2 and 0.25. Mixing statistics are not independent of degree, so they must be computed conditional on it.
Our claim has three parts. First, a degree-preserving null for the patient–code bipartite graph fixes the problem. This graph links each child to the codes that child received, and the null randomly rewires it while keeping every child's code count and every code's prevalence. Second, the difference between the COVID-positive and COVID-negative cohorts in organ-system assortativity, measured after this conditioning, is a direct and falsifiable test of whether pediatric post-acute sequelae (PASC) form a cross-system syndrome. Third, the code pairs that drive this difference give a pediatric phenotype definition that is cleaner than the one obtained from raw co-occurrence.
Write $O_{ab}$ for the number of children in a cohort who carry both codes $a$ and $b$ in the post-acute window. Write $E_{ab}$ for its mean over degree-preserving rewirings. We weight the code network by the excess $w_{ab} = O_{ab} - E_{ab}$ and compute Newman's attribute assortativity $r_{\text{sys}}$ over organ-system labels from these weights. The test statistic is
$$\Delta = r_{\text{sys}}^{\text{COVID+}} - r_{\text{sys}}^{\text{COVID−}},$$
with uncertainty from a bootstrap over children. A syndrome predicts $\Delta < 0$: excess co-occurrence shifts toward cross-system pairs. Pure utilization inflation predicts $\Delta \approx 0$ once degree is conditioned on, even if the raw network shows a large shift.
Research Program
Role of the cited work. The proposal draws on its sources in different ways. Lorman et al. (TreeScan on RECOVER PEDSnet) motivate it. They found 304 significant cuts in the ICD-10 tree, with the strongest signal in the "symptoms and signs" chapter (dyspnea, chest pain, fatigue). They also noted that 55.8% of cases had a pre-existing chronic condition. A tree scan finds over-represented branches but cannot express that dyspnea and palpitations sit on different branches yet co-occur, and that is exactly our object of study.
The two Zhang et al. RECOVER studies supply the cohort design and the per-outcome comparators. The cardiovascular study (Zhang et al., cardiovascular outcomes) models 18 outcomes independently, reporting for example any cardiovascular outcome at RR 1.63. The disparities study (Zhang et al., racial/ethnic differences) uses difference-in-differences with negative control outcomes, and we reuse its negative-control logic.
Dong et al. (Segregated interactions; Segregated interactions in urban and online space) supply the mixing-matrix machinery, which we transfer from income groups to organ systems.
Estiri et al. (temporal sub-phenotypes; Characterization of long COVID temporal sub-phenotypes by distributed representation learning from electronic health record data: a cohort study) supply the strongest clustering comparator: representation learning plus Louvain community detection, with AUROC generally above 0.8 in hospitalized adults. Their work never conditions on patient degree, and it does not include children.
Turalska et al. (Cascading failures; Cascading failures in scale-free interdependent networks) motivate the core-versus-periphery question. Their evidence is limited to synthetic sandpile models, so we treat "rich club" language about clinical symptoms as a hypothesis to test rather than a mechanism to assume. The interactome work (Gulbahce et al., HuRI) is deliberately out of scope. Structured pediatric diagnosis codes do not map cleanly onto proteins, and adding that layer would make the main test less clean.
Cohorts. We work in RECOVER PEDSnet with the Lorman et al. design: PASC cases and COVID-positive and COVID-negative children, propensity-matched with an 18-month pre-infection washout. We extend this to the larger COVID-positive and COVID-negative populations used by Zhang et al.
Building the graph. For each child we form the bipartite graph of distinct codes recorded in the 28–179 day post-index window, mapped to a fixed depth of the ICD-10 hierarchy.
The null model. Rewiring is stratified within site, age band and calendar period. This prevents site coding habits and variant era from leaking into the null. Rewiring uses degree-preserving swaps on the bipartite graph, which keep each child's code count and each code's prevalence.
Analyses. We first compute raw and degree-conditioned versions of three quantities: degree assortativity of the code network, organ-system assortativity, and $\Delta$. We then extract the cross-system pairs with the largest COVID-specific excess and define a phenotype from them. Finally, we test that phenotype for discrimination and transportability.
Baselines
Do-nothing reference. This is the raw co-occurrence network, with organ-system assortativity computed from $O_{ab}$ alone, contrasted between cohorts. If raw and conditioned $\Delta$ agree, the conditioning adds nothing and we will say so.
Prevalence-only null. This ablation keeps code prevalences but lets each child's code count float. It isolates the key ingredient, conditioning on utilization. Our claim requires that this null and the full degree-preserving null disagree in a COVID-positive versus COVID-negative contrast.
TreeScan. This is the strongest existing discovery method on this exact data. It is rerun with the Lorman et al. configuration on identical cohorts and windows. The reported reference is 304 significant cuts for PASC versus COVID-negative children. The method yields no discrimination metric, so we compare what each method discovers. We beat it if conditioned cross-system pairs recover a coherent multi-chapter phenotype that TreeScan's within-branch cuts split apart, and if that phenotype replicates across sites where individual cuts do not.
Per-outcome Poisson risk ratios. This follows Zhang et al., with RR 1.63 for any cardiovascular outcome, and gives the outcome-by-outcome view.
Louvain clustering on raw co-occurrence. This follows Estiri et al. and is given the same input codes and the same tuning budget, a resolution sweep over equal grid sizes.
Phenotype discrimination. The target is held-out U09.9 diagnosis in children indexed after October 2021. The comparator is a gradient-boosted model on code counts and demographics under the same features and tuning budget. No reported result was found on Lacuna for pediatric PASC prediction C-index. The nearest figure is adult: C-index around 0.6 for subjective symptoms in Zang et al.. Beating the comparator means a higher held-out C-index at transfer to unseen sites, not only within the training sites.
Why not just…?
Why not just add visit count as a covariate in pairwise logistic or log-linear models? This is the most serious threat. A log-linear model with child and code effects is close in spirit to a degree-preserving null. We will fit it alongside the null and report agreement. If the two agree, the contribution is the cross-system mixing statistic $\Delta$ and the COVID contrast, not the null itself. The rewiring null keeps two practical advantages. It conditions exactly on code counts without a functional-form assumption about how utilization scales co-occurrence. It also yields a single network-level test rather than tens of thousands of pairwise tests.
Why not just rely on propensity matching on prior utilization, as RECOVER studies already do? Matching balances utilization before infection. Infection itself raises post-infection visits, so post-infection code count differs between cohorts by construction. Raw co-occurrence then grows for reasons unrelated to any syndrome. Conditioning on post-infection degree asks the narrower question that matters: given how much care a child received, is the burden structured? The prevalence-only ablation shows directly how much of the raw signal this accounts for.
Why not just use TreeScan or Louvain clustering? TreeScan is bound to the billing tree. Louvain on raw co-occurrence partitions a graph that, by our hypothesis, is dominated by high-utilization children. Both remain baselines. If Louvain on the conditioned network matches our phenotype, that counts as support for the conditioning, not as a defeat.
Decisive Checks
The first check is synthetic. We simulate bipartite data with realistic heavy-tailed code counts and plant a cross-system cluster in one cohort only. In a second scenario we inflate code counts in that cohort with no planted cluster. Raw $\Delta$ should fire falsely in the second scenario. Conditioned $\Delta$ should fire only in the first.
The second check is the main estimate, $\Delta$ in RECOVER PEDSnet with bootstrap intervals. It is repeated within age bands and variant eras, since Zhang et al. found risks persisting across pre-Delta, Delta and Omicron.
The third check is a negative control. We recompute $\Delta$ restricted to code pairs with no plausible post-viral link, for example injury codes paired with other chapters. A nonzero $\Delta$ there signals residual bias and invalidates the main result.
The fourth check is site transportability. The phenotype is fit on some sites and evaluated on held-out sites. The motivation is the strong regional heterogeneity seen in adults: 38 versus 11 significant categories in Zang et al..
The outcome that would falsify our hypothesis is a raw $\Delta$ that is clearly negative while the conditioned $\Delta$ sits at zero. That would mean pediatric "multi-system" findings derived from co-occurrence largely reflect utilization. Even this result would be worth publishing.
References
- Author Correction: The paradox of second-order homophily in networks
- Understanding pediatric long COVID using a tree-based scan statistic approach: An EHR-based cohort study from the RECOVER Program
- Post-Acute Cardiovascular Outcomes of COVID-19 in Children and Adolescents: An EHR Cohort Study from the RECOVER Project
- Racial/ethnic differences in post-acute sequelae of SARS-CoV-2 in children and adolescents in the United States
- Segregated interactions in urban and online space
- Characterization of long COVID temporal sub-phenotypes by distributed representation learning from electronic health record data: a cohort study
- Cascading failures in scale-free interdependent networks
- Viral Perturbations of Host Networks Reflect Disease Etiology
- A reference map of the human protein interactome
- Identification of risk factors of Long COVID and predictive modeling in the RECOVER EHR cohorts
- Data-driven analysis to understand long COVID using electronic health records from the RECOVER initiative