Do Post-COVID Symptoms in Children Cascade or Sort? A Capacity-Aware Attachment Model for Pediatric EHRs

When a child develops lingering problems after a SARS-CoV-2 infection, a fatigue diagnosis is often followed months later by a gastrointestinal complaint, then a cardiac workup. Two very different stories fit this pattern. In the first, the problems cascade: each newly affected organ system makes the next one more likely. In the second, they sort: some children are simply more fragile from the start and pick up diagnoses everywhere, while others pick up few. Both stories produce the same skewed snapshot, in which a minority of children carry most of the multi-organ burden, so cross-sectional tools cannot tell them apart. The distinction still matters clinically. A cascade argues for early intervention after the second affected system. Sorting argues for identifying fragile children at baseline.

Network science has a tool built for this exact confusion. In growing networks, "rich-get-richer" growth, called preferential attachment, competes with intrinsic node "fitness" as the explanation for why a few nodes gather most links. The PAFit method separates the two by estimating, without assuming a functional form, both the attachment curve (how link acquisition rate depends on current degree) and a per-node fitness (Joint estimation of preferential attachment and node fitness in growing complex networks). On a Facebook network it found that attachment was strongly sub-linear (exponent about 0.43) once fitness was accounted for. We propose to treat each child's post-acute record as a growing set of links from the child to organ systems. We would then estimate the attachment curve and child fitness jointly, using two corrections that matter specifically for medical records. The first counts only currently active diagnoses, because cumulative coding inflates apparent accumulation. The second measures time in clinical encounters, so that more visits do not masquerade as cascading disease.

Core Claim

The claim is that the shape of the attachment curve in pediatric post-acute records is identifiable and decision-relevant. We can measure whether, after infection, a child's per-encounter hazard of a new organ-system diagnosis rises with the number of systems already active (cascade), stays flat once fitness is accounted for (sorting), or rises and then plateaus (saturating cascade). We can also measure whether that curve differs between infected children and matched uninfected controls. The contrast with infected-versus-control curves is the new scientific quantity. Existing pediatric studies report which diagnoses are enriched (Understanding pediatric long COVID using a tree-based scan statistic approach) or how many visits increase (Healthcare utilization following SARS-CoV-2 infection in children and adolescents with chronic conditions). None reports how prior burden changes future acquisition once frailty is held fixed.

We expect the attachment curve to be sub-linear and to saturate, as it does in social and citation networks (A Unified Framework for Complex Networks with Degree Trichotomy Based on Markov Chains). The analogy is only a hypothesis, and the experiments are designed so that a flat curve would be an equally publishable and useful answer.

Research Program

The basic object is a two-sided growing network. One side holds children and the other holds organ systems (ICD-10 chapters) or, at finer resolution, ICD-10 three-character blocks. A link appears at the first post-acute encounter where a child receives a code in a target not active for them in the preceding window of $W$ days. We count time in encounters rather than calendar days. The model is a discrete-time hazard: at encounter $t$ of child $i$, the probability of a new link to target $s$ is

$$\lambda_{is}(t) = A\big(k_i^{W}(t)\big)\,\eta_i\,\beta_s ,$$

where $k_i^{W}(t)$ is the number of targets with a code in the last $W$ days (the child's active degree), $A(\cdot)$ is the attachment curve estimated nonparametrically at each degree, $\eta_i$ is child fitness, and $\beta_s$ is the baseline attractiveness of target $s$. Estimation adapts PAFit's regularized likelihood and minorize–maximization updates (Joint estimation of preferential attachment and node fitness in growing complex networks) to this two-sided, encounter-indexed setting, with a smoothness penalty on $A$ and a shrinkage prior on $\eta_i$.

Three design choices carry the substance. The first is active versus cumulative degree, which comes from evidence that cumulative edge retention manufactures false densification in social networks and that sliding windows correct it (Evidence of Temporal Artifacts in Social Networks). Pediatric symptoms resolve and wax and wane (Clinical manifestations and long-term symptoms associated with SARS-CoV-2 omicron infection in children aged 0–17 years in Beijing), so a cumulative count would keep a resolved cough "active" forever. That would inflate degree for children with long records and could bend the estimated curve upward artificially. The second choice is a mechanical ceiling. With roughly twenty chapters, degree is bounded, so a plateau could be pure arithmetic. We therefore report $A$ per still-available target and repeat the analysis at block resolution, where the ceiling is far away. A plateau that survives both is real. The third choice is to index time by encounter, which conditions on the visit stream. This partly absorbs the "provider caution" explanation the utilization study raises, since a child seen more often gets more chances at a code but not a higher per-encounter rate.

Two extensions draw on existing work. One adds the exploration/reinforcement idea from Uncovering User Interaction Dynamics in Online Social Networks, which weights ties by volume divided by a power of age. Here it would replace the hard window with a decaying activity weight, and we would test whether the conclusions depend on $W$. The other lets fitness depend on recorded pre-existing chronic conditions. This makes it possible to ask whether the doubled post-acute inpatient rate in chronically ill children reflects higher fitness (a fragile baseline) or a steeper curve (faster cascade from a nonzero starting degree).

Baselines

The naive reference point is a cross-sectional comparison. It counts the distinct organ systems each child accumulates in the post-acute window and compares the count distributions of infected children and controls with a negative binomial model, in the style of the utilization rate ratios of Healthcare utilization following SARS-CoV-2 infection in children and adolescents with chronic conditions. That study reports an overall rate ratio of 1.21 (95% CI 1.18–1.24). This baseline answers "how much more" but not "why." It will be run on the identical matched cohorts and windows.

The strongest competitor is a recurrent-event model with a random child-level frailty and the lagged count of active systems entered as a spline covariate. This is the epidemiologist's standard tool for separating state dependence from heterogeneity, and it is close kin to our model. We will fit it to the same encounter-indexed data, with the same window definition, the same spline flexibility budget as our penalized $A$, and the same cross-validated tuning grid. No reported result was found on Lacuna for this model on pediatric post-acute data. We would beat it only if our model gives higher held-out per-encounter log-likelihood for new-link events, or recovers simulated curves more accurately at equal flexibility. If the two tie, we will say so: the frailty model then serves as a parametric cross-check, and the contribution reduces to the scientific estimate and the windowing correction.

Two ablations isolate the key ingredients. A kernel-only model sets every $\eta_i$ equal, which is plain preferential attachment. A fitness-only model fixes $A$ to be flat. If kernel-only fits nearly as well as the full model, cascade is the story; if fitness-only does, sorting is. A third ablation swaps active degree for cumulative degree, which isolates the windowing correction. For context on curve shape we cite the Facebook exponent of about 0.43 (Joint estimation of preferential attachment and node fitness in growing complex networks) and the 34%–97% Kolmogorov–Smirnov improvement of saturating models over power laws on citation graphs (A Unified Framework for Complex Networks with Degree Trichotomy Based on Markov Chains). These are not targets on medical data.

Why not just…?

Why not just run TreeScan or latent class analysis? Both are excellent at what they do. TreeScan found 304 diagnostic clusters enriched in pediatric post-COVID cases (Understanding pediatric long COVID using a tree-based scan statistic approach). Latent class analysis split MIS-C into severe, moderate, and mild groups (Spectrum of severity of multisystem inflammatory syndrome in children). Both work on snapshots. A cascade and a fragile-child population produce the same snapshot, so neither can answer the question posed here. We see them as complementary: TreeScan's enriched branches tell us which finer-grained targets to model.

Why not just fit a frailty recurrent-event model and skip the network machinery? This is the real threat, and we treat it as the main baseline rather than arguing around it. The network framing contributes three things the standard model usually lacks: a nonparametric curve with a tested saturation shape, active-degree bookkeeping motivated by the temporal-artifact result (Evidence of Temporal Artifacts in Social Networks), and target-level attractiveness $\beta_s$. If an equally flexible frailty model matches us on held-out likelihood and simulation recovery, the methodological novelty is small. The finding about pediatric cascades stands regardless.

Why not just attribute everything to provider caution? Caution would raise visit counts. In an encounter-indexed model it should leave the per-encounter curve flat while raising fitness, which gives a testable signature. A negative-control target set, injuries and external causes, should show a flat curve and no infected-versus-control difference. If it shows the same curve as the post-acute targets, our signal is behavioral rather than biological, and we will report it as such.

Proposed Experiments

We will start with simulation. Synthetic cohorts matched to the size of RECOVER PEDSnet (1,309 cases and 6,545 each of COVID-positive and COVID-negative controls, per Understanding pediatric long COVID using a tree-based scan statistic approach) and to realistic encounter counts will be generated under cascade, sorting, and saturating-cascade truths. Recovery of $A$ is measured by integrated squared error, and recovery of $\eta$ by rank correlation. If the method cannot distinguish the three regimes at this sample size, we stop there and report the limit.

On RECOVER PEDSnet, the primary outcome is the infected-minus-control difference in estimated $A$ at each active degree, with bootstrap confidence bands. We will report it at chapter and block resolution and across $W$ of 30, 90, and 180 days. The decisive checks follow from this design. The cumulative-degree ablation should show a steeper apparent curve than the windowed version if the artifact matters. Held-out log-likelihood will rank the full, kernel-only, fitness-only, and frailty models. The negative-control targets should give a flat curve. Fitted models should reproduce the observed distribution of organ-system counts, judged by Kolmogorov–Smirnov distance. Replication on the 748,692-child chronic-conditions cohort (Healthcare utilization following SARS-CoV-2 infection in children and adolescents with chronic conditions) will test whether pre-existing conditions act through fitness or through the curve.

As an exploratory check only, we will ask whether a saturating curve with heterogeneous fitness can produce a bimodal split in organ count resembling the severe-versus-mild MIS-C classes (median five versus two systems). Bistability results for capacity-free network models (A Hierarchy of Network Models Giving Bistability Under Triadic Closure) suggest such a split is possible, but they explicitly lack carrying capacity, so we treat this check as hypothesis-generating rather than confirmatory.