Separating Flares from Follow-ups in Pediatric Post-COVID Care

After a SARS-CoV-2 infection, children with chronic conditions use much more hospital care. In the large pediatric network PEDSnet, inpatient admissions roughly double and emergency visits rise by about half. Nobody yet knows what that excess is made of. One reading is biological: a post-viral flare in which one urgent visit is followed by more urgent visits for the same organ system. The other reading is behavioral: worried parents and cautious clinicians book extra check-ups and admit at a lower threshold, and this effect should fade as uncertainty about COVID fades. The count models used so far collapse months of follow-up into a single rate ratio, so they cannot tell these two stories apart.

We propose to model each child's encounter history as a self-exciting point process. This is a model of event timing in which every visit can temporarily raise the chance of further visits, on top of a slowly varying background rate. The quantity of interest is a within-child, before-versus-after-infection change in which visits trigger which other visits. The two explanations leave different temporal fingerprints. Flares should show up as stronger triggering from unscheduled visits to unscheduled visits within one organ system, and that triggering should grow with acute severity. Caution should show up as a raised background rate and generic visit-to-outpatient triggering that weakens with calendar era. The aim is to estimate the share of excess pediatric post-acute utilization attributable to each mechanism, which is a question about care planning rather than a leaderboard number.

Core Claim

Excess post-acute utilization in children can be split into three parts: a shift in background visit rate, a change in visit-to-visit triggering, and a transient kick that starts at the infection date and decays afterward. These parts can be identified from timestamped encounters if two conditions hold. Each child's own 18-month pre-infection history serves as their reference, and per-child heterogeneity in baseline rate is modeled explicitly. The heterogeneity matters because, without it, children who simply visit often at a steady rate look like children whose visits trigger more visits. The burstiness literature has long warned that apparent clustering can come from mixing heterogeneous rates. That confound is the main technical obstacle here, and the design is built around it.

Grounding in Prior Work

Prior work plays three roles here: some papers motivate the question, some supply methods, and some constrain the design.

The motivating evidence comes from the RECOVER pediatric studies. Pajor et al. fit propensity-matched negative binomial models to 748,692 children with chronic conditions. They report a total utilization rate ratio of 1.21, an inpatient ratio of 2.03, an emergency ratio of 1.54, and an inpatient ratio near 3.7 after severe acute illness. They also raise the caution hypothesis themselves, but their aggregated counts cannot test it. Lorman et al. scanned the ICD-10 hierarchy with a tree-based scan statistic and found 304 enriched diagnostic clusters, dominated by nonspecific symptoms such as dyspnea, chest pain and fatigue. That result tells us which organ-system groupings carry signal, but the study treats the post-acute window as a static presence/absence snapshot. We reuse its ICD-chapter grouping as our event vocabulary. Adult work points the same way. Zhang et al. find subphenotypes with Poisson factor analysis, which assumes memoryless counts. The 4CE temporal sub-phenotype study bins time into coarse windows beyond 90 days. Ribeiro et al. show that such coarse aggregation distorts the dynamics of time-varying networks.

The methods come from point-process work. Miscouridou et al. build multivariate Hawkes models that separate a base rate from mutual excitation. A Hawkes process is a point process in which each event adds a decaying bump to future event rates. De Bie et al. separate transient spikes from a Poisson background. Navaroli and Smyth warp clock time to remove deterministic rhythms before modeling stochastic burstiness. Our analogue is to absorb weekday, holiday and school-calendar effects into the background, so that a back-to-school check-up season is not read as self-excitation.

The closest existing work, which we must clearly differ from, is the Embedded Neural Hawkes Process of Zhao and Engelhard. It fits Hawkes processes with learned, possibly non-monotonic impact kernels to 34,000+ pediatric Duke records, studying perinatal triggers of neurodevelopmental diagnoses. It assumes a fixed base intensity, has no infection shock, and has no within-child pre/post contrast. Its question is trigger discovery, not decomposing an exposure's effect. Its finding that adding Transformer context gave only marginal likelihood gains is useful to us: it is evidence that pairwise kernels carry most of the signal. MEDCAS shows that visits form non-consecutive cascades and that patient histories are sparse, about 2.6 visits per patient in MIMIC-III. It is an adult readmission predictor with no exposure contrast. O'Connor et al. show day-to-day symptom triggering in adults with Long COVID using regression on survey data. That supports the plausibility of self-reinforcing flares but involves neither records nor children.

Malmgren et al. constrain the design. Their hidden-Markov account explains bursts through switching between latent active and dormant states rather than events triggering events. A biological flare could just as well be a latent state as a self-excitation, so we treat their model as a serious competitor and do not dismiss it.

Model

Events are encounters, each typed by $m = (e, s)$. Here $e$ is the encounter setting (outpatient, emergency, inpatient), and $s$ is the organ-system chapter of its primary diagnosis, following the grouping from Lorman et al.. This keeps the number of types $K$ around 45, which avoids the quadratic blow-up of Hawkes models over thousands of codes without needing learned embeddings. For child $i$ with index date $t_0$ (positive test, or pseudo-index for controls), the rate of type-$m$ encounters is

$$\lambda_{im}(t) = \exp\big(u_i + b_m + c_m(t) + \delta_m \mathbb{1}[t>t_0] + h_m(t-t_0)\big) + \sum_{t_k<t} \big(\alpha_{m_k m} + \Delta_{m_k m}\mathbb{1}[t_k>t_0]\big)\, g_{m_k m}(t-t_k).$$

The terms in the background part are as follows. $u_i$ is a per-child random effect on baseline rate (a frailty term). $b_m$ is a type intercept. $c_m(t)$ captures calendar effects such as weekday, holidays, school terms and pandemic era. $\delta_m$ is a persistent post-infection shift in background rate. $h_m$ is a decaying infection kick.

In the triggering part, $\alpha_{m'm}$ is the pre-infection triggering strength from type $m'$ to type $m$. $\Delta_{m'm}$ is its post-infection change. $g$ is a normalized kernel. We use a mixture of gamma shapes so that a delayed bump, such as a two-week follow-up, is representable, in the spirit of the non-monotonic kernels in ENHP.

Infected children and matched COVID-negative controls share all parameters except $\delta$, $h$ and $\Delta$. In controls these are fit and expected to be near zero, which gives a difference-in-differences contrast. Expected post-acute counts then decompose as

$$\mathbb{E}[N] = (I - A^{\text{post}})^{-1} \mu^{\text{post}},$$

where $A^{\text{post}}$ is the matrix of integrated triggering strengths $\alpha + \Delta$ and $\mu^{\text{post}}$ is the vector of integrated background counts. The identity is the standard branching representation of a stable Hawkes process. It lets us say how much of the excess over controls comes from $\delta$ (background), from $h$ (shock) and from $\Delta$ (triggering), by switching each off counterfactually. Fitting uses maximum likelihood with the frailty integrated out by a Laplace approximation. Uncertainty comes from a site-clustered bootstrap.

Research Program

The program has three stages.

The first stage is simulation, and it decides whether the approach can work at all. We generate PEDSnet-like cohorts at realistic sparsity, with a few visits per child per year. They come from three generators: heterogeneous-rate Poisson with no triggering, Hawkes with an increase in triggering, and a two-state latent flare model. We then ask whether our estimator recovers the true split, and whether simpler statistics confuse the three. The inter-visit burstiness coefficient and negative binomial overdispersion should fail to separate them, because all three produce heavy-tailed gaps.

The second stage fits the model on RECOVER PEDSnet, using the Pajor chronic-condition cohort definition and the 18-month washout.

The third stage tests pre-registered predictions that separate the two hypotheses. Under the flare hypothesis, $\Delta$ concentrates on emergency/inpatient-to-emergency/inpatient channels within the same organ system. It grows across acute-severity strata, matching the ~3.7 inpatient ratio. It is stable across calendar eras. Under the caution hypothesis, the excess concentrates in $\delta$ for outpatient types and in non-system-specific any-to-outpatient channels of $\Delta$. It shrinks in later eras, after the U09.9 code introduced in October 2021 and growing familiarity reduced uncertainty. Both can be partly true, and the output is a quantified split with intervals, not a verdict. PEDSnet does not flag scheduled versus unscheduled visits directly. We use encounter setting as the proxy and, where sites record appointment type, run a sensitivity analysis.

Baselines

The naive reference point is the negative binomial rate-ratio model of Pajor et al., refit on our exact cohort and split. Their reported figures are total URR 1.21 (95% CI 1.18–1.24), inpatient 2.03 and emergency 1.54. Reproducing these within interval is a prerequisite: if we cannot, our cohort differs and nothing downstream is comparable. As a predictive model it supplies a count distribution over any held-out window, against which we compare calibration of predicted 90-day encounter counts.

The strongest existing temporal method is ENHP from Zhao and Engelhard, run with its published architecture on the same typed events. It gets the same training children, the same compute, and a matched hyperparameter search of equal trial count. We found no reported result on Lacuna for ENHP on Duke-EHR pediatric log-likelihood or timing error, so no external number can be quoted. ENHP is the strongest competitor because it already has flexible kernels and handles pediatric records. If its fixed base rate and pooled kernels can absorb the infection effect, our extra structure is unnecessary. Beating it means two things. First, higher held-out log-likelihood per event on post-index encounters, with bootstrap intervals excluding zero. Second, and more importantly, correct recovery of the simulated split in stage one, where ENHP has no parameter that can represent it.

The latent-state competitor is a two-chain hidden-Markov model of active and dormant periods in the style of Malmgren et al., with infection allowed to change its transition rates. Beating it means better held-out likelihood and better calibration of the distribution of within-30-day revisit gaps.

Three ablations isolate the key ingredients. The first removes the frailty $u_i$; we expect it to inflate estimated triggering, reproducing the heterogeneity confound. The second removes the pre-infection anchor, so the model is fit on post-index data only and $\Delta$ is not identified separately from $\alpha$. The third removes the calendar term $c_m(t)$. Each is judged by bias in recovering the simulated split and by the shift in the PEDSnet split.

Why not just…?

Why not just fit the negative binomial model in fine time bins, stratified by encounter type? An event-study version of the Pajor et al. model gives the time profile of excess after infection, and it is cheap. It answers part of the question: whether excess decays or persists, and whether it differs by setting. But a rise in counts in a bin is equally consistent with a higher background rate and with visits triggering visits. Only conditioning on each child's own recent events separates those. If our decomposition adds nothing beyond the binned event study, the latter should be preferred. We therefore report both and check whether the era-by-setting pattern from binned counts already sorts the two hypotheses. Stage-one simulation shows directly whether binned counts can distinguish the Poisson and Hawkes generators.

Why not just run ENHP off the shelf? It is the closest method and a real threat on predictive likelihood. Its finding that context layers add little suggests pairwise kernels are enough. The difference is the estimand. ENHP's invariant base rate cannot express a post-infection background shift, and it has no per-child frailty. Pooled kernels fit to sparse pediatric histories, the cold-start problem that Hettige et al. document even in adults, will absorb rate heterogeneity as triggering. The frailty ablation and the simulation show whether that bias appears. If ENHP plus a simple infection indicator recovers the split as well as our model does, we adopt it.

Why not just model latent flare states instead of triggering? This could be the right answer biologically. A child may enter an inflamed period that raises urgent visits regardless of which visit came first, which is the latent-state picture of Malmgren et al.. Held-out likelihood and revisit-gap calibration settle which description fits better. If the hidden-Markov model wins, we restate the flare-versus-caution test in its terms: does infection raise entry into the urgent-visit state, with severity dose-response and era stability, versus raising outpatient background? We report that outcome rather than forcing a Hawkes interpretation.

Decisive Checks

Five results decide whether the proposal stands. First, in simulation at PEDSnet sparsity, the estimator must recover the background, shock and triggering shares within ten percentage points. Without the frailty term it should fail. If neither holds, the approach is not identifiable and we stop. Second, a placebo test: fit on COVID-negative controls with pseudo-index dates. Estimated $\delta$ and $\Delta$ must be indistinguishable from zero. Third, a severity gradient: triggering changes on urgent channels should be ordered by acute severity if flares are real. Fourth, an era test: the background and outpatient-triggering components should decline across pandemic eras if caution is real. Fifth, reproducibility across sites: the split should be stable when fit separately at each PEDSnet center. The large NYC-versus-Florida discordance reported by Zang et al. warns that access friction can masquerade as disease signal. A split that flips between sites would be read as access artifacts, not mechanism.