Which Diagnoses Did the Infection Cause? Signed Hawkes Attribution for Pediatric Long COVID

Long COVID in children is hard to count. The dedicated diagnosis code is rarely used, children's symptoms are vague ("brain fog", fatigue, lingering cough), and clinic visits arrive irregularly, whenever symptoms flare. The standard epidemiological tools give population answers. One example is a hazard ratio, the ratio of disease risk in infected versus matched uninfected patients over a 30–180 day window. They cannot say which particular visit in a particular child's record was caused by the infection, and which was background illness that would have happened anyway. We propose a model that answers the event-level question.

The model is a self-exciting point process, a model of event streams in which each event raises the short-term rate of future events, in the way earthquakes trigger aftershocks. We make it "identity-revealing": every later diagnosis carries a probability of descending, possibly through a chain of intermediate diagnoses, from a specific root event such as a SARS-CoV-2 infection, an influenza infection, or a pre-existing chronic condition. The infection-rooted share of a child's record is then a per-patient, per-event Long COVID attribution that needs no diagnosis code. Two ingredients make this workable on real pediatric records. The first is a way to let some conditions suppress others without destroying the parent–child attribution. The second is a structured prior on which condition groups excite or suppress one another, borrowed from structural balance theory, whose role is tested rather than assumed.

Core Claim

Source attribution in self-exciting processes depends on a branching interpretation: each event has exactly one "parent", either the background rate or an earlier event. That interpretation exists only when all triggering effects are non-negative. Real clinical streams contain suppression. An acute hospitalization, for example, temporarily pushes down the rate of routine outpatient visits. Neural point processes model suppression, but they give up attribution. We claim that factoring the intensity into an excitatory branching part, which carries identity, and a multiplicative suppressive modulation, which rescales but never parents events, preserves exact attribution while fitting suppression. For condition category $k$, the event rate at time $t$ is

$$ \lambda_k(t) = \exp\Big(-\sum_{j:\,t_j<t} \beta_{k c_j}\, h(t-t_j)\Big)\Big(\mu_k + \sum_{j:\,t_j<t} \alpha_{k c_j}\, g_{c_j k}(t-t_j)\Big). $$

Here $t_j$ and $c_j$ are the time and category of past event $j$, $\mu_k$ is the background rate, $\alpha \ge 0$ and $\beta \ge 0$ are excitation and suppression strengths, $h$ is an exponential decay, and $g$ is a triggering kernel. For edges leaving the infection root, $g$ is the sigmoid-gated delayed kernel of Hawkes Process with Flexible Triggering Kernels, so that risk can ramp up weeks after the acute illness rather than peaking at once. Given the suppression factor, each event's parent is drawn in proportion to the terms inside the second bracket. Root identity propagates along parent chains exactly as in a branching process.

Where This Sits Relative to Prior Work

The identity-tracking idea comes from COEVOLVE (COEVOLVE; Farajtabar et al.). COEVOLVE follows which original source a Twitter cascade came from, and lets exposure raise the hazard of forming a direct link to that source. It recovered excitation parameters with rank correlation up to 0.82 and beat the traffic-shortcut baselines TRF and WENG by up to 2.5x in link-prediction success. We reuse its source tracking and its minorization–maximization style of fitting. COEVOLVE allows excitation only and has never been applied to clinical streams. We add suppression in a form that keeps the attribution. We also drop the network-growth half of COEVOLVE, because nothing in a medical record plays the role of a newly formed follow link.

Neural point processes constrain the design. The Neural Hawkes Process shows that inhibition matters: it reached roughly −8 cross-entropy per event on MemeTrack, against −800 for a classical Hawkes model. Transformer Hawkes Process improved likelihood further, for example −2.04 versus −5.60 log-likelihood per event on Retweets. Both models encode history in hidden states, so they cannot say which earlier event caused a later one. This is the gap the multiplicative factorization targets. SGHP supplies our delayed kernel and makes a parameter-efficiency argument (3.1k versus 417.8k parameters for THP), which matters for small pediatric cohorts. SGHP itself models flat event types with no source identity.

Structural balance theory enters as a prior, and this is its most speculative use in the proposal. Prediction and Clustering in Signed Networks shows that a balanced signed network splits into factions: positive edges inside a faction, negative edges between factions, which makes the signed matrix low-rank. We translate this into a clinical hypothesis. Condition categories form a few "factions". Within a faction, conditions mutually excite ($\alpha$ is active). Across factions, conditions suppress one another ($\beta$ is active), for instance through competition for care or a shift in which specialty sees the child. This shrinks the roughly $2K^2$ interaction parameters to faction memberships plus low-rank magnitudes.

Signed graph learners such as SiGAT (0.9333 sign-prediction area under the curve on Epinions), wsGAT (0.832 on Bitcoin-Alpha) and DynamiSE show that balance structure helps sign prediction on social graphs. All of them work on static graphs or binned snapshots. None has an event likelihood, and they say nothing about clinical data. DynamiSE also notes that balance gains shrink when positive edges dominate, which warns us that the prior may not pay off. The directional partial-balance measure of Structural balance in real-world social networks gives us a way to check whether a freely learned signed trigger matrix is more balanced than chance before we impose the prior.

Pediatric Long COVID is the application, and these studies motivate it. Clinical manifestations … in children aged 0–17 years in Beijing reports persistent symptoms in 55% of 2,001 children at 4–12 weeks, led by concentration difficulty (22.1%) and cough (22.1%). It found no significant association with vaccination ($P = 0.286$). The delayed onset is what motivates the gated kernel. LATCH shows how badly the diagnosis code U09.9 performs: F-score 15.9% against 75.4% for LATCH at the Veterans Health Administration, and 16.5% against 52.1% at UPMC. LATCH, however, classifies whole patients and does not attribute individual events. RECOVER produces inverse-probability-weighted hazard ratios, such as pulmonary fibrosis at aHR 2.49 in INSIGHT, but these are compressed into a single window. 4CE temporal sub-phenotypes clusters binned codes after the fact. It finds that neurological deficits appear earlier (mean 289 days) than fatigue or dyspnea (330 days), and it studies hospitalized, non-pediatric patients.

Research Program

We will build the model in three stages. The first is the attribution-preserving signed process, fit by expectation–maximization over latent parents. Because the suppression factor has a closed form under exponential decay, the likelihood integral stays tractable by piecewise integration; we avoid Monte Carlo, the scaling bottleneck that neural point processes face. The second stage adds multiple root identities: SARS-CoV-2, influenza, and respiratory syncytial virus infections, plus a "chronic baseline" root. Comparing infection roots lets us separate COVID-specific sequelae from generic post-viral sequelae, a question that matters clinically. The third stage adds the faction prior, fit jointly through soft faction memberships. Before using it we run a pre-registered diagnostic: fit the free signed model, measure partial balance against degree-preserving sign shuffles, and impose the prior only if the learned matrix is measurably balanced.

The main dataset is the pediatric (0–17) subset of the RECOVER EHR networks, INSIGHT and OneFlorida+, coded into the RECOVER diagnosis categories with infection dates as roots. The file does not report pediatric-specific counts for these networks, so cohort size is a first deliverable. MIMIC-II and the Duke COVID-19 emergency-department sequences serve as public benchmarks for the point process itself. Simulated cascades with known parents and signs serve for verifying recovery.

Baselines

The naive reference is a time-window rule: every new diagnosis 30–180 days after infection is attributed to the infection. At the patient level, a second naive reference is the presence of U09.9, which LATCH found reaches only about 16% F-score. Our model must beat the window rule on event-level precision in chart review, or it is not worth its complexity.

The strongest population-level competitor is RECOVER-style inverse-probability-weighted Cox screening against uninfected pediatric controls (RECOVER). We run it on exactly the same pediatric cohort, categories and window. The file has no reported pediatric result for this method; the adult INSIGHT figures (pulmonary fibrosis aHR 2.49, hair loss aHR 2.10) are context only. We do not treat it as something to beat but as a calibration target. Summed over patients, our infection-attributed event counts per category should match the excess events its weighted contrast implies.

The strongest patient-level competitor is LATCH, which reports F-scores of 75.4% at the Veterans Health Administration and 52.1% at UPMC without language processing. We retrain it on the same pediatric labels, without notes so the inputs match, and give it the same hyperparameter-search budget. Beating it means a higher F-score on chart-reviewed pediatric labels when patients are ranked by their infection-attributed event mass.

The strongest sequence-modeling competitors are NHP, THP and SGHP. They share our splits and get equal tuning budgets. On Retweets, THP's reported log-likelihood per event is −2.04. On MIMIC-II and the Duke ED data, the file reports no numeric likelihood for these models ("no reported result found on Lacuna" for those values), so we rerun them ourselves. We do not expect to beat THP on likelihood. The bar is to come within a small margin of its held-out likelihood while providing attribution it cannot. For NHP and THP attribution, we use gradient-based attribution of the infection token.

Three ablations isolate our ingredients. The first is the identical model with $\beta = 0$: an unsigned, COEVOLVE-style identity-revealing Hawkes process. The second replaces the faction prior with a generic low-rank-plus-sparsity penalty that has the same parameter count. The third replaces the gated delayed kernel with an exponential one.

Why not just…?

Why not just use the hazard-ratio screen RECOVER already runs? It answers the population question well, which is why we use it as the calibration anchor. It cannot rank individual children, cannot say whether new cough led to later fatigue, and cannot tell a COVID-rooted chain from an influenza-rooted one within a patient. If our attributions fail to reproduce its excess counts, our model is wrong. If they reproduce them, we offer strictly more.

Why not just use a Transformer Hawkes process, which fits better? It probably does fit better. Its hidden state, though, mixes the whole history, and gradient attribution is not a causal parent. The synthetic experiment settles this directly: with known parents, we compare parent-recovery accuracy between our model and THP or NHP attribution. If THP attribution matches ours, the factorization adds nothing and we would say so.

Why not just use an unsigned identity-revealing Hawkes process? This is the most serious threat. If suppression is rare in pediatric records, the $\beta = 0$ ablation will match us and should be preferred because it is simpler. We expect suppression to matter mainly around acute hospitalizations, so we test it on held-out likelihood and on attribution calibration in the time windows following admissions. If those tests show no gain, we drop the signed component.

Why not just use a generic low-rank penalty instead of balance theory? Balance is a social-psychology idea, and its relevance to medicine is unproven; DynamiSE already reports that its benefit fades under positive skew. That is why the faction prior is gated on the partial-balance diagnostic and compared against a matched low-rank penalty. If the two perform equally, we report that balance adds no information beyond rank, and we keep the low-rank version.

Decisive Checks

On simulated data, the model must recover infection-rooted parents and the signs of the trigger matrix markedly better than the window rule and the attributions from neural point processes. On pediatric RECOVER data, the infection-attributed event counts per category must fall inside the confidence intervals of the weighted excess estimates, while the window rule overshoots them. Influenza-rooted attributions act as a negative-control contrast. On 300 chart-reviewed pediatric records, attribution-based ranking must exceed the U09.9 baseline and match or beat LATCH. Finally, the learned delays from the infection root should place cognitive and respiratory onsets in the 4–12-week range reported in the Beijing cohort, without that range having been supplied to the model. Failure on the calibration check refutes the core claim. Failure of only the balance or suppression ablations reduces the contribution to an attribution-preserving delayed Hawkes model, which would still be the first event-level attribution of pediatric Long COVID.