Projects & Collaboration
Active Research Programs
We welcome collaborations across institutions. Below is a summary of ongoing and completed projects using the PDA framework.
Replicating the RISK Study: Predicting Stricturing and Penetrating Behaviors in Pediatric Crohn’s Disease
https://pmc.ncbi.nlm.nih.gov/articles/PMC11002871/
Background
Stricturing and penetrating complications account for much of the morbidity and surgical burden in pediatric Crohn’s disease, and the prospective RISK study established which baseline factors predict them. Reproducing that result from routinely collected records would show that a distributed network can stand in for a purpose-built prospective cohort. The two complications compete with one another, so modeling either alone distorts both.
Method
ODACoR-H handles competing risks when sites disagree, allowing the baseline subdistribution hazard for each endpoint to vary by institution. A surrogate likelihood combines patient-level data at the local site with aggregated summaries from external sites, in one shot. Four national children’s hospitals within PEDSnet contributed, and the analysis followed the RISK study’s endpoint definitions so the two sets of results would be directly comparable.
Main finding
The subdistribution hazard ratios recovered from routine electronic health records were consistent with those reported by the original prospective RISK cohort, reproducing its risk stratification for penetrating and stricturing behavior. This supports using existing pediatric network data to replicate and extend prospective cohort findings in rare pediatric conditions where enrolling a new cohort is slow and expensive.
Evaluating Long-Term Cardiovascular Risk Following COVID-19 in a Pediatric Network
https://pubmed.ncbi.nlm.nih.gov/41259033/
Background
Whether SARS-CoV-2 infection leaves a durable cardiovascular signature in children was an open question, and answering it required both a large pediatric cohort and a way to handle the fact that participating institutions serve very different populations. Ignoring that heterogeneity risks reading institutional composition as a biological effect. The analysis drew on a United States pediatric centralized network covering 719,383 patients.
Method
COLA-GLM-H is the heterogeneity-aware member of the COLA family. It reconstructs the global likelihood for generalized linear models from institution-level summary statistics alone, while modeling site-specific effects directly rather than assuming a common intercept. Estimation is lossless and completes in one round. The same algorithm was exercised on an internationally decentralized network of 120,429 hospitalized patients across seven databases in three countries.
Main finding
Elevated long-term cardiovascular risk following infection was detectable in the pediatric cohort after case-mix adjustment, with the magnitude varying across contributing institutions. Because the site-specific effects were modeled rather than averaged away, the analysis could separate genuine post-infection risk from differences in which children each institution serves.
Hospital Profiling for COVID-19 Mortality Across International Health Systems
https://www.ohdsi.org/wp-content/uploads/2023/11/Nov7-Tong.pdf
Background
Hospital profiling asks whether patient outcomes differ across hospitals after accounting for differences in the patients they treat. This use case studied COVID-19 mortality among hospitalized adults from 12 OHDSI sites in the United States, Spain, and the Netherlands. Because sites differed in sample size, age distribution, and comorbidity burden, crude mortality rates alone could give a misleading picture of hospital performance.
Method
We applied a distributed hospital profiling framework that estimates risk-standardized mortality rates without sharing patient-level data. The framework combines dGEM, a communication-efficient distributed generalized linear mixed-effects model, with counterfactual standardization, which compares hospitals as if they had treated the same reference patient population. This enables privacy-preserving, risk-adjusted benchmarking across heterogeneous health systems.
Main finding
After case-mix adjustment, the 12 hospitals showed a stable tier-wise performance pattern across the two pandemic waves. Top, middle, and lower-performing groups remained visible even when individual ranks shifted. At the same time, 8 of 12 sites improved their risk-standardized mortality rates from Wave A to Wave B, suggesting broad improvement in COVID-19 inpatient outcomes over time.
Discovering Subphenotypes of Multisystem Inflammatory Syndrome in Children Across Nine Hospitals
https://pubmed.ncbi.nlm.nih.gov/38343837/
Background
Multisystem inflammatory syndrome in children is rare and clinically heterogeneous, with presentations overlapping Kawasaki disease and macrophage activation syndrome. Resolving it into subphenotypes would sharpen recognition and treatment. Latent class analysis is the natural tool, but it estimates many parameters and the smallest contributing hospital in this study identified only 32 cases, far short of what a site-level model would need.
Method
Running latent class analysis separately at each site fails because the number of classes need not agree and, even when it does, the classes are not comparable. The distributed multi-site latent class analysis approach instead learns a single set of latent classes shared across institutions while allowing the class proportions to differ by site. Conditions, laboratory results, and procedure information served as input features. Nine United States children’s hospitals in PEDSnet participated.
Main finding
From 864 identified cases drawn out of roughly 3.5 million patients, three clinically interpretable subphenotypes emerged. The first, covering about 46 percent of cases, presented mildly with minimal cardiac involvement and no intensive care requirement. The second, about 25 percent, carried high risk of shock along with cardiac and renal involvement. The remainder formed a third group with an intermediate profile.
Estimating Hospital-Specific Post-Acute Healthcare Demand Following SARS-CoV-2 Infection
https://pmc.ncbi.nlm.nih.gov/articles/PMC11573960/
Background
National estimates of post-acute healthcare demand are of limited use to a hospital deciding how to staff a follow-up clinic. Each institution sees a different patient mix and delivers care differently, so a pooled national figure can be wrong locally in both directions. The question here was how much additional healthcare each individual health system should expect its own patients to require after COVID-19 infection.
Method
The latent transfer learning framework proceeds in three stages. It first identifies latent patient subpopulations defined by multimorbidity patterns using data from all participating systems. It then performs causal estimation tailored to the target hospital’s own population. Finally it adaptively integrates information from the remaining hospitals, borrowing more where the populations resemble the target and less where they do not. Data came from eight PEDSnet health systems.
Main finding
Four distinct patient subpopulations emerged, with post-acute healthcare demand varying both across subpopulations and across hospitals. Patients with complex chronic conditions required more services after infection than other subpopulations. The heterogeneity across hospitals was large enough that a global or national estimate would be inadequate for local decision-making, which argues for stratified rather than uniform post-acute care planning.
Estimating the Effect of COVID-19 Vaccination on Post-Acute Sequelae Under High-Dimensional Confounding
https://pmc.ncbi.nlm.nih.gov/articles/PMC12269483/
Background
Whether vaccination reduces the risk of post-acute sequelae in pediatric populations was poorly characterized, and answering it requires a causal contrast rather than an association. Two obstacles compound each other. Electronic health record confounders number in the hundreds, and the sites contributing data differ in their covariate distributions, so an estimator tuned on one network’s population will not transport cleanly to another.
Method
DisC2o-HD estimates the average treatment effect for high-dimensional data while explicitly accounting for covariate shift across sites. Using the surrogate likelihood approach, it calibrates the propensity score and outcome models so that the desired covariate balancing property is approximately attained under the shifted distributions. The estimator remains consistent if either the propensity score model or the outcome regression is correctly specified, and it attains the semiparametric efficiency bound when both hold.
Main finding
Applied to RECOVER data, the method produced an average treatment effect for vaccination on post-acute outcomes in a diverse pediatric population, a setting where prior evidence had been thin. Because covariate distributions differed materially across contributing sites, an estimator that ignored that shift would have attributed part of the between-site composition difference to the vaccine itself.
Quantifying Risk Factors for Post-Acute Sequelae of SARS-CoV-2 Infection in Children and Adolescents
https://pubmed.ncbi.nlm.nih.gov/39489875/
Background
Post-acute sequelae of SARS-CoV-2 infection in children are individually uncommon and clinically diverse, which puts any single hospital far below the sample size needed for stable estimates. Two related questions were pursued: how much does infection raise the risk of specific syndromic and systemic outcomes, and how do risk factors behave once competing events are accounted for. Pediatric research networks supplied the scale.
Method
ODAP-B estimates adjusted relative risks through modified Poisson regression with a robust variance estimator, chosen because relative risk is more interpretable than an odds ratio when outcomes are rare. ODACoR extends the one-shot approach to competing risks, estimating subdistribution effects when several event types can preempt one another. The analyses covered 184,501 children across eight academic medical centers and, for the competing risks work, over 6.5 million pediatric patients.
Main finding
Elevated post-acute risk was detectable across both syndromic and systemic outcome categories, with age group, medical complexity, and intensive care during the acute episode among the factors carrying the strongest associations. Because event rates in several outcomes fall near half a percent, the pooled-equivalent estimates were the only ones stable enough to support these comparisons at the network level.
Identifying Clinical Risk Factors for Opioid Use Disorder Across a Statewide Research Network
Background
Which patients receiving an opioid prescription go on to develop opioid use disorder, and how soon? Five participating sites of the OneFlorida Clinical Research Consortium contributed records from 2012 through early 2019 for patients with at least one opioid prescription and no prior cancer or opioid use disorder diagnosis. A case was defined by a first diagnosis within twelve months of the initial prescription. The sites differ substantially in baseline prevalence.
Method
Four algorithms addressed different facets of the same question. ODACH fits a site-stratified Cox model for time to diagnosis, letting baseline hazards differ across sites. ODACAT and its heterogeneity-aware variant handle categorical and temporally resolved outcomes. ADAP fits penalized regression across roughly 15,000 patients when candidate predictors are numerous. A heterogeneity-aware inference framework examined the contribution of anxiety and depression specifically.
Main finding
Anxiety, alcohol use disorders, depression, sleep disorders, and other substance-related diagnoses emerged as consistent correlates of opioid use disorder alongside demographic and insurance factors. In the high-dimensional analysis, the variables selected under the distributed penalized model matched those recovered from pooled patient-level data, indicating that the risk profile is stable across the five sites rather than an artifact of any one population.