About

Jason K.C. Tong, MD, MSHP, is an Assistant Professor of Surgery at the Hospital of the University of Pennsylvania, within the Department of Surgery. He completed his undergraduate studies at Dartmouth College in 2012, majoring in Biology with a minor in Religion. He earned his MD from the Perelman School of Medicine at the University of Pennsylvania in 2017 and subsequently obtained a Master of Science in Health Policy (MSHP) from the same institution in 2022. His professional focus includes surgical management, with particular expertise in colon and rectal surgery. Dr. Tong has contributed to research on various topics such as the management of chronic mesh erosion, outcomes following inflammatory bowel disease surgeries, and health system structures addressing structural racism. His work also involves evaluating surgical outcomes, healthcare costs, and disparities, with a strong emphasis on improving patient care and advancing surgical practices.

Research topics

• Pathology
• Virology
• Medicine

Selected publications

• A latent transfer learning method for estimating hospital-specific post-acute healthcare demands following SARS-CoV-2 infection

  Patterns · 2025-01-23

  erratumOpen access

  Publisher of over 50 scientific journals across the life, physical, earth, and health sciences, both independently and in partnership with scientific societies including Cell, Neuron, Immunity, Current Biology, AJHG, and the Trends Journals.

  PublisherDOI

• A communication-efficient federated learning algorithm to assess racial disparities in post-transplantation survival time

  Journal of the American Medical Informatics Association · 2025-08-02 · 1 citations

  article

  OBJECTIVE: Patients of different race have different outcomes following renal transplantation. Patients of different race also undergo renal transplantation at different hospitals. We used a novel decentralized multisite approach to quantitatively assess the effect of site of care on racial disparities between non-Hispanic Black (NHB) and non-Hispanic White (NHW) patients in post-transplantation survival times. MATERIALS AND METHODS: In this study, we develop a communication-efficient federated learning algorithm to assess site-of-care associated racial disparities based on decentralized time-to-event data, called Communication-Efficient Distributed Analysis for Racial Disparity in Time-to-event Data (CEDAR-t2e). The algorithm includes 2 modules. Module I is to estimate the site-specific proportional hazards model for time-to-event outcomes in a distributed manner, in which the Poissonization is used to simplify the estimation procedure. Based on the estimated results from Module I, Module II calculates how long the kidney failure time of NHB patients would be extended had they been admitted to transplant centers in the same distribution as NHW patients were admitted. RESULTS: With application to United States Renal Data System data covering 39 043 patients across 73 transplant centers, we found no evidence suggesting the presence of site-of-care associated racial disparities in post-transplantation survival times. In particular, restricting to one year after transplantation, the counterfactual graft failure time would have been extended by only 0.61 days on average if NHB had the same admission distribution to transplant centers as NHW patients. DISCUSSION: The proposed approach offers a quantitative measure to evaluate site-of-care associated racial disparities. CONCLUSION: Our approach has the potential to be extended to investigate site-of-care related disparities in other time-to-event outcomes, thus promoting health equity and improving patient health in various fields.

  PublisherDOI

• Semiparametric sieve estimation for survival data with two-layer censoring

  Biometrika · 2025-01-01 · 1 citations

  article

  Summary Disease registry data provide important information on the progression of disease conditions. However, reports of death or drop-out of patients enrolled in the registry are always subject to a noticeable delay. Reporting delays, together with the administrative censoring that arises from a freeze date in data collection, lead to two layers of right censoring in the data. The first layer results from random drop-out and acts on the survival time. The second layer is the administrative censoring, which acts on the sum of the reporting delay and the minimum of the survival time and random drop-out time. The heterogeneities among patients further complicate data analysis. This paper proposes a novel semiparametric sieve method based on phase-type distributions, in which covariates can be readily accommodated by the accelerated failure time model. A well-orchestrated EM algorithm is developed to compute the sieve maximum likelihood estimator. We establish the consistency and rate of convergence of the proposed sieve estimators, as well as the asymptotic normality and semiparametric efficiency of the estimators for the regression parameters. Comprehensive simulations and a real example of lung cancer registry data are used to demonstrate the proposed method. The results reveal substantial biases if reporting delays are overlooked.

  PublisherDOI

• Meta-Analysis and Federated Learning over Decentralized Distributed Research Networks

  Annual Review of Biomedical Data Science · 2025-08-11 · 2 citations

  reviewOpen access

  Distributed research networks have transformed modern clinical research by enabling large-scale, multi-institutional collaborations while maintaining patient privacy. Two prominent methodologies within these frameworks-meta-analysis and federated learning-address the challenges of synthesizing evidence from decentralized data. Meta-analysis aggregates study-level results to provide robust, interpretable estimates, making it a cornerstone of evidence synthesis for association studies. Federated learning complements this by enabling complex downstream tasks, such as predictive modeling and counterfactual inference, while preserving data privacy through privacy-preserving distributed algorithms. Federated learning facilitates communication-efficient computation and adapts seamlessly to heterogeneous datasets across diverse institutions. This review emphasizes the complementary strengths of federated learning's scalability, flexibility, and readiness for implementation alongside meta-analysis's robust frameworks for evidence synthesis and aggregation in clinical research. Integrations of synthetic data, artificial intelligence (AI)-enhanced harmonization, and hybrid human-AI frameworks are proposed as future directions, promising to further advance both methodologies and enhance their combined impact on privacy-conscious, data-driven healthcare research.

  PublisherDOI

• Splenectomy for immune thrombotic thrombocytopenic purpura: a systematic review and meta-analysis

  Journal of Thrombosis and Haemostasis · 2025-07-18

  review
  PublisherDOI

• Unlocking efficiency in real-world collaborative studies: a multi-site international study with one-shot lossless GLMM algorithm

  npj Digital Medicine · 2025-07-18 · 1 citations

  articleOpen access1st authorCorresponding

  The widespread adoption of real-world data has given rise to numerous healthcare-distributed research networks, but multi-site analyses still face administrative burdens and data privacy challenges. In response, we developed a Collaborative One-shot Lossless Algorithm for Generalized Linear Mixed Models (COLA-GLMM), the first-ever algorithm that achieves both lossless and one-shot properties. COLA-GLMM ensures accuracy against the gold standard of pooled data while requiring only summary statistics and completes within a single communication round, eliminating the usual back-and-forth overhead. We further introduced an enhanced version that employs homomorphic encryption to reduce the risks of summary statistics misuse at the coordinating center. The simulation studies showed near-exact agreement with the gold standard in parameter estimation, with relative differences of 7.8 × 10−6%–3.0% under various cell suppression settings. We also validated COLA‑GLMM on eight international decentralized databases to identify risk factors for COVID‑19 mortality. Together, these results show that COLA‑GLMM enables accurate, low‑burden, and privacy-preserving multi‑site research.

  PublisherOA PDFDOI

• Evaluating the Bias, type I error and statistical power of the prior Knowledge-Guided integrated likelihood estimation (PIE) for bias reduction in EHR based association studies

  Journal of Biomedical Informatics · 2025-02-02

  articleOpen access

  • Question: How does PIE perform in various types of real-world scenarios, in terms of estimation and hypothesis testing? • Findings: Under non-differential misclassification, PIE had a smaller bias in estimated associations compared to the naïve method, but it had similar type I error and power. • The bias reduction of PIE was superior when the prior distribution of sensitivity and specificity of the phenotyping algorithm is more accurate (i.e., close to the true operating characteristics of the phenotyping algorithm). The impact of prior is relatively small when the outcome has low prevalence and is larger when the outcome is common. • PIE can effectively reduce the bias due to phenotyping error under a wide spectrum of real-world settings. However, its main advantage is in the reduction of bias in estimation but not in hypothesis testing. Binary outcomes in electronic health records (EHR) derived using automated phenotype algorithms may suffer from phenotyping error, resulting in bias in association estimation. Huang et al. [1] proposed the Prior Knowledge-Guided Integrated Likelihood Estimation (PIE) method to mitigate the estimation bias, however, their investigation focused on point estimation without statistical inference, and the evaluation of PIE therein using simulation was a proof-of-concept with only a limited scope of scenarios. This study aims to comprehensively assess PIE’s performance including (1) how well PIE performs under a wide spectrum of operating characteristics of phenotyping algorithms under real-world scenarios (e. g., low prevalence, low sensitivity, high specificity); (2) beyond point estimation, how much variation of the PIE estimator was introduced by the prior distribution; and (3) from a hypothesis testing point of view, if PIE improves type I error and statistical power relative to the naïve method (i.e., ignoring the phenotyping error). Synthetic data and use-case analysis were utilized to evaluate PIE. The synthetic data were generated under diverse outcome prevalence, phenotyping algorithm sensitivity, and association effect sizes. Simulation studies compared PIE under different prior distributions with the naïve method, assessing bias, variance, type I error, and power. Use-case analysis compared the performance of PIE and the naïve method in estimating the association of multiple predictors with COVID-19 infection. PIE exhibited reduced bias compared to the naïve method across varied simulation settings, with comparable type I error and power. As the effect size became larger, the bias reduced by PIE was larger. PIE has superior performance when prior distributions aligned closely with true phenotyping algorithm characteristics. Impact of prior quality was minor for low-prevalence outcomes but large for common outcomes. In use-case analysis, PIE maintains a relatively accurate estimation across different scenarios, particularly outperforming the naïve approach under large effect sizes. PIE effectively mitigates estimation bias in a wide spectrum of real-world settings, particularly with accurate prior information. Its main benefit lies in bias reduction rather than hypothesis testing. The impact of the prior is small for low-prevalence outcomes.

  PublisherDOI

• Retiring the Term “Weighted Mean Difference” in Contemporary Evidence Synthesis

  Cochrane Evidence Synthesis and Methods · 2025-09-01 · 2 citations

  articleOpen accessSenior authorCorresponding

  Evidence synthesis frequently involves quantitative analyses of continuous outcomes. A cross-sectional study examining Cochrane systematic reviews identified 6672 out of 22,453 meta-analyses (29.7%) involved continuous outcomes [1]. The primary effect measures employed in meta-analyses of continuous outcomes are the mean difference (MD) and standardized mean difference (SMD) [2]. The MD is appropriately applied when all included studies measure outcomes using identical scales (e.g., body weight in kilograms). In contrast, the SMD serves as a solution when studies utilize different measurement scales (e.g., varied questionnaire scoring methods). Although alternative measures (e.g., the ratio of means) exist [3], their application remains relatively infrequent. Despite this conceptual clarity, the term “weighted mean difference” (WMD) appears frequently in the systematic review literature [4], which can lead to confusion about its relationship to the MD. In this article, we first clarify the distinction between MD and WMD, then describe the historical factors underlying the term's adoption and persistence, discuss why contemporary methods render it unnecessary, illustrate examples of misuse, and conclude with practical recommendations for clearer reporting. The MD represents the straightforward difference between group means (e.g., intervention vs. control) for a continuous outcome. Although the true MD value relates to unknown population-level differences, practical research relies on sample estimates from individual studies. Meta-analysis systematically synthesizes these study-level MD estimates to derive an overall summary effect across studies. The term WMD emerged historically to emphasize the weighted averaging process of meta-analyses, wherein each study contributes a sample MD weighted by its statistical precision (i.e., inverse variance) [5]. Typically, larger studies with smaller variances or narrower confidence intervals are assigned greater weights. Traditional meta-analytical methods, performed through either fixed-effect (also known as common-effect) or random-effects models, follow this inverse-variance weighting principle. Under fixed-effect models, study weights directly reflect the inverse of their variances, whereas random-effects models incorporate both within-study and between-study variances. To contextualize the widespread adoption of WMD, we conducted a brief literature search using Google Scholar on June 12, 2025. Using exact-phrase queries in quotation marks, for each calendar year from 1990 to 2024, we recorded the counts for “weighted mean difference” AND “systematic review” and separately for “systematic review,” then calculated the yearly proportion (Figure 1). Google Scholar indexes titles, abstracts, and, when available, full texts, so counts reflect occurrences anywhere in the indexed record, and these counts are approximate. We did not screen individual records for correct versus incorrect usage because our objective was to describe the prevalence of terminology rather than to quantify misuse. We therefore documented the evolution of usage over time in the proportions reported in Figure 1. This analysis revealed an observable increase in WMD usage around 1996, closely following the establishment of the Cochrane Database of Systematic Reviews (CDSR) in April 1995. The influential role of Cochrane reviews likely contributed greatly to disseminating this terminology. Chapter 6.5 of the latest Cochrane Handbook [6] confirms the prevalence of the term “weighted mean difference” in early editions of CDSR, with such cautionary notes appearing in handbook versions since at least 2008: “Analyses based on this effect measure have historically been termed [WMD] analyses in the [CDSR]. This name is potentially confusing: although the meta-analysis computes a weighted average of these differences in means, no weighting is involved in the calculation of a statistical summary of a single study. Furthermore, all meta-analyses involve a weighted combination of estimates, yet we don't use the word ‘weighted’ when referring to other methods.” Another plausible factor contributing to the continued use of the term is the citation of Andrade's statement in 2020 that “the pooled MD is more accurately described as a weighted mean difference or WMD.” [7] While this interpretation is not technically incorrect in describing the statistical process behind meta-analytic pooling, it may inadvertently encourage broader or careless use of the term WMD. Despite existing notes on WMD in the literature, Figure 1 illustrates the continued widespread use of WMD. Specifically, while the total number of systematic review publications increased until peaking around 2018 and then declined (Figure 1C), both the number and proportion of publications mentioning WMD continued to rise through 2024 (Figure 1A,B). Although the term is not misused in all instances, this trend suggests that existing cautions have had limited impact and underscores the value of clearer terminology. These historical and descriptive observations motivate a focus on current analytic practice and terminology, as discussed next. The explicit emphasis on weighting inherent to the term WMD can be misleading because weighting is fundamental to conventional meta-analytical methods, regardless of the outcome type (continuous, binary, time-to-event, etc.). Nevertheless, analogous terms such as “weighted odds ratio” or “weighted hazard ratio” are rarely used. Hence, more general terms such as “pooled MD,” “combined MD,” “overall MD,” or “meta-analytical MD” may be more appropriate and consistent. Moreover, contemporary methodological advancements in evidence synthesis frequently extend beyond traditional inverse-variance weighting. Modern meta-analyses, including pairwise and network applications, are often fit as one-stage generalized linear mixed or Bayesian hierarchical models in which treatment effects are estimated jointly from the likelihood [8-10]. In these models, precision is incorporated through the model structure rather than through explicit study-specific inverse-variance weights. Consequently, when outcome scales are identical, the pooled estimate is more clearly reported as a pooled MD or another clear descriptor, such as meta-analytic MD; the term WMD is unnecessary and may suggest a distinct effect measure. Imprecise usage nonetheless persists in current literature, as illustrated below. Critically, MD specifically pertains to individual study outcomes, while WMD exclusively represents the meta-analytical synthesis. Despite this clear distinction, some systematic reviews incorrectly label individual study effects as WMD [11-14]. For example, a systematic review published recently in JAMA inaccurately reported “pooled weighted mean differences” for systolic and diastolic blood pressures between screening and control groups [11]. Here, the pooled MD inherently indicates weighting, making the addition of “weighted” redundant and misleading. Moreover, a recent article in the American Journal of Ophthalmology captions a forest plot as “weighted mean differences (WMD) … across each study.” [12] Another applied paper captions a forest plot as “WMD and 95% CI,” both implying study-level WMDs [13]. In addition, a methods book chapter explicitly states, “Table 3.4 presents the WMD and the 95% confidence interval for each study.” [14] Such misuse persists in systematic reviews over time, including many published in various high-impact journals [15]. Labeling study-level effects as “WMD” can blur the distinction between a study's MD and the pooled meta-analytic estimate. For instance, a figure caption that states “WMD across each study” may suggest that each study yields a WMD rather than an MD, which can confuse evidence users about what is being pooled. Clearer labeling (e.g., “MD per study” with a “pooled MD”) reduces this risk and improves interpretability. This article underscores the potential inappropriateness of the term WMD, particularly its incorrect application to individual studies in evidence synthesis. Originating largely from early practices in Cochrane systematic reviews, WMD no longer aligns with contemporary methodological needs and rigor. Consequently, we recommend retiring the term WMD and adopting clearer terminology, using MD for study-level effects and pooled MD or meta-analytic MD for the synthesized estimate, to promote clearer, methodologically sound communication. Lifeng Lin: conceptualization, funding acquisition, investigation, writing – original draft, visualization, writing – review and editing. Xing Xing: investigation, writing – review and editing. Wenshan Han: data curation, writing – review and editing, visualization. Jiayi Tong: conceptualization, writing – review and editing. L.L. was supported in part by the US National Institute on Aging grant R03 AG093555, the US National Library of Medicine grant R21 LM014533, and the Arizona Biomedical Research Centre grant RFGA2023-008-11. The content is solely the responsibility of the authors and does not necessarily represent the official views of the US National Institutes of Health and the Arizona Department of Health Services. The authors declare no conflicts of interest. The peer review history for this article is available at https://www.webofscience.com/api/gateway/wos/peer-review/10.1002/cesm.70051. Data sharing is not applicable to this article as no data sets were generated or analyzed during the current study.

  PublisherOA PDFDOI

• Association of Pre–COVID-19 Body Mass Index with Postacute Cardiovascular, Gastrointestinal, and Neuropsychiatric Outcomes Among Children and Young Adults: An EHR-Based Cohort Study from the RECOVER Initiative

  SSRN Electronic Journal · 2025-01-01

  preprintOpen access
  PublisherDOI

• COLA-GLM: collaborative one-shot and lossless algorithms of generalized linear models for decentralized observational healthcare data

  npj Digital Medicine · 2025-07-15 · 2 citations

  articleOpen access

  Clinical insights from real-world data often require aggregating information from institutions to ensure sufficient sample sizes and generalizability. However, patient privacy concerns only limit the sharing of patient-level data, and traditional federated learning algorithms, relying on extensive back-and-forth communications, can be inefficient to implement. We introduce the Collaborative One-shot Lossless Algorithm for Generalized Linear Models (COLA-GLM), a novel federated learning algorithm that supports diverse outcome types via generalized linear models and achieves results identical to a pooled patient-level data analysis (lossless) with only a single round of aggregated data exchange (one-shot). To further protect aggregated institutional data, we developed a secure extension, secure-COLA-GLM, utilizing homomorphic encryption. We demonstrated the effectiveness and lossless property of COLA-GLM through applications to an international influenza cohort and a decentralized U.S. COVID-19 mortality study. COLA-GLM and secure-COLA-GLM offer a scalable, efficient solution for decentralized collaborative learning involving multiple data partners and diverse security requirements.

  PublisherOA PDFDOI

Frequent coauthors

• Yong Chen

  Chongqing Normal University

  106 shared

• Rui Duan

  Harvard University

  31 shared

• Chongliang Luo

  Washington University in St. Louis

  25 shared

• Yiwen Lu

  Applied Mathematics (United States)

  24 shared

• Jiang Bian

  Microsoft Research (United Kingdom)

  23 shared

• Christopher B. Forrest

  Children's Hospital of Philadelphia

  22 shared

• Nathan M. Pajor

  Cincinnati Children's Hospital Medical Center

  18 shared

• Bingyu Zhang

  California University of Pennsylvania

  16 shared

Labs

• Department of Surgery, University of PennsylvaniaPI

Education

• Ph.D. student, Department of Biostatistics

  University of Pennsylvania