Development and validation of a federated learning framework for detection of subphenotypes of multisystem inflammatory syndrome in children

Naimin Jing, Xiaokang Liu, Qiong Wu, Suchitra Rao, Asunción Mejías, Mitchell Maltenfort, Julia Schuchard, Vitaly Lorman, Hanieh Razzaghi, Ryan Webb, Chuan Zhou, Ravi Jhaveri, Grace M. Lee, Nathan M. Pajor, Deepika Thacker, L. Charles Bailey, Christopher B. Forrest, Yong ChenmedRxiv 2024 · DOI

TL;DR

Researchers have developed a new federated learning framework called dMLCA (distributed Multi-site Latent Class Analysis) to identify clinical subphenotypes of Multisystem Inflammatory Syndrome in Children (MIS-C). By analyzing data from nine hospitals without sharing sensitive individual patient records, the study identified three distinct clinical profiles—ranging from mild to multi-organ failure—providing a roadmap for more personalized pediatric care.

What Problem Does This Paper Address?

Multisystem Inflammatory Syndrome in Children (MIS-C) is a rare but life-threatening condition that appears weeks after a SARS-CoV-2 infection. Because the clinical presentation is incredibly diverse—affecting everything from the heart to the skin—doctors struggle to predict which children will recover quickly and which will require intensive care.

To understand these patterns, researchers typically use Latent Class Analysis (LCA), a statistical method that groups patients into "subphenotypes" based on shared symptoms. However, LCA requires large datasets to be accurate. Because MIS-C is rare, individual hospitals often have too few cases (some as few as 32) to generate reliable results. While the obvious solution is to pool data from multiple hospitals, strict privacy regulations and institutional "data silos" often make sharing individual-level patient data impossible.

Methods

The authors introduced dMLCA, a federated learning approach designed specifically for unsupervised clustering across multiple sites.

Key technical highlights include:

  • Lossless Integration: The algorithm uses a distributed Expectation-Maximization (EM) framework. The authors proved that the mathematical updates for the model are "exactly decomposable" by site. This means a lead site can calculate global parameters using only aggregated summary statistics from other hospitals, achieving the exact same accuracy as if all data had been pooled into one folder.
  • Heterogeneity-Awareness: Unlike standard meta-analysis, dMLCA accounts for the fact that different hospitals see different types of patients. It allows for "site-specific mixing proportions," acknowledging that one hospital might have a higher percentage of severe cases than another due to local demographics or referral patterns.
  • Privacy-Preserving: No patient-level data (EHR) ever leaves the local hospital's firewall; only encrypted or aggregated summary statistics are exchanged.

Datasets

The study utilized the PEDSnet network, a national collaboration of pediatric health systems in the U.S.

  • Primary Cohort: 864 hospitalized MIS-C patients (under age 21) across nine major children’s hospitals (including CHOP, Cincinnati Children’s, and Stanford) between March 2020 and December 2021.
  • Validation Cohort: A massive dataset of 59,740 children who tested positive for acute COVID-19 but did not have MIS-C. This was used to ensure the identified subphenotypes were unique to MIS-C and not just general "sick child" patterns.

Key Insights and Ideas

  • MIS-C is not a Monolith: The study confirmed that MIS-C isn't just one disease; it’s a spectrum. The "aha moment" here is that the subphenotypes are remarkably consistent across different hospitals, even though the proportion of patients in each group varies wildly by location.
  • Distinct from Acute COVID: By comparing the MIS-C groups to the general COVID-19 cohort, the researchers visualized a clear "distance" between the two (using fixation index mapping). This proves that MIS-C subphenotypes are de novo—they are unique clinical entities that require their own specific diagnostic criteria.
  • The "Site Effect": The algorithm revealed substantial heterogeneity across hospitals. For example, in "Site A," 42.7% of patients fell into the severe Class 2, while in "Site C," almost no patients did. This highlights how local hospital protocols or patient populations can bias multi-site studies if not properly accounted for.

Key Results

The dMLCA model identified three distinct subphenotypes:

  1. Class 1 (46.1% - Mild): Patients with minimal cardiac involvement who generally did not require intensive care.
  2. Class 2 (25.3% - Intermediate/Shock): High risk of shock, cardiac, and renal involvement, with a medium risk of respiratory symptoms.
  3. Class 3 (28.6% - Severe/Multi-organ): The most critical group, requiring intensive care and exhibiting failure in more than four organ systems (respiratory, GI, renal, hematologic, and skin).

Quantitatively, the authors demonstrated that the dMLCA estimates for manifest variables were identical to centralized "pooled" data analysis, proving the framework's efficiency.

Limitations and Future Work

While dMLCA is a major step forward for rare disease research, the authors noted several areas for improvement:

  • Communication Overhead: The current EM-based algorithm requires multiple rounds of back-and-forth communication between the lead site and participating hospitals. Future work could focus on "few-shot" algorithms to reduce this traffic.
  • Sample Size: Even with nine hospitals, 864 patients is a "moderate" sample size for such a complex condition. Validation in even larger, perhaps international, cohorts is needed.
  • Advanced Privacy: While aggregated data is shared, the authors suggest integrating more formal privacy techniques like differential privacy or homomorphic encryption in future iterations to further harden the system against potential data leakage.

Summary updated

Abstract

Our identified three MIS-C subphenotypes have profound implications for personalized treatment strategies, potentially influencing clinical outcomes. Further, the proposed algorithm facilitates federated subphenotyping while accounting for the heterogeneity across hospitals.
BibTeX
@misc{jing2024development,
  author = {Naimin Jing and Xiaokang Liu and Qiong Wu and Suchitra Rao and Asunción Mejías and Mitchell Maltenfort and Julia Schuchard and Vitaly Lorman and Hanieh Razzaghi and Ryan Webb and Chuan Zhou and Ravi Jhaveri and Grace M. Lee and Nathan M. Pajor and Deepika Thacker and L. Charles Bailey and Christopher B. Forrest and Yong Chen},
  title = {{Development and validation of a federated learning framework for detection of subphenotypes of multisystem inflammatory syndrome in children}},
  year = {2024},
  doi = {10.1101/2024.01.26.24301827},
  howpublished = {medRxiv},
}
Figure