Pre-trained knowledge transfer with application in multisystem inflammatory syndrome subphenotyping in children

Area of Science:

• Pediatric infectious diseases
• Computational biology
• Epidemiology

Background:

• Multisystem Inflammatory Syndrome in Children (MIS-C) is a severe post-COVID-19 condition with diverse clinical presentations.
• Identifying MIS-C subphenotypes is crucial for risk stratification and targeted therapy development.
• Analyzing combined data from multiple studies is ideal but hindered by privacy and heterogeneity issues.

Purpose of the Study:

• To develop a novel framework for analyzing multisite MIS-C data without sharing patient-level information.
• To enable knowledge transfer of pre-trained subphenotypes from large source studies to a target study.
• To account for and explain between-study heterogeneity using study-specific mixing proportions.

Main Methods:

• A one-shot summary statistics-based framework for knowledge transfer.
• Utilizing pre-trained subphenotypes from large-scale source studies.
• Employing study-specific subphenotype mixing proportions to model heterogeneity.

Main Results:

• The proposed framework successfully transfers knowledge of shared subphenotypes across studies.
• Study-specific mixing proportions effectively explain clinical heterogeneity between populations.
• Simulations and MIS-C data analysis confirm the method's benefits.

Conclusions:

• The framework offers a privacy-preserving approach for joint analysis of distributed health data.
• It facilitates the identification of MIS-C subphenotypes even with limited target study data.
• This method can advance understanding and treatment of MIS-C and similar complex pediatric conditions.