Learning competing risks across multiple hospitals: one-shot distributed algorithms.

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Analyzing the interplay of multiple clinical conditions, particularly in pediatric populations, presents significant statistical challenges.
• Post-acute sequelae of SARS-CoV-2 (PASC) in children and adolescents requires robust methods to identify risk factors and understand disease progression.
• Existing methods like meta-analysis may struggle with rare events and complex data structures from multi-institutional electronic health records.

Approach:

• Developed two novel one-shot distributed algorithms for competing risk models (ODACoR) to handle multi-institutional time-to-event data.
• Applied ODACoR to electronic health record (EHR) data from eight national children's hospitals, encompassing over 6.5 million pediatric patients.
• Validated ODACoR's accuracy and reliability against traditional meta-analysis and pooled data estimates through extensive simulation studies.

Key Points:

• ODACoR algorithms demonstrated substantially lower relative bias (∼0.2%) compared to meta-analysis (∼40%) for rare clinical conditions.
• ODACoR accurately identified risk factors for PASC, including age, gender, chronic conditions, and obesity, which were missed by meta-analysis.
• Algorithm performance was comparable to pooled data analysis, highlighting the reliability of this federated learning approach.

Conclusions:

• The proposed ODACoR algorithms are communication-efficient, highly accurate, and suitable for characterizing complex clinical condition interplay.
• ODACoR offers a powerful, scalable solution for multi-institutional time-to-event analysis, applicable to various clinical research questions.
• This approach enhances the ability to analyze large-scale pediatric EHR data for improved understanding of disease risk and outcomes.