Federated Knowledge-Driven Biomedical Dataset for Heterogeneous Modeling of Neonatal Drug-Adverse Reaction Associations

The Federated Knowledge-Driven Biomedical Dataset for Heterogeneous Modeling of Neonatal Drug-Adverse Reaction Associations comprises a meticulously curated and enriched dataset intended for advanced analysis and predictive modeling of neonatal Adverse Drug Reactions (ADRs). This dataset has been derived from the FDA Adverse Event Reporting System (FAERS), with a targeted focus on neonatal cases across 140+ high-risk health conditions.

Key Statistics:
Total Instances (Rows): 203616
Total Unique Instances: 118512
Number of Unique Drugs: 1,976
Number of Unique Adverse Reactions: 4,993
Total Drug-ADR Associations: 203,616
Unique Drug-ADR Pairs: 118,512

The dataset is well-suited for federated learning frameworks, enabling distributed predictive analytics from multiple biomedical sources while ensuring patient privacy. Each instance is anonymized and structured in a format conducive for training machine learning models across nodes without centralizing sensitive data. Given the structured, multi-relational nature of drug–ADR associations, the dataset is well-suited for heterogeneous graph modeling. It supports the following;

Knowledge graph construction
Link prediction
Multi-hop reasoning
Embedding-based learning
Subgraph extraction for focused ADR clusters

The dataset package includes:
The full dataset (CSV/XLS formats)
Python scripts for data parsing, preprocessing, exploratory data analysis, and graph generation