PFed-Signal: An ADR Prediction Model based on Federated Learning

Tao Li, Peilin Li, Kui Lu, Yilei Wang, Junliang Shang, Guangshun Li, Huiyu ZhouarXiv 2025 · DOI · arXiv

Abstract

Adverse Drug Reactions (ADRs) are a critical public health concern, yet the primary data source for monitoring them—the FDA’s Adverse Event Reporting System (FAERS)—is plagued by noise and structural bias. This paper introduces PFed-Signal, a framework that treats data bias as a "poisoning attack" within a Federated Learning (FL) environment. By analyzing model weight deviations, PFed-Signal identifies and removes biased data splits before training a high-performance Transformer model for ADR prediction. Experimental results demonstrate that PFed-Signal achieves superior accuracy (0.887) and AUC (0.954) compared to standard machine learning approaches, while also improving the reliability of traditional statistical safety signals.

Introduction

Adverse Drug Reactions (ADRs) are a leading cause of hospitalization and death worldwide. To monitor these events, the FDA maintains the FAERS database, a massive repository of spontaneous reports. However, FAERS is notoriously "noisy," suffering from under-reporting, duplicate entries, and incorrect labeling.

Traditional statistical metrics like the Reporting Odds Ratio (ROR) and Proportional Reporting Ratio (PRR) rely on raw counts; if the data is biased, the resulting safety signals are misleading. While machine learning models can identify patterns, they typically cannot correct the inherent structural bias of the data source. PFed-Signal addresses this by systematically cleaning the data using federated learning principles before the prediction phase begins.

The PFed-Signal Framework

The authors propose a two-stage architecture designed to purify and then analyze medical data. The first stage, Pfed-Split, simulates a federated environment by cleaning raw data and segmenting it into sub-datasets based on specific ADR types (e.g., headache, tachycardia, nausea). The second stage, ADR-Signal, performs the core purification and final prediction.

The end-to-end PFed-Signal architecture, from data segmentation to Transformer-based prediction.

Data Purification via Federated Learning

The core innovation of PFed-Signal is treating naturally occurring data bias as if it were a malicious "poisoning attack" in a federated network. Local binary classifiers are trained on different data splits. The system then compares the weights of these local models to a global model using Euclidean distance. If a local model’s weights deviate significantly from the global average (exceeding a threshold $\epsilon$), the underlying data split is flagged as "biased" and discarded.

Detailed Federated Learning process for identifying and removing biased datasets across multiple clients.

Once a "clean" dataset is generated, a multi-head attention Transformer model is trained. This model captures complex, non-linear correlations between 38 distinct drug features (such as patient age, dosage, and administration routes) and reaction intensities to provide the final ADR prediction.

Experimental Evaluation

The study utilized 5,868 adverse event records from the FAERS database spanning Q1 2010 to Q3 2024, covering 10 types of adverse reactions. PFed-Signal significantly outperformed baseline models across all major metrics after 100 training epochs:

Metric SVM BCPNN Random Forest PFed-Signal
Accuracy 0.813 0.855 0.841 0.887
AUC 0.884 0.845 0.891 0.954
F1 Score 0.854 0.862 0.885 0.895
Recall 0.784 0.818 0.801 0.820

Beyond predictive performance, the model achieved an identification accuracy of 0.97 in correctly flagging biased data points. Interestingly, the authors found that after removing the biased data, even traditional statistical methods (ROR and PRR) showed stronger, clearer signals, proving that data quality is the primary bottleneck in drug safety monitoring.

Conclusion and Future Work

PFed-Signal demonstrates that federated learning security techniques can be repurposed to solve data quality issues in centralized medical databases. By filtering out biased records, the framework provides a more reliable foundation for both statistical and deep learning-based ADR prediction.

Future work will focus on optimizing the distance threshold ($\epsilon$) to avoid discarding rare but valid data and scaling the system to handle the millions of records present in the full FAERS database. The authors also plan to enhance the Pfed-Split strategy to improve generalization across more heterogeneous medical datasets.

Summary updated

BibTeX
@misc{li2025pfed,
  author = {Tao Li and Peilin Li and Kui Lu and Yilei Wang and Junliang Shang and Guangshun Li and Huiyu Zhou},
  title = {{PFed-Signal: An ADR Prediction Model based on Federated Learning}},
  year = {2025},
  eprint = {2512.23262},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.23262},
}
Figure