Federated learning using flower framework for enhanced medication safety in intensive care units

Document Type

Article

Publication Title

International Journal of Data Science and Analytics

Abstract

Errors in prescribing medications to patients in intensive care units (ICUs) pose a danger to their health. This paper presents a privacy-preserving federated learning paradigm that provides a strategy to minimize these medication prescription errors. A federated learning model that synthesizes knowledge from several ICU facilities while maintaining patient data privacy is proposed. The proposed study used the Flower framework with differential privacy that would help doctors and nurses choose the right medicines for critically ill patients. Incorporating patient-specific data from several ICUs into the proposed model enables the creation of a powerful predictive system to detect medication prescription errors. The model was trained with a neural network with three fully connected layers with ReLU and Sigmoid activation functions. Various experiments were conducted considering different numbers of ICUs. The added differential privacy would protect the model parameters sent from client to server from various attacks. The model achieves an accuracy of 98% and an F1-score of 90% with minimum losses. The accuracy gradually drops from 98.35 to 94.4% when the noise multiplier is increased from 0.1 to 15. This validates a trade-off in differential privacy as the model accuracy decreases with more privacy. Also, the model is validated with the bootstrapping technique, and performance is observed. This study resolves a significant healthcare problem by collaborating with several ICUs at a time and providing a generalized model for all ICUs. This discovery paves the way for using federated learning in advanced medical systems to lower pharmaceutical prescription errors and enhance patient care.

First Page

7039

Last Page

7053

DOI

10.1007/s41060-025-00877-x

Publication Date

12-1-2025

This document is currently not available here.

Share

COinS