Related Experiment Video
Updated: Jun 27, 2026
12:15
Optimized System for Cerebral Perfusion Monitoring in the Rat Stroke Model of Intraluminal Middle Cerebral Artery Occlusion
Published on: February 17, 2013
23.2K
A generalizable physiological model for detection of Delayed Cerebral Ischemia using Federated Learning
Ahmed Elhussein1, Murad Megjhani2, Daniel Nametz2
1Department of Biomedical Informatics, Columbia University, New York Genome Center, New York, NY, USA.
Proceedings. IEEE International Conference on Bioinformatics and Biomedicine
|February 23, 2024
Summary
Federated Learning (FL) trains a delayed cerebral ischemia (DCI) classifier across hospitals. FL improves DCI detection when data distributions are similar, but can hinder it if distributions differ significantly.
Area of Science:
- Neurology
- Artificial Intelligence
- Medical Informatics
Background:
- Delayed cerebral ischemia (DCI) is a critical complication following subarachnoid hemorrhage, often detected late and predicting poor patient outcomes.
- Machine learning shows promise for early DCI detection, but small sample sizes due to the condition's rarity hinder model training.
- Current data sharing limitations across institutions prevent the development of robust DCI prediction models.
Purpose of the Study:
- To develop and evaluate a Federated Learning (FL) approach for training a DCI classifier across multiple institutions.
- To address the challenge of limited data availability for rare conditions like DCI by enabling collaborative model training without direct data sharing.
- To investigate the impact of feature distribution similarity on the performance of FL models in DCI prediction.
Main Methods:
- A Federated Learning framework was implemented for collaborative training of a DCI classifier across three distinct institutions.
- A federated feature selection method was developed and integrated into the FL pipeline.
- A federated ensemble classifier was constructed and its performance compared against models trained independently at each site.
Main Results:
- The FL model demonstrated significant performance improvements at two of the three participating sites.
- Performance gains were correlated with the similarity of feature distributions across institutions.
- Heterogeneous feature distributions across sites led to performance degradation in some cases, highlighting potential challenges of FL.
Conclusions:
- Federated Learning offers a viable strategy for enhancing DCI detection by leveraging multi-institutional data.
- The effectiveness of FL is contingent upon the similarity of data distributions between participating sites.
- Assessing dataset distribution similarity prior to FL implementation is crucial for optimizing model performance and avoiding potential negative impacts.
