Ensemble Learning Using Individual Neonatal Data for Seizure Detection.

Area of Science:

• Medical informatics
• Machine learning in healthcare

Background:

• Sharing medical data between institutions is hindered by data protection laws and institutional procedures.
• This data sharing limitation often results in smaller datasets for training algorithms, potentially reducing prediction accuracy.
• Electroencephalogram (EEG) analysis is particularly affected by these data limitations.

Purpose of the Study:

• To develop and evaluate an ensemble learning approach for neonatal EEG analysis that does not require inter-institutional data sharing.
• To compare the performance of different data aggregation schemes for ensemble models.

Main Methods:

• Simulated a scenario where data cannot be shared by splitting a public EEG dataset into institution-specific subsets.
• Trained local detectors within each simulated institution.
• Aggregated local predictions using four schemes: majority vote, mean, weighted mean, and Dawid-Skene.
• Validated the ensemble method on an independent dataset using a subset of EEG channels.

Main Results:

• Ensemble models achieved accuracy comparable to a single model trained on the entire dataset, provided sufficient data was available locally.
• The weighted mean aggregation scheme demonstrated the best performance.
• The Dawid-Skene method showed comparable performance to the weighted mean when local detectors approached the accuracy of a centralized model.

Conclusions:

• Ensemble learning offers a viable solution for training reliable neonatal EEG analysis algorithms without the need to share sensitive EEG data.
• The weighted mean aggregation strategy is effective for combining predictions from distributed local models.