Rationale:
Automated algorithms for EEG analysis hold the promise for greatly improving the speed and availability of diagnosis of seizures as well as offer the hope of finding new methods for improving treatment. While these methods can achieve high accuracies and low false positive rates in the domains in which they are trained, even subtle changes in the data acquisition or population characteristics may disrupt the algorithm. In contrast, human beings appear to be able to adapt to such changes quickly and automatically.
One, brute-force, approach to this problem is to create ever larger datasets, which can encompass all possible data conditions. This can be successful in some cases, but it is resource intensive, and always vulnerable to the appearance of new data types or knowledge classes in the future. Here, we build upon prior work showing that transfer learning improved performance on EEGs from different hospital centers and different patient populations (Saab 2020), as we explore our ability to adapt knowledge learned from EEGs coming from different centers and older ages groups to neonatal patients.
Method:
First, we trained a graph neural network (Li et al. 2018) (Figure 1A) for seizure detection on 26,217 adult EEGs collected at Stanford Hospital. Next, we applied transfer learning to 3,016 neonatal EEGs collected at Lucile Packard Children's Hospital (LPCH). Transfer learning was performed by initializing the model parameters with the weights pre-trained on the adult EEGs. To further improve knowledge generalization, we developed a meta-learning framework (Figure 1B) by incorporating different ages and downsampling scales into the model architecture. We carried out the following experiments to explore the effectiveness of knowledge transfer from adult EEGs with the meta-learning framework: (a) initialize the model parameters randomly; (b) initialize the model parameters with pre-trained weights on the adult EEGs and fine-tune all the model parameters; (c) initialize the model parameters with pre-trained weights and freeze them. For both models, the inputs are 60s epochs of EEGs, and the outputs are binary labels indicating seizure or no seizure.
Results:
Transfer learning from adult EEGs greatly improved seizure detection ROC-AUC on neonatal EEGs, with the largest gain from 0.58 to 0.74 ROC-AUC when the input signals were downsampled by a scale of 2. Moreover, by using the original 60s signals and signals downsampled by scales of 2 and 4, our meta-learning approach generally improved model performance compared to that without meta-learning, and achieved 0.85 ROC-AUC with knowledge transferred from adult EEGs.
Conclusion:
We show that transfer learning and meta-learning approaches can be used to generalize deep learning algorithms for EEG analysis across centers and age populations. Importantly, our results suggest that knowledge learned on adult EEGs can be effectively transferred to neonatal EEGs and inform neonatal EEG interpretation.
Funding:
:This work was supported in part by the Stanford Wu Tsai Neuroscience Foundation
FIGURES
Figure 1