Deep Representation Learning for Optimizing Individualized Treatment Decisions

Music City Center 

Antidepressant treatment response remains highly unpredictable, with moderate
remission rates due to heterogeneous disease mechanisms. Electroencephalography
(EEG) biomarkers has shown utilities for guiding optimal treatment rule but poses
challenges for direct use due to the need for manual feature extraction and lack of
interpretability for its complex high-dimensional property. In this paper, we propose a
deep representation learning framework EEG-Variational Autoencoder (EEGVAE) ar-
chitecture that incorporates a convolution neural network based encoder base designed
for EEG (i.e., EEGNet), and extends to a multi-head EEGVAE variant for estimating
heterogeneous treatment effects (HTE) to improve treatment response prediction. Our
approach offers several advantages by providing 1) an end-to-end learning framework
to eliminate the need for manual feature engineering; 2) flexibility in incorporating
multi-modality data; 3) the ability to learn robust, interpretable EEG representation.
Through simulations based on real data and application to a large randomized clin-
ical trial EMBARC, we demonstrate our method's superior performance in predict-
ing responder status and optimal treatment policy estimation compared to existing
approaches. The extracted latent encoding provides insights into spatial and tempo-
ral EEG patterns that influence treatment outcomes, advancing our understanding of
treatment response mechanisms in depression.