Abstract
Electroencephalogram (EEG) provides noninvasive measures of brain activity and is found to be valuable for diagnosis of some chronic disorders. Specifically, pre-treatment EEG signals in alpha and theta frequency bands have demonstrated some association with anti-depressant response, which is well-known to have low response rate. We aim to design an integrated pipeline that improves the response rate of major depressive disorder patients by developing an individualized treatment policy guided by the resting state pre-treatment EEG recordings and other treatment effects modifiers. We first design an innovative automatic site-specific EEG preprocessing pipeline to extract features that possess stronger signals compared with raw data. We then estimate the conditional average treatment effect using causal forests, and use a doubly robust technique to improve the efficiency in the estimation of the average treatment effect. We present evidence of heterogeneity in the treatment effect and the modifying power of EEG features as well as a significant average treatment effect, a result that cannot be obtained by conventional methods. Finally, we employ an efficient policy learning algorithm to learn an optimal depth-2 treatment assignment decision tree and compare its performance with Q-Learning and outcome-weighted learning via simulation studies and an application to a large multi-site, double-blind randomized controlled clinical trial, EMBARC.
Introduction
Treating Major Depressive Disorder (MDD) is often a frustrating process of trial and error. While antidepressants like sertraline are common, patient response rates are notoriously low and unpredictable. Clinicians lack reliable tools to determine which patient will benefit from which treatment.
While biomarkers like Electroencephalograms (EEG) offer a non-invasive window into brain activity, using them in clinical practice is difficult. EEG data is inherently "noisy," and recordings vary significantly between different hospital sites due to equipment differences. Furthermore, standard statistical methods often fail to capture the complex, non-linear interactions between a patient’s unique biology and their response to a drug—a phenomenon known as Heterogeneous Treatment Effect (HTE).
Methodology
The proposed pipeline consists of three main stages designed to transition from raw neuroimaging data to actionable clinical decisions.
Automated Site-Specific Preprocessing
To address the "garbage in, garbage out" problem, an automated pipeline was developed to detect and repair faulty sensors (channels) and noisy time segments (epochs). The multitaper method was then used to extract specific power signatures from alpha and theta frequency bands. As shown in Figure 1, this preprocessing significantly reduces noise and stabilizes the signal across various EEG channels, which is critical for accurate downstream analysis.
Causal Forests and Policy Learning
Instead of standard regression, the researchers employed Causal Forests (an extension of Random Forests) to uncover how treatment effects vary across a high-dimensional feature space. This non-parametric approach is particularly effective at capturing non-linear interactions. To turn these insights into action, a doubly robust technique was used to estimate the Average Treatment Effect (ATE), followed by a policy learning algorithm. This process culminates in a depth-2 decision tree—a simple, interpretable "if-this-then-that" tool suitable for clinical use.
Datasets
The pipeline was validated using the EMBARC (Establishing Moderators and Biosignatures of Antidepressant Response in Clinical care) dataset, a multi-site randomized controlled trial involving 309 participants. Additionally, synthetic simulations were conducted to test the algorithm's accuracy against known ground truths.
Results
The integrated pipeline demonstrated significant improvements in both signal quality and treatment assignment accuracy.
Simulation Performance
In synthetic tests, the proposed Policy Tree consistently outperformed traditional approaches such as Q-Learning and Outcome-Weighted (O) Learning. This superiority was observed across both "weak" and "large" effect size settings, as well as varying training sample sizes (N=200 vs. N=500).
Clinical Application to EMBARC
When applied to the EMBARC study, the model revealed a significant treatment effect that conventional methods missed. While standard empirical methods yielded a non-significant p-value ($p=0.08$) for sertraline, the doubly robust framework identified a significant improvement in response rate (17.4%, $p < 0.05$). The model identified fc2.close.theta (theta band power in the frontal/central region during eyes-closed states) as the most critical biomarker for predicting treatment response.
Conclusion
This research highlights that effective preprocessing is a causal necessity, significantly reducing error in treatment effect estimation. By prioritizing simplicity through a depth-2 decision tree, the authors provide a tool that is both high-performing and interpretable for clinicians. Future work will explore multi-modal fusion, combining EEG with fMRI and genetic data, and validation in external cohorts to further refine these personalized treatment policies.
Summary updated