PII-223 - BRAIN MEASUREMENTS, PERIPHERAL BIOMARKERS, OR BOTH? IMPROVED PREDICTIONS OF RESPONSE TO SERTRALINE THROUGH MULTI-MODAL DATA INTEGRATION.

C. Grant¹, M. Barac², T. Mayes³, T. Carmody³, A. Minhajuddin³, M. Jha³, P. Croarkin², W. Bobo², R. Toll³, C. Chin Fatt³, A. Athreya², M. Trivedi³; ¹Mayo Clinic, Rochester, MN, USA, ²Mayo Clinic, ³UT Southwestern.

Background: Prior work demonstrates that outcomes to antidepressant pharmacotherapy can be predicted using either brain or blood-derived markers. However, integration of brain- and blood-derived markers to improve predictability of outcomes has not yet been achieved. This work develops a machine learning (ML) pipeline to integrate brain- and blood-derived measures and improve predictability of treatment outcomes using antidepressant pharmacotherapy (Sertraline) in major depressive disorder (MDD) as a case study.

Methods: Patients with MDD (N=146) treated with Sertraline from the ‘Establishing Moderators and Biosignatures of Antidepressant Response for Clinical Care for Depression’ (EMBARC) study were included (response rate = 52%, defined as score <= 2 on the Clinical Global Impression Scale [CGI-I]). Resting state fMRI, EEG, plasma protein, and clinical items were collected (8,949 items). Feature selection was performed via an integrative network analysis-based approach. Using the selected features, supervised ML approaches (random forest, XGBoost, penalized regression, naïve bayes, K-nearest neighbors) were implemented to compare predictability of response to Sertraline by each data modality alone (clinical, protein, fMRI, EEG) and all possible combinations of modalities (15, in total). Nested cross validation was used to train and test models, with 75% of the sample randomly allocated for training. The random split was repeated 10 times, and results were averaged over splits.

Results: Network-based feature selection yielded 303 features (of the 8,949) for inclusion in ML prediction of response to Sertraline. The highest area under the receiver operating characteristic curve (AUC) was achieved by random-forest based integration of clinical, blood, EEG, and fMRI brain connectivity measures (mean AUC [95% CI] = 70 [0.62-0.79]). Sensitivity was 0.79 [0.71-0.87], and specificity was 0.53 [0.39-0.67]. Top predictors include fMRI, EEG, and blood markers.

Conclusion: This work demonstrates the power of multi-modal measures integrated through a ML pipeline to improve predictability of response to Sertraline. Future work will strive to understand the functional interplay of top brain (fMRI, EEG) and blood-derived (protein) predictors.