Using a SSVAE to Predict Antidepressant Treatment Response Among Older Adults with Depression

1708 

Abstract Submission 

Linghai Wang¹, Jihui Diaz¹, Akiko Mizuno¹, James Wilson², Andrew Gerlach¹, Carmen Andreescu¹, Minjie Wu¹, Howard Aizenstein¹

¹University of Pittsburgh, Pittsburgh, PA, ²University of San Francisco, San Francisco, CA

Linghai Wang  
University of Pittsburgh
Pittsburgh, PA

Jihui Diaz  
University of Pittsburgh
Pittsburgh, PA

Akiko Mizuno, PhD  
University of Pittsburgh
Pittsburgh, PA

James Wilson  
University of San Francisco
San Francisco, CA

Andrew Gerlach, PhD  
University of Pittsburgh
Pittsburgh, PA

Carmen Andreescu, MD  
University of Pittsburgh
Pittsburgh, PA

Minjie Wu, PhD  
University of Pittsburgh
Pittsburgh, PA

Howard Aizenstein, M.D., PhD.  
University of Pittsburgh
Pittsburgh, PA

Late-life depression (LLD) represents a significant concern in geriatric mental health due to its association with heightened risks of suicide, cardiovascular disease, and cognitive impairment including dementia [1]. The effectiveness of antidepressants for LLD is modest, typically achieving only a 50% remission rate for the first trial [2]. Resting-state functional magnetic resonance imaging (fMRI) is frequently used to understand responses to antidepressant treatment in LLD [3]. However, the analysis of such data often relies on aggregate measures due to challenges in interpreting raw data, which are characterized by noise, nonlinearities, and high dimensionality. Recent advancements in deep learning have demonstrated considerable promise for data-driven analysis of complex datasets, harnessing more nuanced information. In this study, we tested the performance of an analytic strategy integrating baseline resting-state fMRI functional connectivity data using a semi-supervised variational autoencoder (SSVAE) to predict depression remission of driven by standard antidepressant medication.

In this study, remission was encoded as a binary variable, defined by a final MADRS score of 10 or less for two consecutive weeks following 12 weeks of treatment, with assessments by a clinician blinded to the treatment conditions. A total of 80 resting-state fMRI scans were analyzed (mean±SD age 66.2±6.8 years; baseline MADRS 24.2.2±6.8). The image acquisition and processing details were outlined in the study by Wilson et al. [4]. Region-to-region functional connectivity (FC) was calculated for each scan using the Shen50 atlas. For the main model, demographic, and clinical factors, such as age, race, education, cumulative illness burden, baseline MADRS, and FC were used as inputs for an SSVAE [5], which was modified to predict remission status at the end of treatment (Fig. 1). The SSVAE combines an autoencoding network with a prediction network to efficiently learn information relevant for prediction while minimizing overfitting. The SSVAE was trained for an average of 22 epochs with early stopping at three epochs and a learning rate of 0.00001. The hyperparameters were selected using a held-out validation set comprising 10% of the data. Monte Carlo cross-validation over 30 repetitions was used to evaluate the model performance, with 20% of the data reserved for testing. The SSVAE's results were compared to a random forest classifier using the same cross-validation methods and evaluating the area under the curve (AUC). To evaluate the predictive significance of FC, we trained both SSVAE and random forest models with an identical set of input variables, excluding FC, as a secondary model for each.

These results are promising, indicating that a pretreatment scan can predict the treatment response. From our testing, the SSVAE outperformed the random forest model when including pre-treatment FC but performed worse when it was excluded. This indicates the ability of the SSVAE to effectively handle a larger number of variables in the prediction. This is consistent with the ability of neural networks to scale effectively for complex inputs. However, further work is needed to refine the interpretation of these predictions.

These results are promising, indicating that a pretreatment scan can predict the effectiveness of first-line antidepressant treatments. From our testing, the SSVAE outperformed the random forest model when including pre-treatment FC but performed worse when it was excluded. This indicates the ability of the SSVAE to effectively handle a larger number of variables in the prediction. This is consistent with the ability of neural networks to scale effectively for complex inputs. However, further work is needed to refine the interpretation of these predictions.

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Aging

Classification and Predictive Modeling ²

fMRI Connectivity and Network Modeling ¹

Task-Independent and Resting-State Analysis

Aging

Design and Analysis

FUNCTIONAL MRI

Psychiatric Disorders

Other - Deep Learning; Depression; Resting-state fMRI; Predictive modeling; Semi-supervised learning