Using Semi‐supervised Variational Autoencoder and Acute fMRI Changes to Predict Antidepressant Treatment Responses Among Older Adults with Depression

• Published in: Alzheimer's & dementia Vol. 19; no. S17
• Main Authors: Wang, Linghai, Lee, Jihui, Mizuno, Akiko, Wilson, James D, Lamb, Shannon, Gerlach, Andrew, Wu, Minjie, Aizenstein, Howard J
• Format: Journal Article
• Language: English
• Published: 01.12.2023
• ISSN: 1552-5260; 1552-5279
• DOI: 10.1002/alz.079181

Background Late‐life depression (LLD) has been consistently linked to an increased risk of dementia, with studies showing that older adults with depression have a two fold risk of developing dementia. Dementia has proven difficult to treat, yet depression has several viable treatment options, particularly antidepressants. However, the efficacy of antidepressants in LLD has proven modest as antidepressants typically achieve around a 50% remission rate and only after 4‐6 weeks of treatment. In this analysis, we investigate the effectiveness of combining fMRI data with deep learning for predicting antidepressant driven depression remission. Method In two LLD studies using venlafaxine (n = 51) and escitalopram or levomilnacipran (n = 29), resting state fMRI scans were collected at baseline and at day one of antidepressants. Region‐to‐region functional connectivity (FC) was calculated for each scan using the Shen50 atlas, and differential connectivity (DC) was calculated to compare baseline and day 1 FC. Clinical factors including age, race, education and cumulative illness burden and DC were used as input to a semi‐supervised variational autoencoder (VAE) to predict remission status at treatment end (Fig.1). Remission was encoded as a binary variable, defined as a final MADRS score of 10 or less after 12 weeks of treatment and subject to blinded clinician assessment. Monte Carlo cross‐validation over 30 repetitions was used to evaluate the model performance with 10% of the data held out for testing. The VAE was compared to a random forest classifier with the same cross validation methods using the area under the curve (AUC). Result The VAE achieved an average mean‐squared error of .005 and AUC of 0.87 on a held out test set during cross validation. The VAE outperformed the random forest classifier which achieved an average AUC of 0.68. Conclusion The VAE using DC in fMRI scans between baseline and day one shows the potential in accurately identifying responders to antidepressants as early as in one day. Our results show that acute changes to functional networks triggered by a single dose of antidepressants are promising as a means to improve patient outcomes. These results can contribute to optimizing treatment for LLD in the future.