Sequential autoencoders for feature engineering and pretraining in major depressive disorder risk prediction

Area of Science:

• Computational psychiatry
• Machine learning in healthcare
• Deep learning for risk prediction

Background:

• Major Depressive Disorder (MDD) poses a significant health burden, necessitating improved prognostic risk prediction.
• Traditional risk models often rely on aggregate features, potentially missing complex temporal relationships in patient data.
• Autoencoders offer a novel approach for feature engineering and pretraining to capture these temporal dynamics.

Purpose of the Study:

• To evaluate autoencoders as a feature engineering and pretraining technique for enhancing MDD prognostic risk prediction.
• To compare the performance of autoencoder-based strategies against established benchmarks like Restricted Boltzmann Machines (RBM) and random forests.
• To assess the temporal predictive performance of autoencoders with sequential structures, including Attention and LSTM layers.

Main Methods:

• MDD patient data from Vanderbilt University Medical Center was utilized.
• Autoencoder models incorporating Attention and LSTM layers were trained to generate latent data representations.
• Predictive performance was assessed temporally using random forest models and by initializing neural network layers with autoencoder weights.
• Area Under the Precision-Recall Curve (AUPRC) trends were evaluated throughout the patient treatment course.

Main Results:

• The pretrained LSTM model demonstrated superior predictive performance over pretrained Attention models and benchmarks for 3 out of 4 outcomes, including self-harm/suicide attempt (AUPRCs: LSTM pretrained=0.012, Attention pretrained=0.010, RBM=0.009, random forest=0.005).
• Autoencoders used for feature engineering yielded varied results, with benchmarks outperforming LSTM and Attention encodings for the self-harm/suicide attempt outcome.
• While pretraining improved prediction, temporal feature encodings did not show additive benefits, suggesting potential information loss during the encoding process.

Conclusions:

• Pretrained LSTM models using autoencoder weights show clinically useful performance, outperforming current state-of-the-art predictors for MDD.
• The findings suggest that pretrained LSTM autoencoder weights are a valuable tool for future MDD risk modeling research.
• Further investigation is warranted to understand the mechanisms behind predictive information retention and potential loss during autoencoder encoding.