Background: Major Depression is a serious and prevalent mental illness. While many treatments exist, they are not equally effective for all patients. Clinicians lack an evidence-based method to predict which treatment is best for an individual patient, leading to significant delays before some patients improve. Our goal is to take advantage of deep learning and large datasets to create a clinical decision aid to help physicians select treatments based on predictions of individual patient responses.
Methods: We will use data from research studies and clinical trials to ensure high data quality. Data will include patient features such as treatment history, sociodemographic information, symptom profile, imaging results, peripheral markers, and genetics. Input parameters will be selected based on literature, expert opinion, and contribution to model performance. Data will be combined into a final dataset with the curated features. Use of deep learning will help ensure robustness to missing inputs. Training on heterogeneous data, with subtle dependencies between a large number of features, requires deep layers to represent complex dependencies in the data. We have built an open-source deep learning framework, Vulcan (https://github.com/rfratila/Vulcan), focused on modularity, optimization, and interpretability. Our network architecture uses a feed-forward Deep Belief Networks (DBN) with self-normalizing Scaled Exponential Linear Units (SELU) to achieve higher-order representation. We take advantage of function space non-convexity by using Snapshot learning to learn separate nuanced features in the data. Final output will be a list of the most effective predicted treatments. Patient outcomes will be assessed in terms of validated depression rating scales. Model validation methods will include sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and the receiver operating curve (ROC) along with the area under the ROC (AUC). K-fold cross validation will be used to ensure model validity. Dropout will be used during training to prevent overfitting. Clinicians require interpretability in decision-making tools, something that has been traditionally difficult with deep learning. Our tools for addressing this include: cross-validating the final model input features and relating them to existing literature, saliency maps, and using receptive field analysis for all layers in the network to gain a sense of low-level feature groupings. This level of sophistication is roughly analogous to the factor analyses available in the medical literature. Some of the feature clusters we identify may also produce novel research questions. Such clusters correspond to similarities between certain patient types that we would be able to extract with t-SNE.
Our model will be the first to combine patient symptom and history information with biomarkers into a system designed to be used by physicians for the treatment of depression. We focus on building a clinician-intelligible model by curating excellent data with a clinically accessible feature space to produce an optimal clinical decision aid.
PRECISION MEDICINE & DRUG DISCOVERY
Author: Sonia Israel
Coauthor(s): David Benrimoh, Robert Fratila, Kelly Perlman, Eleonore Fournier-Tombs
Status: Work In Progress