SA40 - A TWO-DIMENSIONAL APPROACH FOR A VENLAFAXINE TREATMENT EFFICACY PREDICTION MODEL

Abstract

Background Patient response to antidepressant treatment is frequently defined at the end of treatment, as an improvement of 50% or greater from the initial depression score. However, this definition does not take into account the ‘time to response’, therefore not allowing for any differentiation between response rates. Of course, time to response can be an important factor when trying to determine features influencing the efficacy to antidepressant treatments as the genetic backgrounds of ‘fast’ and ‘slow’ responders may be different. By taking into consideration both depression scores and time to response, we can find new features that may serve as predictors. Additionally, many genetic differences may impact the efficacy to antidepressants, and analysis of combinations of genetic factors is needed to define novel predictors of response to antidepressant treatments. We applied novel methods, based on this logic, to design a prediction model of efficacy to Venlafaxine treatment. Methods Using the raw data of the Sequence Treatment Alternatives to Relieve Depression (STAR⁎D) clinical trial, we have designed two response models for the Venlafaxine treatment group: 1. A percentage of depression score reduction using the first and last scores. 2. An exponential formula, taking into account all depression scores for each patient, and plotted against their number of days in treatment. For genetic feature selection, we have applied a genome--wide association approach, and then developed a prediction model by applying machine learning algorithms to define combinations of parameters that predict response for Venlafaxine. Results Out of the 135 patients who were treated with Venlafaxine, the 50% depression score reduction approach found that 51 patients were responders and 84 patients were non-responders. The exponential approach however, labeled 66 patients as responders and 69 patients as non-responders. We split the Venlafaxine group into train and validation groups. The split was performed randomly to remove bias. Genome-wide association analysis identified two genetic features for the Venlafaxine model only for the exponential response model groups, but not for the 50% depression score reduction approach. After feature selection was performed, a Support Vector Machine (SVM) machine learning algorithm with a linear kernel was applied. The prediction model yielded the following results on the validation group: Area Under the Curve (AUC) of 0.8511, accuracy of 77.50%, sensitivity of 76.19%, specificity of 78.95%, positive predictive value of 80.00% and negative predictive value of 75.00%. Discussion Genome-wide association analysis on a two-dimensional response model taking into consideration both changes in depression score and time of response, revealed novel features which predict the response to Venlafaxine. These features could not be found using the commonly used one-dimensional 50% depression score reduction model, emphasizing the strengths of using a two-dimensional approach. Furthermore, by using machine learning algorithms we have managed to design a highly accurate prediction model of response to Venlafaxine, based on the combinatorial approach between the genetic features we identified by genome-wide association. Applying new models and new algorithmic methods on currently available large datasets can lead to novel findings which will advance our understanding of psychiatric disorders, and advance the design of accurate prediction models for psychiatric medications.