Association Between Alu Insertion/Deletion Polymorphism On The Tpa Gene And Mirtazapine Response In Koreans With Major Depression

Objectives: Although single clinical predictors have repeatedly been associated with TRD (treatment resistant depression), they have not proven sufficient for predicting treatment outcome1. Thus, attention shifted to interaction-based models but only few multivariate investigations have been performed in TRD so far2. Using the data pool of the Group for the Study of TreatmentResistant-Depression (GSRD) and a machine learning algorithm we intended to draw new insights and back up previous results featuring a set of 66 clinical and demographical predictors for treatment outcome. Methods: 415 patients recruited between 2011 and 2015 in 11 participating centers showed full availability for all 66 predictors. Treatment response was defined by MADRS-score below 22 and a reduction of 50% or more. A score higher than 21 after at least two antidepressant trials of adequat dosage and length was considered as treatment resistance. After generating importance values for all predictors the prediction algorithm was trained in a sample of 385 patients. Subsequently, prediction was performed in a sample of 30 new patients not featured in the model generation. Results: The accuracy for predicting treatment outcome in TRD was at 0.75 using all 66 predictors. Importance measurement revealed chronicity, i.e. full or partial intraepisodic recovery or chronic MDD, number of depressive episodes, age of first and last lifetime depressive episode, total time of hospitalization, education and occupation status, suicidal risk, marital status and number of children and cigarettes smoked per day as the most useful predictors. Conclusion: Exploiting a machine-learning algorithm, we scored an accuracy of 0.75 for treatment outcome using a sample of 415 patients. Reaching a probability of 83.4% for a correct prediction for treatment resistance and 66.6% for response we exceeded the predictive capabilities of clinicians. Thus, these results strengthen our previous data mining approaches and suggest keeping the focus on interaction-based statistical approaches3.