Dealing with mixed data types in the obsessive-compulsive disorder using ensemble classification

• Published in: Neurology, psychiatry, and brain research Vol. 32; pp. 77 - 84
• Main Authors: Hasanpour, Hesam, Ghavamizadeh Meibodi, Ramak, Navi, Keivan, Asadi, Sareh
• Format: Journal Article
• Language: English
• Published: Elsevier GmbH 01.06.2019
• Subjects: Ensemble classification; Fluvoxamine; Mixed data types; Predictor; Treatment response; Mixed data types; Treatment response; Ensemble classification; Predictor; Fluvoxamine
• ISSN: 0941-9500
• DOI: 10.1016/j.npbr.2019.04.004

•New ensemble method for classification of obsessive-compulsive disorder pharmacotherapy data set has been proposed.•The proposed strategy introduced classification method to deal with mixed data types.•The proposed algorithm predicts treatment response in OCD with good accuracy, sensitivity and specificity. Obsessive-compulsive disorder (OCD) is a psychiatric disorder characterized by recurrent obsessions and/or compulsions. Applying classification algorithms for prediction of treatment response helps to individualize treatment with more effectiveness. OCD data set is heterogeneous including continuous and discrete variables which presents challenges for most of the traditional classifiers to avoid data over-fitting. Here, we aimed to develop an ensemble classifier which is suitable for mixed data types for prediction of treatment response in OCD. One hundred fifty-one subjects with OCD aged between 18–65 underwent fluvoxamine pharmacotherapy for 12 weeks and categorized into two groups (responder, non-responder) based on the reduction in their symptom severity following treatment. Decision tree and support vector machines (SVM-tree) were combined to deal with discrete and continuous variables and were used as base classifiers to build an ensemble of classifiers. Some of the attributes such as sexual obsessions and occupation, factor 1 (aggressive, contamination, sexual, religious, symmetry obsessions), initial obsession score, age at onset and illness duration are the high ranked predictors of treatment response. Comparing accuracy, precision, sensitivity, specificity and f-measure of the new algorithm with traditional classification algorithms such as decision tree, support vector machines (SVM), k-nearest neighbor and random forest showed a stronger performance of the proposed algorithm in the prediction of OCD treatment response. The proposed strategy introduced an effective classification method to deal with medical datasets with mixed data types which can be of great significance in medical datasets and personalized medicine.