Effective Classification of Major Depressive Disorder Patients Using Machine Learning Techniques

Abstract

Background: Major Depressive Disorder (MDD) in simple terms is a psychiatric disorder which may be indicated by having mood disturbances which are consistent for more than a few weeks. It is considered a serious threat to psychophysiology which when left undiagnosed may even lead to the death of the victim so it is more important to have an effective predictive model. The major Depressive disorder is often termed as comorbid medical condition (medical condition that co-occurs with another), it is hardly possible for the physicians to predict that the victim is under depression, timely diagnosis of MDD may help in avoiding other comorbidities. Machine learning is a branch of artificial intelligence which makes the system capable of learning from the past and with that experience improves the future results even without programming explicitly. As in recent days because of the high dimensionality of features, the accuracy of the predictions is comparatively low. In order to get rid of redundant and unrelated features from the data and improve the accuracy, relevant features must be selected using effective feature selection methods. Objective: This study aims to develop a predictive model for diagnosing the Major Depressive Disorder among the IT professionals by reducing the feature dimension using feature selection techniques and evaluate them by implementing three machine learning classifiers such as Naïve Bayes, Support Vector Machines and Decision Tree. </P><P> Method: We have used Random Forest based Recursive Feature Elimination technique to reduce the feature dimensions. Results: The results show a considerable increase in prediction accuracy after applying feature selection technique. Conclusion: From the results, it is implied that the classification algorithms perform better after reducing the feature dimensions.

Realizing a Stacking Generalization Model to Improve the Prediction Accuracy of Major Depressive Disorder in Adults

“…The filter method ranks every feature, based on their relevance score. Filter methods are faster than other feature selection methods, but they are not computationally effective as they lack interaction with the learning model [20].…”

“…In embedded methods, the training of a machine learning model itself will have a feature selection step as a part of it (decision tree, LASSO) [14]. Although it is more effective than filter and wrapper methods, it is highly dependent on the learning model, thus it is difficult to use with any other learning model [20].…”

“…Feature selection is the process of selecting the best subset of features that have a significant impact on the predicted outcomes. For prediction/classification to be efficient and reduce the computation time, feature selection is the most needed one [20]. They are a form of the black-box method [14], where we feed them with features and do not know the internal structure or functionality.…”

“…It is done using random searching, in which cross-validation is done to predict the advantage of adding a feature or removing it from the subset. The wrapper method wraps the implementation of the chosen model, along with the search technique [20]. They sometimes lead to over-fitting and time-consuming [19].…”

“…Here the model is introduced with data from the real world and made to gain knowledge on their own and find hidden meaning from the data [15]. Widely used unsupervised algorithms are Clustering and Association rule mining [20].…”