Testing Bayesian Classifiers on Adult Depression Data: A Study to Handle Uncertainty Related to Its Grading

作者：Chattopadhyay Subhagata^*; Rajput S S; Prajesh A R

来源：Journal of Medical Imaging and Health Informatics, 2013, 3(4): 607-614.

DOI：10.1166/jmihi.2013.1211

• Depression
• Baye's Net
• Significant Symptoms
• Bayesian Classifier
• Depression Grading
• Hill Climbing Search Algorithm

Accurate grading of depression is a research challenge. It is important to decide on the treatment plans. This paper proposes a depression classification model using Bayesian approach. In order to achieve this task, 302 real-world depression cases have been collected and used to develop the initial model where the number of symptoms equals to 15 for each case. The load of symptoms is quantified [0, 1] by specialist doctors (psychiatrists) to measure the cumulative grade of the illness, which is labeled as 'Mild/Low,' 'Moderate/Medium,' and 'Severe/High.' In order to reduce the data dimension and extracting significant factors, Paired t-test (PTT), Principal Component Analysis (PCA), Multiple Linear Regressions (MLR), Analysis of Variance (ANOVA) and Factor Analysis (FA) is performed. All these techniques are tried because we have no prior idea which would effectively serve the purpose. After rigorous experiments, it is noted that 4 symptoms are found significant (p value < 0.05) using PTT. It reduces the data matrix from 302 x 15 to 302 x 4. It helps reducing the complexity in the intended classifier design. At first a Naive Baye's classifier is designed. It can only grade depression with 19% accuracy. Hence, it is further trained using Hill Climbing Search (HCS) algorithm to develop a Learned Bayesian Classifier. Throughout the design and developmental processes, 70% of the data has been used for training and the rest tests the classifier's performance. Experimental results show that the average accuracy achieved by the learned classifier is close to 80%.

• 出版日期2013-12