A multi-label learning model for psychotic diseases in Nigeria

Abstract

Abstract The goal of Multi-Label Classification (MLC) is to allot an instance to a set of different labels. This task is usually addressed by either transforming the problem into several binary problems, adapting machine learning models to fit multi-label data or create an ensemble of models that can classify multi-label datasets. The communal relationship between Bipolar, Insomnia, Schizophrenia, Vascular Dementia (VD) and Attention-Deficit/Hyperactivity Disorder (ADHD) in the Psychotic Disorder Diseases (PDD) motivate the research for a diagnostic method that classifies and evaluates each psychotic disorder simultaneously. This study experimentally evaluates 15 MLC methods using 10 evaluation measures over a new PDD dataset. The performance of these methods is measured with four ranking - based, three example-based and three label-based measures. Also, the efficiency of these methods is measured by their 90-10 Train-Test split with the 10 evaluation measures. The results show that the Label Powerset (LC) and Pruned Sets (PS), MLC methods with Naive Bayes (NB) and Naive Bayes Tree (NBTree) consistently performed best in terms of the evaluation measures on the PDD dataset. Schizophrenia has the highest classification accuracy with Bipolar the lowest in the data split of 90-10. Logistic model tree (LMT) is the best algorithm for Insomnia and Bipolar while Naive Bayes (NB) is the best for Schizophrenia, VD and MBD. Support vector machines (SVM) with ensemble learning and classification (ELC) and Ensemble of Pruned Set (EPS) are the best classifiers for Bipolar while SVM with regression and threshold (RT) is the least. The classifiers are statistically significantly different for Insomnia, VD and ADHD only.