Tri-model ensemble with Grid Search optimization for Bipolar Disorder Diagnosis

Abstract

Bipolar disorder is one of the common mood disorders and diagnosis is the most important part of mood disorders. This research involves sequence classification of bipolar 1, bipolar 2, and cyclothymia using psychiatric cliff notes, there were 200 samples of bipolar 1, 200 samples of bipolar 2, and 200 samples of cyclothymia. This work uses a novel tri-model based ensemble for the diagnosis of bipolar disorder with grid search based optimization. The paper involved several textual preprocessing techniques like lower casing, punctuation removal, and lemmatization and it involved the tfidf approach for feature extraction of important attention words from paragraph based textual data. After preprocessing the ensemble was created using three models Decision tree, Random Forest, and Adaboost. The ensemble was optimized with grid search optimization with an early stopping mechanism to prevent overfitting. The ensemble’s classification prediction was determined by the highest vote from the 3 individual models. The tri-model ensemble produced excellent results with an accuracy of 99% and precision, recall, and f1-score of 98%, outperforming other studies on text based mental health disorders. This work is the first work to include cyclothymia variants of bipolar disorder and involved complete coverage of all the 3 types of bipolar disorder. This work can help to facilitate bipolar patients and provide an extremely accurate diagnosis of all types of bipolar disorder in a real world scenario.