Information theory characteristics improve the prediction of lithium response in bipolar disorder patients using a support vector machine classifier

• Published in: Bipolar disorders Vol. 25; no. 2; pp. 110 - 127
• Main Authors: Tripathi, Utkarsh, Mizrahi, Liron, Alda, Martin, Falkovich, Gregory, Stern, Shani
• Format: Journal Article
• Language: English
• Published: Denmark Wiley Subscription Services, Inc 01.03.2023
• Subjects: Bipolar disorder; Bipolar Disorder - diagnosis; bipolar disorders; Dentate gyrus; entropy; Granule cells; Humans; Information Theory; iPSC; Lithium; Lithium - therapeutic use; Lithium Compounds - therapeutic use; lithium responders and non‐responders; lithium response prediction; machine learning; Mood; Morbidity; mutual information; Patients; Support Vector Machine; Support vector machines; SVM classifier; valproic acid; lithium responders and non-responders; entropy; iPSC; lithium response prediction; bipolar disorders; information theory; mutual information; valproic acid; machine learning; SVM classifier
• DOI: 10.1111/bdi.13282
• ISSN: 1398-5647; 1399-5618

Aim Bipolar disorder (BD) is a mood disorder with a high morbidity and death rate. Lithium (Li), a prominent mood stabilizer, is often used as a first‐line treatment. However, clinical studies have shown that Li is fully effective in roughly 30% of BD patients. Our goal in this study was to use features derived from information theory to improve the prediction of the patient's response to Li as well as develop a diagnostic algorithm for the disorder. Methods We have performed electrophysiological recordings in patient‐derived dentate gyrus (DG) granule neurons (from a total of 9 subjects) for three groups: 3 control individuals, 3 BD patients who respond to Li treatment (LR), and 3 BD patients who do not respond to Li treatment (NR). The recordings were analyzed by the statistical tools of modern information theory. We used a Support Vector Machine (SVM) and Random forest (RF) classifiers with the basic electrophysiological features with additional information theory features. Results Information theory features provided further knowledge about the distribution of the electrophysiological entities and the interactions between the different features, which improved classification schemes. These newly added features significantly improved our ability to distinguish the BD patients from the control individuals (an improvement from 60% to 74% accuracy) and LR from NR patients (an improvement from 81% to 99% accuracy). Conclusion The addition of Information theory‐derived features provides further knowledge about the distribution of the parameters and their interactions, thus significantly improving the ability to discriminate and predict the LRs from the NRs and the patients from the controls.