Multimodal Machine Learning Can Predict Persistent Depression in Adults with Epilepsy: A Pilot Study
Abstract number :
1.233
Submission category :
4. Clinical Epilepsy / 4D. Prognosis
Year :
2022
Submission ID :
2203925
Source :
www.aesnet.org
Presentation date :
12/3/2022 12:00:00 PM
Published date :
Nov 22, 2022, 05:22 AM
Authors :
Guillermo Delgado-Garcia, MD, MSc, CSCN (EEG) – University of Calgary; Jordan D.T. Engbers, PhD – Co-first author, Desid Labs; Pauline Mouches, PhD(c) – University of Calgary; Kimberly Amador, BSc, MSc – University of Calgary; Samuel Wiebe, MD, MSc, – University of Calgary; Nils D. Forkert, MSc, PhD – University of Calgary; James A. White, MD – University of Calgary; Tolulope Sajobi, PhD – University of Calgary; Karl Martin Klein, MD, PhD – University of Calgary; Colin B. Josephson, MD, MSc – University of Calgary; Calgary Comprehensive Epilepsy Program Collaborators, MD – Group author, University of Calgary
Rationale: Depression is common in people with epilepsy and affects seizure outcomes. Accurate prediction of its development and course may mitigate adverse outcomes through early treatment and prevention. The aim of this work was to develop a multimodal machine learning approach for predicting depression in adults with epilepsy.
Methods: We randomly selected 200 patients from the Calgary Comprehensive Epilepsy Program registry and linked their baseline clinical data to their first available EEG and MRI acquired during routine clinical care. Two analyses were performed: (1) predicting a combination of incident and prevalent depression that was persistent across all follow-up visits and (2) predicting incident depression following the baseline clinic visit in those free of it at the first visit. We applied the ReliefF feature selection algorithm on clinical, EEG (empirical Shannon entropy, excessive entropy rate, and entropy of the surrogate Markov microstate model), and MRI (volume, area, and median ADC values from 27 brain regions) features. Seven machine learning algorithms (Table 1) were trained and tested using a stratified k-fold cross-validation. F1 score (F1), Matthew’s correlation coefficient (MCC), Area Under the Receiver Operating Characteristic Curve (ROC AUC), sensitivity, and specificity were reported.
Results: Of 200 patients, 150 (75%) had EEG and MRI data of sufficient quality for adequate feature extraction. Median age was 36 (interquartile range [IQR] 23) years, 50% were female, 14% were 1-year seizure-free, and 21% had depression at baseline. Median total visits were 4 (IQR 3) and median follow-up 2.6 (IQR 1.7) years. Of those free of it at baseline, 27 (18%) developed incident depression during follow-up. Using the ReliefF algorithm, a total of 46 and 32 features were selected for each of the planned analyses (Table 2). No quantitative MRI or EEG variables were selected for the final models following inclusion in the ReliefF algorithm. Random Forest, L2 Penalized Logistic Regression, Gradient Boosting Consensus, and Support Vector Classification performed well (F1: 0.70-0.75; MCC: 0.39-0.48; Table 1) for predicting persistent depression. Model performance was less reliable when predicting incident depression (F1: 0.71-0.75; MCC: 0.05-0.09; Table 1).
Conclusions: Multimodal machine learning can predict the course of depression in epilepsy. The pilot model to predict persistent depression demonstrated good discrimination and calibration, but future efforts will need to be made before these algorithms are ready for clinical practice. Other quantitative MRI and EEG features, as well as the application of deep learning to larger samples, could add value to fully exploit these modalities for depression prediction in people with epilepsy.
Funding: This work was supported by Epilepsy Canada.
Clinical Epilepsy