Modeling and Prediction of Clinical Symptom Trajectories of MCI Subjects Using Longitudinal Data
Bhagwat, Nikhil1,2,3; Raihaan Patel3; Aristotle N. Voineskos2,5; M. Mallar Chakravarty3,4
1. Institute of Biomaterials & Biomedical Engineering, University of Toronto; 2. Kimel Family Translational Imaging-Genetics Research Lab, Research Imaging Centre, Campbell Family Mental Health Research Institute, Centre for Addiction and Mental Health; 3. Cerebral Imaging Centre, Douglas Mental Health University Institute, Verdun, QC, Canada; 4. Department of Psychiatry, McGill University; 5. Department of Psychiatry, University of Toronto
Introduction: Prognostic predictions of subjects from prodromal stages such as mild-cognitive-impairment (MCI) is an area of great clinical interest. Prognostic forecasting of symptom severity is complicated not only by the heterogeneity in demographics and clinical presentation, but also highly variable and nonlinear symptom patterns exhibited in those suffering from MCI. We address these challenges through a novel data-driven modeling of symptom trajectories derived solely from a longitudinal clinical scale measuring cognitive performance. We show that the resultant trajectories classes represent relatively stable and progressive subgroups of MCI subjects. Furthermore, we show that clinical and structural MR imaging data can be combined to predict these trajectories at baseline using machine-learning techniques. We evaluate our clustering and trajectory prediction performance using MCI cohort from publically available ADNI-2 dataset.
Methods: 156 MCI subjects with complete 2-year longitudinal data were used in this analysis. ADAS-13 scale was used as a surrogate cognitive performance measure. MR data for all subjects were processed through MAGeT Brain for hippocampal segmentation (9 subfields) and CIVET for cortical surface extraction for cortical thickness measures (78 AAL regions). ADAS-13 scores at baseline, 6,12, and 24 months were used as a clinical feature vector of a subject. The correlation between two feature vectors was used as a similarity metric among subjects. Then, subjects were divided into two subgroups representing different symptom trajectories using hierarchical clustering. Subsequently, age, sex, APOE4 status, baseline clinical score and MR features were used to predict these trajectory classes for a given individual. Logistic regression with Lasso, support vector machine, and random forest classifiers were compared for predictive performance in a 10-fold nested cross-validation paradigm.
Results: Hierarchical clustering yielded two distinct subgroups of subjects (N=75,81) following two trajectory classes with characteristic progression patterns of ADAS-13 scores. For trajectory prediction task, accuracy, confusion matrix, and area under the curve were used as performance metrics. First, we used age, sex, APOE4 status, and clinical score as input which yielded best scores of 0.58 (ACC) and 0.59 (AUC). Subsequently, we added MR variables to the input which resulted in improved scores of 0.64 (ACC) and 0.69 (AUC).
Discussion: In this work, we showed a novel data-driven approach for defining MCI subgroups using longitudinal clinical scores. The clusters represent distinct trajectories classes pertaining to stable and worsening cognitive symptoms. We then predicted these trajectories using clinical variables (age, sex, APOE4 status, baseline score). Moreover, we showed that this likelihood of trajectory class can be improved significantly by incorporating hippocampal and cortical thickness measures as input features.