(C-94) A Machine Learning Approach to Classify Dementia-Related Disorders Using Magnetoencephalography (MEG) Signals

Introduction:: Dementia is a broadly defined clinical syndrome characterized by acquired losses of cognitive and emotional abilities severe enough to interfere with a person’s daily function and quality of life. The prodromal phase of dementia is known as mild cognitive impairment (MCI), a cognitive disorder characterized by memory impairment growing at a higher rate than expected by age. However, many individuals report noticing subtle changes in memory and cognitive function before MCI is clinically detectable. The subjective perception of cognitive decline compared to previous levels of functioning in individuals with normal cognition is a pre-clinical phase known as subjective memory complaints (SMC). Magnetoencephalography (MEG) is a non-invasive imaging technique that measures the magnetic fields, ranging from Femto-tesla to pico-tesla, induced by electric currents in the brain and delivers accurate temporal and spatial resolution mapping of neuronal activity. MEG has shown great potential in understanding the mechanisms of various dementia disorders. This study aimed at implementing continuous-wavelet-transform (CWT) on MEG signals to classify healthy control, MCI, and SMC subjects using machine learning (ML).


Materials and Methods:: Subjects were recruited as part of an ongoing population-based study, the Dallas Hearts and Minds Study (DHMS). MEG scans were performed using a 306-channel MEGIN Triux-Neo whole-head MEG system at the University of Texas Southwestern Medical Center. Three minutes of artifact-free resting-state MEG signals were extracted from 40 SMC patients, 27 MCI patients, and 48 healthy control. CWT was computed for each channel of MEG data. Radiomic features were extracted from the CWT matrix for five frequency bands (delta, theta, alpha, beta, gamma). The Boruta feature selection algorithm was used to identify the most relevant features for classification. Two classical machine-learning algorithms, Random Forest and XGBoost, were employed for the classification task. The Synthetic Minority Oversampling Technique (SMOTE) was used to address the issue of unbalanced datasets. A simple three-class classification (TCC) and a two-step-three-class classification (TSTCC) were implemented. TSTCC consists of two steps. First, a two-class classification was performed between healthy control and non-control (a combination of SMC and MCI). In step two, the subjects identified as non-control were classified into SMC and MCI. Furthermore, a simple two-class classification between healthy control and MCI was also implemented. The data was split into 80% training and 20% testing sets. This procedure was repeated 100 times by randomly dividing the data for each run.


Results, Conclusions, and Discussions:: The average accuracy of the 100 runs are presented in Table 1. XGBoost achieved an overall accuracy of 55.8% and 53.6% for TCC and TSTCC, respectively. Random Forest improved the overall accuracy to 67.5% and 65.6% for TCC and TSTCC, respectively. For the two-class classification, Random Forest achieved an overall accuracy of 81.9%, whereas XGBoost achieved 72.7%. Although the overall accuracies were not high, these three algorithms were successful in improving the accuracies of each class separately compared to the results before applying SMOTE.

We trained two ML algorithms for classifying healthy control, SMC, and MCI subjects and one for classifying healthy control and MCI. Random Forest outperformed XGBoost for most classification methods. SMOTE improved the classification accuracy by increasing the number of minority group cases in each classification algorithm. SMC has not been well investigated using MEG and ML, despite being an ideal target for pre-clinical interventions. The study demonstrates the potential of the MEG-CWT-ML approach in classifying SMC and MCI.


Acknowledgements (Optional): :

References (Optional): :