Real‐time prediction of working memory performance: A machine learning‐based approach towards dementia prevention

Abstract

AbstractBackgroundMCI is a clinical state that typically precedes Alzheimer’s dementia (AD). Working memory deficits are common in MCI and affect routine activities. Thus, developing an intervention that enhances working memory could enhance day‐to‐day function in these patients and, in turn, prevention progression to dementia. Towards this goal, a model that predicts individual‐specific working memory performance in this population could be instrumental for personalized interventions. Electroencephalography (EEG) captures the time dimension of cognitive events that happen during working memory performance and could predict individual‐specific performance. EEG predictive markers can then be targeted by interventions to enhance performance.MethodWe propose a single‐trial classification process that predicts individuals’ responses i.e. target correct (TC) vs. target noncorrect (TNC) responses during a working memory task, N‐back. We applied this process to EEG data of 15 healthy participants (mean age (SD) = 29.8 (7.6)) while performing the 3‐back task. We used event related (de‐)synchronization (ERD/ERS) from EEG signals 600 milliseconds prior to stimulus presentation as input features to a support vector machine classifier. To avoid overfitting of the model, we applied recursive feature elimination and cross validation to the first two‐thirds of the task. A trained classifier was then tested on the last third of the session. Non‐parametric permutation testing was used to ensure that the extracted pattern is associated with the original data rather than a random pattern.ResultOur model identified the brain regions where ERD/ERS predicted each individual’s working memory performance. Mean (SD) prediction accuracy across 15 participants was 70.1% (5.9). Accuracy was significantly above chance in 12 out of the 15 participants. The total number of the predictive EEG features across all participants ranged between 4 and 9. The mode was 6. As an example, in one participant, we achieved 69.2% accuracy based on 6 features: decreased parietal theta ERS; increased prefrontal theta ERS; decreased right temporal and occipital gamma ERS; and increased frontal and right temporal alpha ERD.ConclusionThis pilot study could lead to a machine‐learning based approach to increase the efficacy of personalized AD preventative interventions by individualizing the targets for these interventions.