Merrill, David A.

AbstractBackgroundMonitoring changes in cognitive impairment is an important aspect of care for aging patients. Traditionally, this is done with pen‐and‐paper assessments such as the Mini‐Mental State Exam (MMSE). EEG is a low‐cost and widely available neurophysiology tool that provides a data‐rich measure of brain health, but signal complexity has limited its impact. Advances in machine learning bring new possibilities to decode those complex signals and leverage EEG in patient care. Here, we engineer novel EEG features for assessing brain health with the [Banded Fractal Variability] (BFV) technique. We use [BFV] and other EEG metrics to assess patients’ cognitive function via a continuous score (a [Cognitive Impairment Index], or [CII]). We validate the diagnostic value of the EEG features by comparing our model against reduced models lacking the EEG features.MethodThe patient sample (N=97, mean (sd) age = 69.1 (11.0), 52.6% female) was collected at the Pacific Brain Health Center. Diagnostic categories included subjective cognitive impairment (n=44), mild cognitive impairment (n=26), and dementia (n=27). Mixed disease (AD, DLB, FTD, vascular, diabetes) was common in the impaired sample, and 39.6% were diagnosed with AD. We developed a continuous [CII] model trained on patients’ MMSE scores (range = 11 to 30, mean = 26.6) using a gradient boosting model with [Banded Fractal Variability] and other EEG features computed from clinical wake EEGs. Model development and testing took place within a rigorous nested cross‐validation (NCV) procedure to prevent overfitting. Performance was compared to reduced (no‐EEG features) models using the non‐parametric Wilcoxon sign‐rank test.ResultThe approach showed a Mean Absolute Error of 1.98 (sd = 0.28) across 5 test folds in the NCV outer loops, significantly outperforming both the baseline intercept‐only model (p = 0.031) and the demographics (age and years of education) model (p = 0.031).ConclusionThese preliminary results suggest that our [CII] derived from machine‐learning and quantitative EEG using [BFV] is a viable approach to assess and track patients’ cognitive impairment regardless of disease status across all impairment levels, raising the possibility that a brief EEG scan could shed light on patient health normally accessible only through extensive cognitive and neuropsychological testing.

AbstractBackgroundAlzheimer’s disease (AD) lacks a fast, easy, reliable, and inexpensive method of diagnosis. Currently, diagnosis is based on time‐consuming behavioral tests and the exclusion of other potential causes of impairment. Several biomarkers show good or promising diagnostic performance (e.g. CSF, tau PET, MRI, blood), but are either expensive, invasive, or still in development and while some can detect preclinical disease stages, all appear slow to change relative to the rate of cognitive decline. Here we develop a prototype diagnostic classifier based on novel metrics of brain activity in resting state electroencephalography (EEG) that correlates well with mental status.MethodArchival resting‐state EEG recordings of older adults (N=248) came from a memory clinic and university‐based clinic with a range of clinical diagnoses including subjective cognitive impairment (SCI), mild cognitive impairment (MCI), and dementia, representing AD, vascular dementia, and Lewy body dementia, TBI, and depression. We developed XGBoost classifiers to detect AD using EEG, age, and sex under increasingly challenging conditions. We computed metrics of periodic and aperiodic brain activity using the FOOOF algorithm. Furthermore, we developed a novel technique called [Banded Fractal Variability], which yields a set of features based on fluctuations in the fractal dimension within canonical frequency bands. We trained classifiers using cross‐validation to avoid overfitting during hyperparameter selection.ResultAlong with ROCAUC, we report an optimal sensitivity, specificity, and accuracy for the point on the ROC curve that maximizes Youden’s j. The classification tasks were healthy vs. probable AD (ROCAUC = 98%, sensitivity = 90%, specificity = 99%, accuracy = 96%), SCI vs. mild AD (ROCAUC = 89%, sensitivity = 76%, specificity = 87%, accuracy = 84%), and AD vs. other pathologies (no AD diagnosis) in MCI and dementia patients (ROCAUC = 82%, sensitivity = 72%, specificity = 87%, accuracy = 80%). All ROCAUC values were stronger than would be expected by chance (ps < 0.001).ConclusionThese preliminary results suggest that AD could be diagnosed in the clinic on the basis of machine‐learning classifiers and resting‐state EEG. Furthermore, they demonstrate that [Banded Fractal Variability] carries clinically‐relevant information about AD.