Performance Evaluation and Fusion of Methods for Early Detection of Alzheimer Disease

Abstract

The number of people that develop Alzheimer's Disease (AD) is rapidly rising, while the initial diagnosis and care of AD patients typically falls on non-specialist and still taking up to 3-5 years before being referred to specialists. An urgent need thus exists to develop methods to extract accurate and robust biomarkers from low-cost and non intrusive modalities such as electroencephalograms (EEGs). Contributions of this paper are three-fold. First we review 8 promising methods for early diagnosis of AD and undertake a performance evaluation using ROC analysis. We find that fractal dimension (AUC = 0.989), zero crossing interval (AUC = 0.980) and spectrum analysis of power alpha/theta ratio (Pwr <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">alpha,thetas</sub> )(AUC = 0.975) perform best, with all three having sensitivity and specificity higher than 94%. We plot ROC curve with 95% confidence contours because of the small size of our data set (17 AD and 24 NOLD). Second, we investigate a fusion approach to combine these methods, using a logistic regression model, into one single more accurate biomarker (AUC = 1.0). Thirdly, to help support the distribution and use of these methods for early detection and care of AD, we developed them as web-services, integrated into online tools available from the BIOPATTERN project portal (www.biopattern.org).