Comparison of methods for early detection of Alzheimer's disease

Abstract

In this paper, six methods for early detection of AD - fractal dimension (FD), source localization (SL), Hjorth analysis, cross mutual information (CMI), pdf of zero-crossing intervals (ZCI) and power spectrum (PS) are compared. We selected these methods because they were relatively insensitive to artefacts and gave promising results. The methods were applied to the EEGs of 38 mild AD patients and 45 normal subjects, to extract markers for early detection of AD. The datasets were obtained from three different countries and hence provided a more rigorous comparison. The performances of each method were measured using ROC analysis, sensitivity, specificity, classification accuracy, z-score and Area under ROC curve. Results showed that indices found using time domain methods, such as FD and ZCI outperform those obtained from frequency analysis. In particular, ZCI had the best overall performance for at least 50% of the datasets. We found that, although the PS results tend to agree with SL and ZCI, it is more sensitive to shifts in the alpha-theta rhythms. Results also show that FD and ZCI could potentially be use for early detection as their performances outperformed those used in current clinical AD diagnosis (sensitivity > 55%, specificity > 83%).