Classification of ADHD Using Heterogeneity Classes and Attention Network Task Timing

Throughout the 1990s ADHD diagnosis and medication rates have increased rapidly, and this trend continues today. These sharp increases have been met with both public and clinical criticism, detractors stating over-diagnosis is a problem and healthy children are being unnecessarily medicated and labeled as disabled. However, others say that ADHD is being under-diagnosed in some populations. Critics often state that there are multiple factors that introduce subjectivity into the diagnosis process, meaning that a final diagnosis may be influenced by more than the desire to protect a patient's wellbeing. Some of these factors include standardized testing, legislation affecting special education funding, and the diagnostic process.

In an effort to circumvent these extraneous factors, this work aims to further develop a potential method of using EEG signals to accurately discriminate between ADHD and non-ADHD children using features that capture spectral and perhaps temporal information from evoked EEG signals. KNN has been shown in prior research to be an effective tool in discriminating between ADHD and non-ADHD, therefore several different KNN models are created using features derived in a variety of fashions. One takes into account the heterogeneity of ADHD, and another one seeks to exploit differences in executive functioning of ADHD and non-ADHD subjects.

The results of this classification method vary widely depending on the sample used to train and test the KNN model. With unfiltered Dataset 1 data over the entire ANT1 period, the most accurate EEG channel pair achieved an overall vector classification accuracy of 94%, and the 5th percentile of classification confidence was 80%. These metrics suggest that using KNN of EEG signals taken during the ANT task would be a useful diagnosis tool. However, the most accurate channel pair for unfiltered Dataset 2 data achieved an overall accuracy of 65% and a 5th percentile of classification confidence of 17%. The same method that worked so well for Dataset 1 did not work well for Dataset 2, and no conclusive reason for this difference was identified, although several methods to remove possible sources of noise were used. Using target time linked intervals did appear to marginally improve results in both Dataset 1 and Dataset 2. However, the changes in accuracy of intervals relative to target presentation vary between Dataset 1 and Dataset 2. Separating subjects into heterogeneity classes does appear to result in good (up to 83%) classification accuracy for some classes, but results are poor (about 50%) for other heterogeneity classes. A much larger data set is necessary to determine whether or not the very positive results found with Dataset 1 extend to a wide population.