Therefore, self-reported depressive symptoms did not improve the SVM prediction accuracy. Including data on CART CPE also failed to improve the prediction. For the scenario where log10 HIV RNA was included, the accuracy of the prediction was 75% for impairment and 72% for NP nonimpairment. These same accuracies were also achieved for the scenario where detectable vs. undetectable HIV RNA was used. Hence inclusion of CPE did not improve

prediction accuracy. Our study was conducted with the intention of generating an extra-brief tool to assist HIV physicians in referring HIV-positive persons at risk for NP impairment. We believe that our study provides a preliminary but robust solution to this first objective. Indeed, we found that our SVM-derived check details models yielded adequate prediction accuracy for NP impairment (sensitivity 78%; n=28/36) and NP nonimpairment (specificity 70%; n=43/61). These figures are certainly adequate for use of the algorithm as an adjunct clinical tool. Moreover, we believe that the predictions were quite good in comparison with predictions of HAND provided by brief paper-and-pencil NP instruments. Davis et al. [28] reported

70% sensitivity and 71% specificity for the HIV-dementia scale. Carey et al. [29] showed 78% sensitivity, 85% specificity and 83% overall prediction accuracy using two selected NP tests. The California Computerised Assessment Package (Calcap), a brief cognitive computerized test, yielded 68% sensitivity and 77% specificity [30]. Lastly, the brief computerized battery CogState demonstrated 81% sensitivity, MI-503 in vivo 70% specificity, and an overall prediction accuracy of 77% [31]. These accuracy rates provide preliminary support for application of these models in a clinical setting. In addition, this algorithm can be easily implemented on a web-interface platform (under construction) for which the HIV physician will only have to input

the necessary characteristics [for example when using the model determined from detectable levels of HIV RNA the required characteristics are: age in years; current CD4 T-cell count; presence or absence of past CNS HIV-related diseases (yes or no); and current CART duration in months]. The expected duration of the screening (computation Akt inhibitor of the algorithm including data entry with interactive instructions) is about 3 min. Here we have shown that it is the inclusion of easily ascertainable clinical factors that makes the algorithm practical. However, while the inclusion of the factors might be obvious, the relative weighting of each is certainly not. This study also contributes to the body of evidence on the use of SVM as a robust tool for data classification problems [18]. SVM methods have been increasingly used in a wide variety of medical classification problems.