Abstracts

Rationale: This paper describes a pattern classification paradigm using nonlinear decision functions (NDF) as means to automatically categorize language related fMRI brain activation maps into typical and atypical groups within a large heterogeneous population. Data was provided by a multisite consortium dedicated to pediatric epilepsy research involving 13 hospitals. Methods: NDF under different dimensions and with different degrees of complexity were applied in association with the eigenvectors of the principal component analysis (PCA). 400 synthetic datasets were generated based on real datasets collected from 122 subjects. The well-established support vector machines (SVM) method is also used for comparative purposes. Results: In the testing phase using synthetic data, high classification results were obtained with an accuracy of 96%, a sensitivity of 97%, a specificity of 95%, and a precision of 95%. These optimal results were obtained with the use of 4 dimensions (eigenvectors) and a degree of complexity of 7. These results are given in Table 1 with SVM included for comparative purposes. Moreover, based on the best NDF classifier, two distinct activation patterns among the 122 real datasets were identified as illustrated in Figure 1. In order to assess the significance of these groupings, the results were compared with those obtained using clinical rating and lateralization index (LI). Good agreements were found for both: 82.79% agreement with LI (Kappa 0.592) and 81.15 % agreement with visual rating (Kappa 0.548). Conclusions: The data-driven mechanism using NDF was found to be effective at classifying typical from atypical language networks activation patterns, even from a heterogeneous population often acquired with different acquisition parameters. The integration of PCA with the NDF classification paradigm results in a data-driven method that is both accurate and computationally appealing (within few seconds in processing time after the weights of the decision function are generated in the training phase). This could promote objective assessments of large data sets and to interrogate data for a multitude of clinical variables.