A Combination of Particle Swarm Optimization and Minkowski Weighted K-Means Clustering: Application in Lateralization of Temporal Lobe Epilepsy

K-Means is one of the most popular clustering algorithms that partitions observations into nonoverlapping subgroups based on a predefined similarity metric. Its drawbacks include a sensitivity to noisy features and a dependency of its resulting clusters upon the initial selection of cluster centroids resulting in the algorithm converging to local optima. Minkowski weighted K-Means (MWK-Means) addresses the issue of sensitivity to noisy features, but is sensitive to the initialization of clusters, and so the algorithm may similarly converge to local optima. Particle Swarm Optimization (PSO) uses a globalized search method to solve this issue. We present a hybrid Particle Swarm Optimization (PSO) + MWK-Means clustering algorithm to address all the above problems in a single framework, while maintaining benefits of PSO and MWK Means methods. This study investigated the utility of this approach in lateralizing the epileptogenic hemisphere for temporal lobe epilepsy (TLE) cases using magnetoencephalography (MEG) coherence source imaging (CSI) and diffusion tensor imaging (DTI). Using MEG-CSI, we analyzed preoperative resting state MEG data from 17 adults TLE patients with Engel class I outcomes to determine coherence at 54 anatomical sites and compared the results with 17 age- and gender-matched controls. Fiber-tracking was performed through the same anatomical sites using DTI data. Indices of both MEG coherence and DTI nodal degree were calculated. A PSO + MWK-Means clustering algorithm was applied to identify the side of temporal lobe epileptogenicity and distinguish between normal and TLE cases. The PSO module was aimed at identifying initial cluster centroids and assigning initial feature weights to cluster centroids and, hence, transferring to the MWK-Means module for the final optimal clustering solution. We demonstrated improvements with the use of the PSO + MWK-Means clustering algorithm compared to that of K-Means and MWK-Means independently. PSO + MWK-Means was able to successfully distinguish between normal and TLE in 97.2% and 82.3% of cases for DTI and MEG data, respectively. It also lateralized left and right TLE in 82.3% and 93.6% of cases for DTI and MEG data, respectively. The proposed optimization and clustering methodology for MEG and DTI features, as they relate to focal epileptogenicity, would enhance the identification of the TLE laterality in cases of unilateral epileptogenicity.