Multistart algorithms for MEG empirical data analysis reliably characterize locations and time courses of multiple sources.

We applied our newly developed Multistart algorithm (M. Huang et al., 1998, Electroencephalogr. Clin. Neurophysiol. 108, 32-44) to high signal-to-noise ratio (SNR) somatosensory responses and low SNR visual data to demonstrate the reliability of this analysis tool for determining source locations and time courses of empirical multisource neuromagnetic data. This algorithm performs a downhill simplex search hundreds to thousands of times with multiple, randomly selected initial starting parameters from within the head volume, in order to avoid problems of local minima. Two subjects participated in two studies: (1) somatosensory (left and right median nerves were stimulated using a square wave pulse of 0.2 ms duration) and (2) visual (small black and white bull's-eye patterns were presented to central and peripheral locations in four quadrants of the visual field). One subject participated in both of the studies mentioned above and in a third study (i.e., simultaneous somatosensory/visual stimulation). The best-fitting solutions were tightly clustered in high SNR somatosensory data and all dominant regions of activity could be identified in some instances by using a single model order (e.g., six dipoles) applied to a single interval of time (e.g., 15-250 ms) that captured the entire somatosensory response. In low SNR visual data, solutions were obtained from several different model orders and time intervals in order to capture the dominant activity across the entire visual response (e.g. , 60-300 ms). Our results demonstrate that Multistart MEG analysis procedures can localize multiple regions of activity and characterize their time courses in a reliable fashion. Sources for visual data were determined by comparing results across several different models, each of which was based on hundreds to thousands of different fits to the data.