Quality estimation of subdurally recorded, event-related potentials based on signal-to-noise ratio.

Our goal is to develop a direct brain interface (DBI) that will provide communication and environmental control to persons who are "locked-in" (or nearly so) as a consequence of brainstem stroke, amyotrophic lateral sclerosis (ALS), or other etiologies. Previously we demonstrated that templates constructed from trigger averaged event-related potentials (ERPs) can be cross-correlated with ongoing electrocorticograms (ECoGs) to detect ERPs associated with the performance of simple motor actions. However, it was difficult to predict a priori which of many candidate ECoG recording site(s) could provide signals that would provide adequate motor action detection. We present here a measure of ERP quality based on an estimate of the signal to noise ratio (SNR) associated with the formation of an ERP template from the performance of consecutive voluntary actions. Detection-theory-based receiver operator characteristics (ROCs) and a database of ECoGs (6000+) recorded from the cortical surface of awake human subjects were used to assess the usefulness of the SNR technique. The SNR method was found to predict the detection efficacy of ERPs when characterized over a wide parameter range, with the majority of ROC curve areas greater than 90%. This method was compared with our previously developed quality measure (the peak-to-baseline ratio) and found to provide significantly better performance (ROC area differences from 4.4% to 13.7%). Thus, the SNR estimate of the ERP is a useful tool to predict the efficacy of ERP templates for cross-correlation-based detection and assist in the selection of viable ERP templates for DBI applications.

Establishing likelihood ratios for patterned garment comparisons from seam measurement data.

It is often challenging to ascribe an objective measure of confidence for identifications based on surveillance imagery from a crime scene. The present work seeks to address this deficiency in the case of garment comparison evidence by developing a quantitative method for establishing a conservative lower bound on the likelihood ratio (LR) for identifications involving patterned garments. The method is based on statistical analysis of pattern offset measurements taken from a sample of garments of the same type (manufacturer, style, and size) as the seized evidence. The developed analysis framework was demonstrated on different types of garments over a range of modeled surveillance imaging scenarios with variable image quality; the lower bounds on the LRs ranged from approximately 10-1 to over 400-1. The statistical model was tested and validated through a large-scale empirical study involving both simulated and human observer-performed garment comparisons.