Fault Detection with the Kullback-Leibler Divergence
Elements: Fault detection, innovative analysis, morphological image processing, science writing, publishing tools and standards, Mathematica, intellectual property, interdisciplinary collaboration
Links: Pre-peer-review manuscript (PDF)
Published journal article (behind paywall)
Published article also available upon request.
Summary

I developed a fault detection technique for large photomicrograph data sets. Protein drugs (such as insulin and an enormous variety of immune modifying drugs) may contain unintended particles that trigger an immune response, and in some patients the immune response can increase over time to a level that requires discontinuing a life-saving drug. I proposed to compare particle image data sets by using the Kullback-Leibler divergence (KLD). In one application, we generated scatter plots reflecting the similarity/dissimilarity of image data sets. In another application, we found that the method succeeds in classifying samples by stress condition, and, once trained, is able to identify the stress that caused particle formation in new samples. It was even able to identify the mixing ratio of different sample types, which could be applied to identifying counterfeit protein drugs.