In February 2009, the National Academy of Sciences published a report entitled "Strengthening Forensic Science in the United States: A Path Forward." The report notes research studies must be performed to "…understand the reliability and repeatability…" of comparison methods commonly used in forensic science. Numerical classification methods have the ability to assign objective quantitative measures to these words. In this study, reproducible sets of ideal striation patterns were made with nine slotted screwdrivers, encoded into high-dimensional feature vectors, and subjected to multiple statistical pattern recognition methods. The specific methods employed were chosen because of their long peer-reviewed track records, widespread successful use for both industry and academic applications, rely on few assumptions on the data's underlying distribution, can be accompanied by standard confidence levels, and are falsifiable. For PLS-DA, correct classification rates of 97% or higher were achieved by retaining only eight dimensions (8D) of data. PCA-SVM required even fewer dimensions, 4D, for the same level of performance. Finally, for the first time in forensic science, it is shown how to use conformal prediction theory to compute identifications of striation patterns at a given level of confidence.

Over the last several decades, forensic examiners of impression evidence have come under scrutiny in the courtroom due to analysis methods that rely heavily on subjective morphological comparisons. Currently, there is no universally accepted system that generates numerical data to independently corroborate visual comparisons. Our research attempts to develop such a system for tool mark evidence, proposing a methodology that objectively evaluates the association of striated tool marks with the tools that generated them. In our study, 58 primer shear marks on 9 mm cartridge cases, fired from four Glock model 19 pistols, were collected using high-resolution white light confocal microscopy. The resulting three-dimensional surface topographies were filtered to extract all "waviness surfaces"-the essential "line" information that firearm and tool mark examiners view under a microscope. Extracted waviness profiles were processed with principal component analysis (PCA) for dimension reduction. Support vector machines (SVM) were used to make the profile-gun associations, and conformal prediction theory (CPT) for establishing confidence levels. At the 95% confidence level, CPT coupled with PCA-SVM yielded an empirical error rate of 3.5%. Complementary, bootstrap-based computations for estimated error rates were 0%, indicating that the error rate for the algorithmic procedure is likely to remain low on larger data sets. Finally, suggestions are made for practical courtroom application of CPT for assigning levels of confidence to SVM identifications of tool marks recorded with confocal microscopy.

No abstract