Abstract:
|
The challenge of DNA mixture interpretation is at the core of forensic genetic identification; efficacy can significantly impact the course of criminal investigations and the quality of intelligence. Forensic scientists have relied on empirical studies and experience to aid in identification of donors in a mixture. Limitations are inherent in such analyses, and probabilistic approaches suffer from computational and time constraints, but expansive mixture data sets can be leveraged via machine learning algorithms. The power in wedding such an approach with expert systems stems from (1) a classifier's ability to learn from data and make predictions about previously unseen instances, and (2) human analysts' experience-derived rule sets, which can reduce data dimensionality without significant information loss. Multiple classifiers were evaluated for suitability for the problem domain. These algorithms learned models that can rapidly predict the number of contributors and deconvolute complex mixtures of at least three contributors. The data-agnostic nature of this approach affords increased flexibility in adapting to analyses of new data such as next generation DNA sequences.
|