Applying an interpretation model for body fluid mixture identification

Confirmatory body fluid identification via the detection of specific messenger RNA (mRNA) species is important in circumstances where a conventional test is not available for the questioned fluid type (such as vaginal material) or when greater sensitivity, or specificity, is required. Samples forwarded for testing will often contain two or more body fluid types from different donors (mixtures). Endpoint reverse-transcription PCR (RT-PCR) or real-time quantitative reverse-transcription PCR (RT-qPCR) are two methods which may be used for confirmatory body fluid identification, with the former currently in use in forensic casework laboratories. The interpretation of such data has been a topic of recent interest, with two main approaches proposed: categorical and probabilistic. Categorical methods, which may use a scoring system or a threshold approach, do not utilise all relevant information. Probabilistic methods are able to incorporate known information and reflect uncertainty as part of classification. This paper explores the use of multi-class machine learning classifiers to predict the components of body fluids in mixed samples, extending prior work on single-source body fluids. When including mixture profiles within the training data as an additional class, we obtained high prediction accuracy. We also determine that quantitative information is more informative than binary (presence/absence) data for the prediction of body fluids in different mixture ratios. Finally, we investigate the modelling and prediction of an "unknown" category, composed of samples which do not have specific features in the data, using two approaches: metric learning and a leave-one-type-out simulation for missing data.