Applying an interpretation model for body fluid mixture identification.

Area of Science:

• Forensic Science
• Molecular Biology
• Computational Biology

Background:

• Accurate body fluid identification is crucial in forensics, especially for complex samples like vaginal material.
• Current methods like RT-PCR may lack sensitivity or specificity, and struggle with mixed-origin samples.
• Interpreting body fluid data involves categorical or probabilistic approaches, with limitations in utilizing all information.

Purpose of the Study:

• To explore multi-class machine learning for predicting body fluid components in mixed samples.
• To enhance body fluid identification accuracy and incorporate uncertainty.
• To evaluate the utility of quantitative versus binary data and model unknown sample categories.

Main Methods:

• Utilized multi-class machine learning classifiers trained on single-source and mixed body fluid samples.
• Compared the informativeness of quantitative (presence/absence) data for predicting mixture ratios.
• Investigated metric learning and leave-one-type-out simulation for modeling unknown sample categories.

Main Results:

• High prediction accuracy was achieved by including mixture profiles as distinct classes in training data.
• Quantitative body fluid information proved more informative than binary data for predicting mixture proportions.
• The study explored effective approaches for modeling and predicting an 'unknown' category.

Conclusions:

• Machine learning classifiers, particularly when trained on mixtures, offer a powerful tool for body fluid identification.
• Quantitative data analysis significantly enhances the accuracy of forensic body fluid mixture analysis.
• Advanced modeling techniques can address the challenge of identifying unknown or ambiguous body fluid samples.