Spectra-based blood species discrimination by machine learning: Between human and non-human

With the fast development of data-processing technologies, more and more machine learning algorithms and a large volume of data were involved in spectra analysis. This study aims to validate the spectra' capability of discriminating blood species between human and non-human with a large volume of data and investigate the predictive ability of the binary classification model for unseen animal species. This study used eight machine learning algorithms and six balancing re-sampling strategies to train the binary classifier on 8431 samples from 14 species, including human. The hold-out experiment was conducted to evaluate the trained model's predictive ability for unseen animal species. The optimal binary classifier between human and non-human produced a testing accuracy of 98.9% and Youden's index of 0.976. This model was trained with the linear discriminant analysis with the re-sampling strategy of cluster centroids. Nine out of thirteen species produced sensitivities greater than 0.9 on the hold-out data for the hold-out experiment. This result indicates that the human and nonhuman blood classification model trained with a limited number of animal species could correctly recognize many unseen animal species as non-human and suggests careful data selection for the model building.

Automated summary

This study validates the capability of spectra to discriminate between human and non-human blood species, and investigates the predictive ability of a binary classification model for unseen animal species. The results show that the model trained with a limited number of animal species can correctly recognize many unseen animal species, highlighting the importance of careful data selection for model building.