Elsevier

Infrared Physics & Technology

Volume 88, January 2018, Pages 200-205
Infrared Physics & Technology

Regular article
Identification of blood species based on diffuse reflectance and transmission joint spectra with machine learning method

https://doi.org/10.1016/j.infrared.2017.11.030Get rights and content

Highlights

  • Visible spectra more discrete than near-infrared for blood species identification.

  • The joint spectra outperform the separate visible or near-infrared spectra.

  • More information used to build the model, more accurate the predictions are.

Abstract

The identification of blood species is of great importance for the forensic science and wildlife investigations. Our previous work has proved that the diffuse reflectance visible spectra and transmission near–infrared spectra are effective in differentiating the blood species non-contactedly. This paper compared these two spectra’s abilities to classify the blood species, and proposed using these two spectra jointly to build the classification model. The machine learning method artificial neural network was the algorithm of model building. 1200 samples from five species were used to verify the recognition model. Results showed that the visible spectra outperform the near–infrared spectra, while the joint spectra perform better than the separate visible spectra. Therefore, the joint spectra of diffuse reflectance visible and transmission near–infrared provide more information of the blood components to distinguish the blood species more precisely.

Introduction

The identification of blood species is important in forensic science [1] and wildlife preservation. To achieve the aim of non-destructive measurement, Raman spectroscopy method was developed and proved to be an efficient technique to discriminate the blood species [2], [3]. However, this method cannot detect a blood sample placed in a specific container, such as an anticoagulation tube, which is a common used container for blood transportation. Our group demonstrated that the diffuse reflectance visible spectra and transmission near–infrared (NIR) spectra are useful in identifying the species of blood samples placed in anticoagulation tubes [4], [5], which could be very useful on screening blood samples for blood supervision.

The underlying reasoning of this method has two parts. Firstly, since Jöbsis [6] for the first time proposed the method of determining the blood components’ contents by NIR spectroscopy in 1970s. Lots of studies on blood spectrum analysis were conducted to establish the prediction models between the blood spectra and the standard contents of blood components [7], [8], [9], [10]. The latter is usually achieved by the biochemical methods. These previously published results indicate that the blood spectra can be used to predict the contents of the components in blood. Secondly, in the sexual creatures, the species is defined as a unified interbreeding group, consisting of populations with actually or potentially reproductive capacities. And they are reproductively isolated from other such groups [11]. Wang et al. [12] compared a number of blood physiological and biochemical indexes of cynomolgus monkeys and macaques, and found that there are significant differences in the blood components of these two groups. Ekser et al. [13] compared the hematologic parameters of pigs, baboons, cynomolgus monkeys, rhesus monkeys and humans. The published results indicate that the differences of blood components’ contents among different species are significant. Therefore, it is logically possible to predict the blood components’ contents firstly, and then discriminate the blood species based on the predicted blood components’ contents.

However, to logically infer a rule to describe the elements in the collection, we must have information about each element in the collection [14]. Which means we need to collect all the information of each blood component. This would be difficult in practice. Therefore, the machine learning methods are proposed to avoid this problem by offering only probabilistic laws, rather than the entirely certain laws used in purely logical reasoning. In this paper, the well-known classifier algorithm artificial neural network (ANN) was used to construct the spectra based blood species recognition models. The diffuse reflectance visible and transmission NIR spectra were jointly used to optimize the performance of the classification. A dataset consisting of 1200 samples was used to verify the proposed method.

Section snippets

Blood samples

1200 blood samples of dog, goat, rhesus monkey, rat and human were collected by Institute of Laboratory Animal Sciences, CAMS & PUMC and refrigerated to deliver to the spectral detection laboratory. Each species has 240 blood samples. Each blood sample was provided by an individual donor with about 5 mL, and placed in a Greiner VACUETTE vacuum blood collection tube. Ethylene diamine detraacetic acid (EDTA) was added into the blood sample to prevent blood clotting. The blood samples were

Prediction results

We compared the performances of classifications with visible spectra, NIR spectra and both visible and NIR spectra. The results shown in Fig. 6 indicate that the accuracy error decreases with the increment of the preserved PCs’ number of the dataset. Because the more PCs are preserved, the more information of the original dataset is transferred to build the classifier. Thus, more exactly is the built model able to describe the generative distribution of the dataset. In addition, the results

Conclusions

This study developed the spectral recognition method to identify a blood sample’s species rapidly and non-contactedly. The diffuse reflectance visible and transmission NIR spectra were jointly used as the dataset. The machine learning method ANN was used to build the recognition model. Though the diffuse reflectance visible spectra are more distinguishable than the diffuse transmission NIR spectra with respect to the species recognition, combination of these two spectra made the best

Conflict of interest

There are no conflicts of interest to declare.

Acknowledgements

This research was supported by National High-Tech R&D Program of China (Grant No. 2015AA021105). It was also supported by PUMC Youth Fund (Grant No. 2017310030).

References (18)

There are more references available in the full text version of this article.

Cited by (0)

1

Co-corresponding author.

View full text