Igor Lednev and his lab at the University at Albany have been working for years to develop forensic methods that allow criminal investigators to narrow down the search for a suspect. His Raman-based methods can determine the time since fluid deposition, differentiate between menstrual and peripheral bloodstains, differentiate between human and animal blood, as well as determine the race, sex and age group of a biological stain donor. And now, thanks to recent research, Lednev can add differentiating between smokers and non-smokers to the toolkit.
In a proof-of-concept study, recently published in the Journal of Biophotonics, Lednev and his collaborators at Kuwait University used non-destructive Raman spectroscopy to analyze dry oral fluid to differentiate between smoker and non-smoker donors—with the help of an artificial neural network.
Nicotine, the major component in tobacco, has a very short half-life and is rapidly absorbed into the body, making it a difficult biomarker to detect in smokers. Thus, the researchers targeted cotinine, nicotine’s main metabolite, which has a much longer half-life.
For the study, the researchers analyzed a total of 32 oral fluid samples from donors of differing gender, age and race. However, they were unable to detect a Raman signature of cotinine.
“The Raman bands of pure cotinine overlap with Raman bands of oral fluid, which made it difficult to differentiate between them,” Lednev explained to Forensic. “Oral fluid is a complex matrix that contains a mixture of secretions from major and minor salivary glands and non-salivary constituents, such as proteins, small molecules and electrolytes. The majority of these components contribute to a complex Raman spectrum of oral fluid.”
So, the researchers turned to machine learning to develop a “spectroscopic” signature instead. They employed a general statistical program first to identify eight spectral regions that contribute the most to the differentiation of smokers and nonsmokers in oral fluid. The bands located in the selected regions correspond to various biomolecules, including proteins, amino acids (such as tryptophan and histidine), as well as acetates.
“Neural networks are computational-based empirical models that resemble the human brain. When an artificial neural network is given sufficient input data, it can learn from it and predict a specific output,” said Lednev. “In this case, the power of machine learning supervised statistics is that it allows for identifying spectral features to differentiate classes by analyzing numerous Raman spectra with a known class assignment, or training data set.”
The team, which included Kuwait’s Entesar Al-Hetlani and Mohamed Amin, built multiple neural networks to determine which network architecture and parameters would provide the best performance results. Ultimately, a classification model was developed based on the artificial neural network that showed 100% accuracy after external validation.
The researchers say further investigations are necessary to conclusively assign the identified spectral features to specific biomolecules and understand why the abundance of these biomolecules changes when people smoke. The proof-of-concept study was also based on a small sample size.
“We need to significantly increase the number and variety of donors involving people of different age, sex and race,” Lednev said. “It is also important to rule out potential interfering factors, such as medical conditions and illnesses, consumption of food and drinks before and even during spilling oral fluid.”
Further tests could also clarify if the method allows for differentiating between regular smokers and someone that just smoked the day of the crime—knowledge that could be exploited by criminals.
The research study was supported by grants from the National Institute of Justice, U.S. Department of Justice and Kuwait University Research Administration.
Photo: Igor Lednev in his laboratory at University at Albany. Credit: Mark Schmidt/UAlbany