Person Identification using E-Nose : A quantitative study on identifying individuals based on exhale patterns
Desai, Ashutosh (2024)
2024
Bioteknologian ja biolääketieteen tekniikan maisteriohjelma - Master's Programme in Biotechnology and Biomedical Engineering
Lääketieteen ja terveysteknologian tiedekunta - Faculty of Medicine and Health Technology
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-12-09
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2024120310713
https://urn.fi/URN:NBN:fi:tuni-2024120310713
Tiivistelmä
Identifying scents through digital means has long been of interest in many fields ranging from forensics to warfare and to medicine. Within medicine, ‘scents’ are essentially created by Volatile Organic Compounds (VOCs). These VOCs are used to detect diseases such as cancer and COVID-19. This study utilizes VOCs to generate exhale patterns via Ion Mobility Spectrometry (IMS) signals, commonly referred to as eNose signals, to uniquely identify individuals.
The research builds upon foundational studies by addressing limitations related to dataset size, session variability, and classifier performance. Breath samples were collected from 30 participants across three sessions held on different days in a controlled laboratory environment, ensuring consistent conditions for temperature, humidity, and pre-exhalation protocols.
The study evaluates six machine learning classifiers—Classification Decision Trees (CDTs), K-Nearest Neighbors (KNN), Naïve Bayes, Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and Support Vector Machines (SVM)—to assess their efficacy in distinguishing breath signatures. Signal preprocessing included detrending, Z-score normalization, and the use of absolute and relative signals to account for baseline variations. Classifiers were tested and validated with varying breath lengths (10s, 20s, and 30s) and different signal types.
Results from classification and Cross-Validation revealed accuracy greater than 90% in most cases. However, the F-Scores of all classifiers seemed to be very underwhelming, which indicated overfitting amongst all classifiers. Possible causes for this could be data imbalance, high dimensionality, and high feature correlations.
This research underlines the potential of IMS for non-invasive and rapid person identification. With applications spanning personalized diagnostics, airport security, and disease screening, the methodology represents a step forward in biometrics. Future directions include exploring neuromorphic systems for enhanced classification efficiency and integrating advanced data augmentation techniques to mitigate data imbalance. These refinements aim to further establish IMS as a viable tool in real-world biometric systems and personalized medicine.
The research builds upon foundational studies by addressing limitations related to dataset size, session variability, and classifier performance. Breath samples were collected from 30 participants across three sessions held on different days in a controlled laboratory environment, ensuring consistent conditions for temperature, humidity, and pre-exhalation protocols.
The study evaluates six machine learning classifiers—Classification Decision Trees (CDTs), K-Nearest Neighbors (KNN), Naïve Bayes, Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and Support Vector Machines (SVM)—to assess their efficacy in distinguishing breath signatures. Signal preprocessing included detrending, Z-score normalization, and the use of absolute and relative signals to account for baseline variations. Classifiers were tested and validated with varying breath lengths (10s, 20s, and 30s) and different signal types.
Results from classification and Cross-Validation revealed accuracy greater than 90% in most cases. However, the F-Scores of all classifiers seemed to be very underwhelming, which indicated overfitting amongst all classifiers. Possible causes for this could be data imbalance, high dimensionality, and high feature correlations.
This research underlines the potential of IMS for non-invasive and rapid person identification. With applications spanning personalized diagnostics, airport security, and disease screening, the methodology represents a step forward in biometrics. Future directions include exploring neuromorphic systems for enhanced classification efficiency and integrating advanced data augmentation techniques to mitigate data imbalance. These refinements aim to further establish IMS as a viable tool in real-world biometric systems and personalized medicine.