Stephens_Alexandra_version7
- 1. Implementation of volatile organic compound identification algorithms using colorimetric sensor array data Alexandra Stephens Mentored by Dr. Alan Samuels and Dr. Charles Davidson Introduction Colorimetric sensor arrays create a way to “see smells.” These small “tickets” consist of 76 colored spots with different chemical compositions, such as metalloporphyrins and hydrogen bonding sites. When exposed to volatile organic compounds (VOCs) or other chemicals, the spots’ molecular structures foster various intermolecular reactions, ranging from Lewis donor/acceptor reactions to Brønsted acid/base reactions (Suslick, 2004). The result of these chemical changes is reflected in the change in color of the dots. The red, green, and blue (RGB) values of each spot are extracted through digital color imaging n times until the reaction is complete, creating an n by 228 (76 times three) matrix. Analysis of VOC ticket data is utilized in chemical identification; some methods of identification currently include computing the dot product of data sets, the k nearest-neighbor algorithm, and hierarchical cluster analysis. However, a permanent and accurate algorithm has yet to be established. The purpose of this investigation was to develop an approach that analyzes colorimetric sensor array data and correctly identifies at least 90% of chemicals. Methods and Materials Three separate approaches to chemical identification were developed and tested with a matrix of 34 chemical signatures, “S”, using the platform MATLAB®. Testing involved processing a copy of one of the chemicals in matrix S, as if it were an unknown substance, and comparing it to S by running the developed algorithm. The first idea was to differentiate between non-reacting spots, or “zeros,” and spots that experienced significant color change due to a chemical reaction. Graph 1 shows how the various amount of color change is reflected in the data. A threshold of what is to be considered an unchanging spot was calculated, and data within the range were assigned the value zero. The remaining spots were set equal to either one or negative one, regardless of magnitude (Graph 2). The same threshold was applied to the signature matrix S, and the program identified the chemical with the most matching values. The next identification method calculated the z-score of every RGB value of the “unknown” element, using the mean standard deviations of all 34 chemicals. The chemical with the smallest z-score summation was identified as the correct chemical, meaning that overall, it was the fewest-standard-deviations away from the mean. Mathematically, Z = x − μ σ . Originally, the dot-product method was used to compare the magnitude and direction of two, 228-dimensional vectors, identifying the substance with the highest dot product as correct. A normalized version of the dot product—the angle between two vectors—was implemented to compare only the direction of the vectors. The formula for computing the angle between vectors is cos θ = u · v u v . Conclusion The purpose of this study was to develop a program that analyzes colorimetric sensor array data and correctly identifies at least 90% of the 34 household chemicals given. The angle-between-vectors approach surpassed the 90% accuracy goal, and drastically improved upon the dot product approach. This is because it eliminates the potential to incorrectly identify a VOC due to high magnitudes of color change, found in chemicals such as Bleach. The one-zero method did surprisingly well, given the simplicity in the program. It performed slightly better than the z-score approach, which is more complex and uses more advanced statistics. The same issue arose in all of the identification programs: since some of these household substances were quite similar, for example, two different versions of OFF® insect repellent were tested, many of these substances identified as one another. This may be because the actual chemical compositions of these substances are so similar, the data varies only slightly, causing confusion in some or all of the identification algorithms. To advance this study, larger sample sizes should be used to ensure consistency of the programs. It is an important piece to the many applications of colorimetric sensor array data analysis. For example, lung cancer and other diseases can be identified through analysis of the breath of patients with colorimetric sensor arrays (Beukemann et. al., 2012). They serve as a less-invasive, less-expensive, and potentially more accurate diagnostic tool. This situation can be life threatening, and an identification program with high accuracy (at least 90%) is necessary. Results References Beukemann, M. C., Kemling, J. W., Mazzone P. J., Mekhail, T., Na, J., Sasidhar, M.,…Xu, Y. (2012). Exhaled breath analysis with a colorimetric sensor array for the identification and characterization of lung cancer. J Thorac Oncol, 7(1):137–142 doi: 10.1097/JTO.0b013e318233d80f Suslick, K. S. (2004). An optoelectronic nose: “seeing” smells by means of colorimetric sensor arrays. MRS Bulletin. Retrieved from www.mrs.org/publications/bulletin Graph 1: This is a graph of the change in the color values of each of the 76 spots on a colorimetric sensor array over time when exposed to the common household item, Permethrin. Graph 2: This graph shows the signature of Permethrin after it is altered by the first program. 5.88% 85.29% 82.35% 94.12% 0% 20% 40% 60% 80% 100% Dot Product One-Zero Comparison Z-Score Angle Between Vectors Chemical Testing Results Graph 3: This graph shows the results of testing an old identification method, the dot product, and the three new methods with 34 different colorimetric sensor array VOC data sets. Results(Continued) The angle-between-vectors approach was the most successful, only misidentifying two chemicals out of the 34. The dot-product was the least successful, as most chemicals were identified as either Bleach, Permethrin, or Hoppes #9, chemicals with generally high magnitudes. The remaining two codes performed relatively well, each identifying 29 to 30 chemicals correctly. Scaled Permethrin Signature 0 50 150 200100 ScaledRGBValue Band Number -1.0 -0.5 0.0 0.5 1.0 Permethrin Signature -1.0 -0.5 0.0 0.5 1.0 1.5 RGBValue Band Number 0 50 150 200100 × 104 PercentCorrectlyIdentified Methods and Materials (Continued)