Description:FIELD OF INVENTION
The main field of invention involves creating and deploying an innovative electronic nose system equipped with advanced sensors to detect illegal drugs like cannabis and methamphetamine with high accuracy and efficiency, contributing to enhanced security measures.
BACKGROUND OF INVENTION
Detecting drugs is vital for law enforcement, workplace safety, and addiction treatment, impacting public health, safety, and crime rates. Failure to detect drugs risks their spread in communities, leading to crime and violence. Drug use at work poses safety risks, notably in transport, construction, and healthcare. Patients using illegal drugs risk health complications, and addiction contributes to social issues like poverty and family problems. Legal consequences include fines, imprisonment, and job loss. Various detection methods like mass spectrometry (MS) and Surface-Enhanced Raman Scattering (SERS) are effective but costly. Microfluidics offer potential but have fabrication challenges, while optical spectroscopy has limitations. Canine detection is reliable but resource-intensive. Developing electronic nose (e-nose) systems for drug detection faces challenges in verifying their VOC-based drug detection ability. Recent studies show promise in distinguishing drugs based on their VOC profiles. E-nose systems offer speed and cost advantages, ideal for on-site screening. To enhance performance, multiple sensors and modalities are recommended. Our study uses 56 sensors, including MOS, EC, NDIR, and PID, aiming for robust detection. We also test in diverse conditions, crucial for real-world applicability, and employ forward-feature selection for sensor optimization, ensuring accuracy.
Commercializing e-nose technology requires validation and cost considerations. Successful companies like MyDx Inc. and Electronic Sensor Technology, Inc. offer strategies such as product innovation, targeted marketing, and continuous improvement to penetrate the market and stay competitive.
The patent application number 202041053820 discloses a method for detection and quantification of nitrosamine impurities in a drug sample.
The patent application number 202017030155 discloses a bioassay for the non-invasive detection of drug use and physiologic conditions.
The patent application number 202017017258 discloses a end-of-dose detection for drug delivery system.
The patent application number 201911046436 discloses a system and method for detection of drug abuse using retinal imaging.
SUMMARY
We proposed a multi-sensor e-nose system and enhanced its performance across different environmental conditions by selecting the most relevant sensors. This research has the potential to create a highly accurate, compact portable device capable of detecting illegal drugs effectively in diverse settings. We acknowledge that our study's tested environments may not fully represent the complexity of real-world scenarios like border crossings or smuggling routes. However, we view our current experimental approach as an initial step toward developing a precise e-nose system for detecting illegal drugs in various environments. To achieve this, additional experimental procedures are necessary, which we intend to explore in future studies. These procedures involve gathering air samples from different locations, analyzing them using an e-nose system, fine-tuning detection parameters, validating results, and implementing the system in the field with robust security measures and real-time data analysis. This comprehensive approach not only facilitates drug detection in practical settings but also provides a roadmap for commercializing e-nose technology. Proper execution of these procedures can significantly contribute to national security efforts by combating drug trafficking and enhancing public safety.
DETAILED DESCRIPTION OF INVENTION
Comprehensive Design and Functional Components of a Multi-Sensor E-Nose System for Precise Drug Detection
Our research endeavors to create a swift and precise approach for identifying illegal drugs utilizing the proposed e-nose system, which mimics olfactory senses to detect and analyze various odors or VOCs in a sample. The system comprises four primary components: a control unit, a sensor unit, a gas supply unit, and a system facility unit, as illustrated in Fig. 1(a). The control unit configures experiment parameters, such as gas type, flow rate, and dilution ratio with ambient air, through a developed GUI (Labview 2015, National Instruments, Texas, USA). The gas supply unit features separate inlets for high-purity air, three target gases, and six distinct normal air sources from diverse laboratory settings. This unit utilizes a mixing chamber to combine each of the three target gases with the six normal air sources sequentially, creating a well-mixed and uniform gas mixture that undergoes predetermined dilution. The sensor unit comprises multiple commercial sensors, including 39 MOS, 8 EC, 8 PID, and 1 NDIR sensor, carefully selected for comprehensive detection, enhanced accuracy, and reduced interference when detecting a broad spectrum of drugs with varying chemical properties emitting diverse gases or volatile organic compounds. The injected gas is evenly distributed among eight flat cylindrical chambers, each consisting of four pairs of two-layer structures. Each gas chamber module, as depicted in Fig. 1(b), is constructed from aluminium and coated with an anodized layer to prevent adsorption and gas reactivity. Additionally, eight sensor boards are positioned around each chamber, with each board housing a sensor placed inside the chamber. Fig. 1(c) illustrates the block diagram outlining the e-nose system's operation. A detailed list of all 56 sensors in our e-nose system is provided in Table S1. The system facility unit acts as a central hub that powers, connects, and transmits real-time data to the control unit, ensuring efficient management of system performance, accurate data transmission, and TCP/IP communication.

Figure 1: Comprehensive View of the E-Nose System Components and Setup. (a) Image depicting the entire e-nose system setup, (b) Flat cylindrical gas chamber with 8 sensor boards, (c) Block diagram illustrating the operation of the e-nose system.

The study utilized drug samples (CA, ME, and TB) to investigate target gases. ME, a highly addictive stimulant made from various ingredients like pseudoephedrine, and CA, containing THC responsible for its intoxicating effects, were among the illicit drugs studied. These drugs are prevalent in illegal drug markets and are frequently trafficked and illegally imported through ports. Fig. 2(a) shows recently confiscated ME (in powder form) and CA (ground hemp leaves) samples from an illegal distribution process, stored separately in unscented plastic bags. TB, also widely trafficked and illegally imported through ports, was included as a target sample (ESSE Prime, KT&G, Korea). All drugs used in the experiments were approved by the Korea Ministry of Food and Drug Safety, and the experiments adhered to drug control laws. To ensure precise and reliable data, gas emitted from the target drugs was collected as shown in Fig. 2(b), with the inflow of external environmental gas blocked, and samples automatically collected at regular intervals.

Figure 2: Visual Representation of Drug Samples. (a) Images displaying drug samples of cannabis (CA), methamphetamine (ME), and tobacco (TB), (b) Illustration of the target drug gas collection process.

Methodology and Data Acquisition for Drug Detection using E-Nose System
We developed a 30-liter compression bombe to collect and store drug samples in separate airtight chambers (20 ×20×20 cm) to prevent interference and facilitate target gas selection. During sample collection, we used an oil-free vacuum pump (ISP-50, ANEST IWATA, Japan) to purge initial gas and maintain chamber pressure with high-purity air for 10 min, ensuring precise data collection. Data acquisition involved measuring the e-nose response to target gases diluted with normal air in six different environments at a flow rate of 250 ml/min. Target gas dilution (10–20%) reflected varied environmental compositions, enhancing adaptability and robustness. Sensor response, defined as Ra/Rg (sensor resistance in high-purity air vs. presence of target gas), was characterized by sensor transient response (STR), encompassing baseline, response, and recovery periods.
For drug detection classification, normalized data from 56 sensors were input into SVM, LDA, KNN, and RF classifier models, analyzing 720 samples over 6 min. Validation involved training models on data from five laboratories and testing on the sixth, ensuring accurate detection in diverse environments. Forward-feature selection optimized sensor choice, reducing overfitting risk and enabling development of a precise, portable device.
Sensor Sensitivity and Selectivity Variation in Drug Detection
Figure 3 illustrates the sensitivity and selectivity differences among four sensors in response to three target gases (CA, ME, and TB), measured by our e-nose system. The data presents the mean and standard deviation of normalized sensor responses. Sensor (a) exhibits similar reactivity to all three drugs, indicating cross-sensitivity and making it challenging to distinguish between target gases. Conversely, sensor (b) demonstrates exceptional selectivity for TB, while sensor (c) shows high reactivity to CA. Similarly, sensor (d) displays the highest reactivity towards ME and exhibits discerning capability, as its responses to the other two targets were also distinct. These results highlight the importance of using a diverse array of sensors and modalities to enhance the accuracy of detecting specific drugs in tested samples, as this approach increases the likelihood of utilizing sensors highly responsive to a particular drug.

Figure 3: Sensor Reactivity to CA, ME, and TB. Mean and standard deviation of normalized sensor responses are depicted for four representative sensors (a-d). Sensor (a) demonstrates cross-sensitivity to all three drugs, while sensor (b) exhibits high selectivity for TB, sensor (c) for CA, and sensor (d) for ME.

Impact of Environmental Variability on E-Nose Detection of Drug Targets
This section examines the influence of diverse environmental conditions on e-nose system performance in detecting drug targets. Fig. 4(a) displays the results of principal component analysis (PCA) based on three different target data diluted solely with air from a specific laboratory (En1). It highlights the ease of classification tasks when data acquisition and validation occur in a controlled experimental environment, contrasting with real-world conditions. Fig. 4(b) illustrates PCA results from 100% unmixed normal air samples collected from six distinct laboratories (En1 ∼ En6), demonstrating unique clustering for each laboratory, signifying varying environmental characteristics. In Fig. 4(c), we present clustering data for each of the three targets when diluted with 20% air from six different environmental sources. The wider distribution of clusters compared to (a) indicates increased difficulty in classification tasks. Fig. 4(d) shows the same data as in (c), with different colors representing each environmental laboratory mixed with each target, highlighting that scattering effects observed in (c) were influenced by environmental air variations across different laboratories. Each target data is mixed with different airs in each environment, making target clustering more challenging. These findings underscore the importance of assessing the ability to detect drug targets across diverse environmental settings.

Figure 4: Influence of Environmental Variables on E-Nose Target Detection. (a) Principal Component Analysis (PCA) based on three different target data diluted exclusively with air from a specific laboratory (En1). (b) PCA analysis conducted on 100% unmixed normal air samples collected from six different laboratories (En1 ∼ En6). (c) Clustering data for each of the three targets when diluted with 20% air from six different environmental sources. (d) Same data as in (c), with distinct colors representing each environmental laboratory mixed with each target.

Comparative Analysis of Machine Learning Models for Drug Detection
We conducted a discriminant classification analysis using machine learning algorithms, namely SVM, LDA, RF, and KNN, utilizing data from all 56 sensors. Our analysis encompassed 5166 datasets, including non-data (0% target gas) and three target gases (CA, ME, and TB), with dataset counts as follows: Non (no target, 0%): 1724, TB: 1708, CA: 1002, ME: 732. The RF model achieved the highest accuracy of 87.77%, as depicted in Fig. 5(a). Fig. 5(b) illustrates individual performance for each environment-based data tested for models trained in different environments, revealing performance variations attributed to environmental components. Fig. 5(c) presents the confusion matrix of the four models, with blue boxes indicating accurate predictions and red boxes indicating inaccuracies. The SVM model showed poor accuracy, particularly in detecting CA and ME, while the RF model exhibited superior performance across all classes without bias towards any specific class. This superiority can be attributed to RF's ability to handle high-dimensional data, nonlinear relationships, missing or noisy data, and a large number of features without overfitting. Additionally, our STR-based approach outperformed traditional feature-based methods, as evidenced by ROC curves and accuracy comparisons shown in Fig. S3 and Table S2, respectively. These findings underscore the unique and superior performance of our STR-based method for drug detection.

Figure 5: Performance of E-Nose System for Drug Detection in Varied Environments. (a) Overall average accuracy of four classifiers (SVM, LDA, RF, KNN). (b) Individual performance across diverse environments. (c) Confusion matrix depicting class-wise performance for different classifiers.

Optimal Sensor Selection Using Forward Feature Selection
The findings reveal that the system's performance improved with the inclusion of more sensors, up to a certain threshold. Using a single sensor alone yielded a performance of 78.83%, highlighting the inadequacy of singular sensor usage for accurate detection. However, the performance significantly enhanced to 88.25% when eight optimal sensors were selected, surpassing the performance achieved with all 56 sensors (87.77%). Notably, the system's performance peaked at 93.03% (with the lowest error rate) when only 24 optimized sensors were utilized, gradually declining with an increase in sensor count and reaching 87.77% with all 56 sensors in use. The list of the 24 selected optimal sensors is detailed in Table S3, showcasing their primary targets and selection order based on priority.
These results suggest that the optimal sensor set for drug detection may not necessarily be the largest available set. The decline in performance with all 56 sensors could be attributed to cross-sensitivity among sensors, leading to increased sensitivity to irrelevant features or noise in the data. Therefore, preemptively eliminating sensors with minimal differentiation in response characteristics could enhance optimization. However, sensor selection should also consider their ability to complement and enhance each other's performance. AI models may identify patterns that are not immediately apparent to human observers.
Furthermore, our analysis indicates a high proportion of EC sensors among the selected 24 sensors, likely due to their high selectivity for vapors from the tested drugs. EC sensors exhibit specific electrode material reactions with certain gases, resulting in highly selective and sensitive detection compared to MOS sensors, which have a broader detection range but may lack selectivity.
In future studies, we plan to delve deeper into analytes with the highest sensor response and conduct more comprehensive analyses of drug gas components, providing valuable insights for enhancing system performance in drug vapor detection.
Performance Improvement Analysis with Optimal Sensor Combination
The comparison between using all 56 sensors and the optimal combination of 24 sensors selected through forward-feature selection is presented here. Fig. 7(a) showcases individual performance improvements for each environment-based data when tested against models trained in different environments using the optimal sensor combination. The results illustrate significant performance enhancements with the optimal sensor combination for each environment-based data.
Fig. 7(b) presents the average performance comparison between using all sensors and the optimal sensor combination, demonstrating improved performance with the latter. Finally, Fig. 7(c) displays the confusion matrix for the case of using the 24 optimal sensor combinations, highlighting superior accuracy in classifying target gases compared to using all sensors. The optimized sensor combination accurately classifies the majority of test samples without significant bias towards any particular class.

Figure 6. E-nose System Performance with Varying Numbers of Optimized Sensors Using Forward Feature Selection.

Figure 7. Performance Comparison between All 56 Sensors and Optimal 24-Sensor Combination using RF Model. (a) Performance in Different Environments, (b) Average Performance, (c) Confusion Matrix for Optimal 24-Sensor Combination.
The analysis of receiver operating characteristic (ROC) curves and area under the curve (AUC) for four different models (SVM, LDA, RF, and KNN) demonstrates that the optimal combination of 24 sensors significantly enhances the performance of all models. As depicted in Fig. 8, the ROC curves for each model using the optimal 24 sensors exhibit improved positioning towards lower false negative and false positive rates, indicating superior performance compared to utilizing all 56 sensors. Table 1 presents comprehensive accuracy and AUC values, highlighting the performance comparison between the optimal sensor combination and the use of all sensors. Notably, the RF model achieved the highest accuracy among the four models, reaching 93.03% accuracy and an AUC value of 0.9891 with the optimal sensor combination. The LDA model also exhibited a substantial increase in accuracy from 74.31% to 80.64%, accompanied by an improved AUC value of 0.9493. Furthermore, the optimal sensor combination significantly enhanced the SVM model's performance from 46.96% to 80.80%, while the KNN model saw an accuracy boost from 81.26% to 83.49%. These findings underscore the efficacy of optimized sensor combinations in enhancing gas sensor system performance, with the RF model proving particularly effective for classification tasks in this domain.

DETAILED DESCRIPTION OF DIAGRAM
Figure 1: Comprehensive View of the E-Nose System Components and Setup. (a) Image depicting the entire e-nose system setup, (b) Flat cylindrical gas chamber with 8 sensor boards, (c) Block diagram illustrating the operation of the e-nose system.
Figure 2: Visual Representation of Drug Samples. (a) Images displaying drug samples of cannabis (CA), methamphetamine (ME), and tobacco (TB), (b) Illustration of the target drug gas collection process.
Figure 3: Sensor Reactivity to CA, ME, and TB. Mean and standard deviation of normalized sensor responses are depicted for four representative sensors (a-d). Sensor (a) demonstrates cross-sensitivity to all three drugs, while sensor (b) exhibits high selectivity for TB, sensor (c) for CA, and sensor (d) for ME.
Figure 4: Influence of Environmental Variables on E-Nose Target Detection. (a) Principal Component Analysis (PCA) based on three different target data diluted exclusively with air from a specific laboratory (En1). (b) PCA analysis conducted on 100% unmixed normal air samples collected from six different laboratories (En1 ∼ En6). (c) Clustering data for each of the three targets when diluted with 20% air from six different environmental sources. (d) Same data as in (c), with distinct colors representing each environmental laboratory mixed with each target.
Figure 5: Performance of E-Nose System for Drug Detection in Varied Environments. (a) Overall average accuracy of four classifiers (SVM, LDA, RF, KNN). (b) Individual performance across diverse environments. (c) Confusion matrix depicting class-wise performance for different classifiers.
Figure 6. E-nose System Performance with Varying Numbers of Optimized Sensors Using Forward Feature Selection.
Figure 7. Performance Comparison between All 56 Sensors and Optimal 24-Sensor Combination using RF Model. (a) Performance in Different Environments, (b) Average Performance, (c) Confusion Matrix for Optimal 24-Sensor Combination. , Claims:1. Design and Implementation of eNose: An electronic nose for Drug Detection claims that the innovative method proposed in this study addresses the challenge of detecting illegal drugs, such as CA and ME, with exceptional accuracy and speed.
2. Our newly developed e-nose system features 56 sensors, including MOS, EC, NDIR, and PID sensors, strategically selected for their ability to capture distinct characteristics of different drugs.
3. Utilizing forward-feature selection, we identified an optimal combination of 24 sensors, reducing the sensor count while maintaining high detection accuracy.
4. Our method achieved an impressive detection and identification accuracy of 93.03%, significantly reducing the classification error rate from 12.23% to 6.97% across 5166 datasets containing CA, ME, and TB.
5. We validated the effectiveness of our method in diverse environments by evaluating diluted drug gas datasets across six different laboratory conditions using a model trained on data from a separate environment.
6. The proposed method has the potential to evolve into a highly accurate and portable device for swift and precise detection of illegal drugs across various settings, leveraging salient feature-based sensor selection.
7. Our research demonstrates promising results for detecting illegal drugs with outstanding accuracy and speed, laying the foundation for future advancements in law enforcement and public safety.
8. The success of our approach opens avenues for extending drug detection capabilities to other illegal substances, contributing significantly to enhancing national and social security measures.
9. Future research endeavors could further build upon these findings to develop more advanced and effective systems for bolstering security measures and combating drug-related crimes.
10. Overall, our study underscores the transformative potential of innovative sensor technologies in tackling complex societal challenges and improving overall safety and security.