Detection of Fentanyl Analogues in Illicit Drug Samples Using Surface-Enhanced Raman Spectroscopy and Random Forest Classification

Abstract

The ongoing overdose crisis is largely fueled by unregulated opioids; therefore, harm reduction measures like drug checking are forced to adapt to the changing supply. Even at low concentrations, fentanyl and its analogues can induce potent effects, particularly when present together. We found that the three most common fentanyl analogues occur individually and in mixtures of each other, and most often below 5% w/w, under the limit of consistent detection by Fourier transform infrared spectroscopy. Using surface-enhanced Raman scattering and random forest binary classifiers, we detected fentanyl (0.2–29.4% w/w), para-fluorofentanyl (0.2–20.0% w/w), and ortho-methyl fentanyl (0.8–27.4% w/w) in real-world samples with high accuracy (0.93, 0.85, 0.93) and precision (0.96, 0.90, 0.96), an improvement over Fourier transform infrared spectroscopy paired with random forest. However, while the surface-enhanced Raman scattering-based model performed exceptionally well for samples with only one target analyte, mixtures of analogues resulted in the suppression of signals from analytes at lower concentrations, complicating their identification. Overall, this demonstrates the need for the detection and distinction of fentanyl analogues below 5% w/w, the applicability of surface-enhanced Raman scattering and random forest classification for real-world samples involving one fentanyl analogue, and the need for future research in competitive binding of fentanyl analogues in mixtures.

Graphical abstract

This is a visual representation of the abstract.

Keywords

Fentanyl analogues drug checking surfance-enhance Raman spectroscopy SERS Fourier transform infrared spectroscopy FT-IR machine learning random forest classification

Introduction

Illicit drug deaths continue to impact communities worldwide.^1,2 A major contributor to overdoses is the unpredictable and toxic drug supply,³ to the extent that unregulated drug toxicity results in nearly 300 deaths per day in the United States.⁴ Since 2016, over 50 000 opioid-related deaths have been recorded in Canada, with fentanyl and its analogues involved in the majority of these cases.⁵ As a result, fentanyl and its analogues are a point of interest for harm reduction initiatives. Among these analogues, para-fluorofentanyl (pFF) and ortho-methyl fentanyl (oMF) are currently the most common.⁶ The potency of pFF has not been defined, with research finding it weaker, comparable, or stronger than fentanyl,^7–9 but pFF has been determined to increase hypothermic response and antinociception compared to fentanyl.⁹ Additionally, fentanyl and pFF can increase toxic effects relative to an equal dose of either drug alone.¹⁰ Because oMF is thought to have emerged in the drug supply around 2023,¹¹ as a new designer fentanyl analogue, little is known about its potency.⁶ oMF is suggested to induce comparable effects to fentanyl,¹¹ but no in vivo studies have investigated oMF or the co-consumption of oMF with fentanyl or pFF.¹² Given the potency of fentanyl and its analogues and the unknown interactions between multiple analogues in a single sample, identifying and differentiating these fentanyls is an important component of overdose prevention efforts.

Drug checking services have been implemented as a harm reduction strategy to address the unregulated drug supply by providing detailed breakdowns of drug sample components.^13,14 These services typically employ techniques such as immunoassay test strips, Fourier transform infrared (FT-IR) spectroscopy, and mass spectrometry.^15,16 Currently, immunoassay test strips and FT-IR are used widely in North America because both are cost-effective, portable, and simple, making them easily accessible for community drug checking.^16–20 Depending on the brand, fentanyl test strips can be sensitive down to 100–200 ng/mL (0.005–0.01% w/w if prepared using 10 mg sample in 5 mL water),^17,21 while FT-IR can broadly search for unexpected components.¹⁶ However, fentanyl test strips cannot distinguish between fentanyl analogues, and FT-IR is considered to have limited detection beginning at 3–10% w/w for mixtures of this nature.^{16,17,22–24} Mass spectrometry can fill the gap in trace analyte identification, but its cost and complexity can make it unattainable for most point-of-care drug checking sites.²⁵ Therefore, a technique that is affordable and simple but can identify and distinguish trace components in opioid mixtures is needed.

Surface-enhanced Raman scattering (SERS) is a cost-effective and sensitive method that can overcome the limitations of traditional Raman spectroscopy.^15,26–28 While normal Raman spectroscopy has been used for drug checking and fentanyl detection, its sensitivity is insufficient for identifying trace components, and opioids and impurities can produce strong fluorescence signals.^{15,25,28–32} SERS leverages surface plasmon resonance of metal nanoparticles to enhance Raman signals and suppress fluorescence.^26,27,32,33 Previous research using laboratory samples has demonstrated that SERS is capable of identifying fentanyl and its analogues, including at low quantities.^34–42 However, real-world opioid samples are often complex mixtures that include many components such as benzodiazepines and tranquillizers, different cutting agents, impurities, etc.,^6,43,44 Ott et al. successfully employed electrochemical-SERS for the detection of fentanyl or fentanyl analogues in authentic illicit drug mixtures with an accuracy of 88% for samples averaging 6–9% w/w, achieving improved identification over colour test methods.⁴⁵ However, the results do not distinguish between fentanyl analogues. Erkok et al. have demonstrated the possibility of distinguishing fentanyl analogues in counterfeit tablets through peak analysis,⁴² but there remains a lack of research using SERS for the distinction of fentanyl and analogues in real-world samples that includes chemometric approaches.

Manual interpretation of SERS spectra is challenging, especially for samples with complex matrices, such as real-world opioid mixtures.^46,47 As a result, multivariate analysis is often used with SERS.^40,47,48 Supervised machine learning methods, such as random forest (RF), support vector machine (SVM), and k-nearest neighbours (k-NN), have been effective for spectral analysis of drug samples due to their ability to handle high-complexity data and identify trace components in contrast to unsupervised methods.^{39,47,49–54} SERS combined with RF has successfully detected trace components in real-world opioid mixtures⁴⁹ and outperformed SVM and k-NN for identifying trace adulterants.⁵⁰ RF classification is an ensemble learning technique that fits multiple decision tree classifiers on subgroups of the training data, then combines the results using majority voting.⁵⁵ By employing many decision trees, RF improves accuracy and reduces the chances of over-fitting.⁵⁵ RF could assist SERS interpretation of fentanyl and analogues while improving sensitivity, specificity, and ease of use.

In this work, we apply binary RF classifiers to SERS spectra of real-world opioid samples to detect and distinguish fentanyl, pFF, and oMF, both individually and in mixtures of each other. Our findings indicate that SERS outperforms FT-IR both in terms of precision and accuracy. This demonstrates that SERS is a viable and practical method for point-of-care analysis of complex opioid samples.

Methods

Sample Selection

The drug samples used in this study were collected from Substance, a drug checking site in Victoria, British Columbia, Canada.⁵⁶ As part of this service, samples were analyzed with benzodiazepine immunoassay test strips (Rapid Response, BTNX), paper spray mass spectrometry, and FT-IR for confirmatory testing. All expected fentanyl samples analyzed and quantified at Substance from January 2023 to August 2025, which contained fentanyl, pFF, or oMF ( $n = 7953$ ), were used for the drug trend analysis.

Samples containing fentanyl, pFF, or oMF below 30% w/w with at least 1.5 mg remaining after confirmatory tests were selected weekly by Substance staff for additional SERS testing ( $n = 321$ ). The number of fentanyl, pFF, and oMF samples could not be effectively balanced because their prevalence and co-occurrence vary in the unregulated supply.^43,44 The components in the samples chosen for SERS testing are shown in Table 1. Samples with positive benzodiazepine test strips that could not be identified by FT-IR or mass spectrometry were labelled as “benzodiazepine (unknown).” Samples with carbohydrate or fentanyl precursor bands in their FT-IR spectra that could not be distinguished were labelled as “carbohydrate (unknown)” or “fentanyl precursor (unknown),” respectively.

Table 1.

Drug components in training and test set samples, identified by benzodiazepine immunoassay test strips, paper spray mass spectrometry, and FT-IR with manual library matching.

	Count
	Training Set	Test Set
Compound	( $n = 203$ )	( $n = 88$ )
Fentanyl	124	54
pFF	101	44
oMF	61	26
Caffeine	189	84
Erythritol	126	63
Bromazolam	78	33
Xylazine	55	19
Carfentanil	47	24
Desalkylgidazepam	35	17
Xylitol	29	15
Medetomidine	18	7
Benzodiazepine (unknown)	17	3
Mannitol	9	6
Heroin	7	1
Acetylmorphine	5	1
N-phenyl propanamide	4	0
Acetylcodeine	3	1
Methamphetamine	3	1
Dimethyl sulfone	2	1
Etizolam	2	0
Phenacetin	2	1
Benzocaine	1	0
Carbohydrate (unknown)	1	1
Cocaine	1	0
Fentanyl Precursor (unknown)	1	0
Flualprazolam	1	1
Flubromazepam	1	1
Flubromazolam	1	1
Morphine	1	1
N-pyrrolidino etonitazene	1	0

High-noise SERS spectra were removed using a fast Fourier transform and the squared amplitude at each frequency to assign high-frequency scores. Spectra with scores higher than typical, determined using the third quartile of score distributions plus two times the interquartile range, were excluded ( $n = 30$ ). The remaining samples ( $n = 291$ ) were separated into a training set ( $n = 203$ ) and a test set ( $n = 88$ ), with 70% of the samples used for model training and 30% for model evaluation. The same sets were used for both the SERS and FT-IR models to ensure consistency. Within the test set samples, $n = 54$ samples contained fentanyl (range 0.2–29.4% w/w, median 4.3% w/w), $n = 44$ samples contained pFF (range 0.2–20.0% w/w, median 3.8% w/w) and $n = 26$ samples contained oMF (range 0.8–27.4% w/w, median 4.3% w/w).

Measurements

SERS measurements were collected using a portable Raman spectrometer (Resolve, Agilent Technologies, Santa Clara, USA) with an 830 nm excitation source. Details of the method and instrumentation are described previously.^36,49,50,57 Briefly, powdered samples were suspended in deionized water (18.2 M $Ω \cdot$ cm, Barnstead Nanopure water, Thermo Fisher Scientific) and vortexed to make a sample solution of 2.14 mg/mL. 70 $μ$ L of the sample solution was added to 1.42 mL of a 50 nm colloidal gold nanoparticle solution (EM.GC50/7, BBI Solutions, UK), which was then vortexed for 30 s. 10 $μ$ L of 1 M MgSO $_{4}$ solution made from magnesium sulfate anhydrous certified powder (MgSO $_{4}$ , Fisher Chemical) in deionized water was added as an aggregating agent. The final solution was vortexed for approximately 10 s, and the spectra were collected through a glass vial using 100–475 mW variable power. The signal was recorded between 200 and 2000 cm⁻¹, with 2 s integration time, 10 averages, and a spectral resolution of 15 cm⁻¹. The instrument software automated baseline, fluorescence, and cosmic ray corrections.

FT-IR measurements of powdered samples were collected using a portable ATR-FT-IR spectrometer (Agilent 4500a, USA) with a 45 $\circ$ diamond attenuated total reflectance element. Infrared (IR) spectra were recorded between 650 and 4000 cm⁻¹ with 32 averages and an effective resolution of 4 cm⁻¹. IR spectra were first used for the confirmatory identification of bulk components by manually matching with library spectra as part of intake at Substance. IR spectra of the training and test set samples were also analyzed by the RF model for the purpose of this study.

All samples used in this study were analyzed using a TSQ Fortis triple quadrupole mass spectrometer equipped with a VeriSpray paper spray ion source (Thermo Fisher Scientific, USA). Details of the method and data analysis are previously described.^58–62 The current calibration curves for fentanyl, pFF, and oMF cannot accurately quantify above 50% w/w, so samples with concentrations above this limit are set to 50% w/w.

Data Analysis

SERS spectra were corrected for common baseline subtraction artifacts using a median filter with a window size of 10, then each spectrum was min-max normalized. Outlier detection on training set samples was performed separately for SERS and IR data. In the first step of outlier detection, training set spectra were standardized by subtracting the mean and scaling to unit variance. Local outlier factor was used for outlier detection on the first two principal components of each spectrum. The local density around a data point was estimated based on its 20 nearest neighbours, and samples with the lowest 2.5% of density scores were identified as outliers and removed. SERS and IR outliers primarily included spectra with anomalies induced by uncommon components like mannitol, acetylmorphine, and medetomidine. After outlier removal, the final training sets ( $n = 197$ ) and test sets ( $n = 88$ ) were used for supervised machine learning.

Separate RF binary classifiers were developed for each of fentanyl, pFF, and oMF using SERS or IR data.⁶³ The spectral derivative and four RF hyperparameters were tuned: the number of decision trees, the maximum depth, the minimum samples per leaf, and the minimum samples per split. Hyperparameters were optimized using a five-fold grid search with cross-validation, which evaluates every combination of hyperparameter values for multiple splits of the data. The combination that achieved the highest F1 score for training set predictions was selected, and the final models were trained using these optimal hyperparameters. The grid search values, optimized hyperparameters, and performance metrics for the training of SERS- and IR-based models are shown in Table S1. The decision threshold for test set predictions was set to 0.5.

Results and Discussion

Trends in the Illicit Drug Supply

We analyzed trends in co-occurrence and w/w percentage distributions of expected-opioid samples ( $n = 7953$ ) with fentanyl ( $n = 6597$ ), pFF ( $n = 3565$ ), and oMF ( $n = 581$ ) to highlight the need to distinguish between fentanyl and its analogues, particularly at concentrations below 5% w/w. Although other analogues also exist, they were less prevalent: carfentanil ( $n = 135$ ), chloroisobutyryl fentanyl ( $n = 6$ ), acetylfentanyl ( $n = 2$ ). As shown in Figure 1, w/w percentage distributions of fentanyl, pFF, and oMF were right-skewed, with analyte quantities between 0.1–5% w/w being the most common. FT-IR has a limited ability to detect compounds at these levels, especially in complex mixtures, so many of these samples would likely go undetected or misidentified.^16,17,23 However, these low concentrations are still significant as a result of the high potency of these synthetic opioids.^7–9,11

Figure 1.

Frequency of fentanyl, pFF, and oMF in expected-opioid samples tested by Substance from January 2023 to August 2025 ( $n = 7953$ ) broken down by w/w percentage with molecular structures: (a) fentanyl ( $n = 6597$ ), (b) pFF ( $n = 3565$ ), (c) oMF ( $n = 581$ ).

Figure S1 shows the distribution of w/w percentages for single-fentanyl samples ( $n = 5315$ ) with fentanyl ( $n = 4016$ ), pFF ( $n = 1173$ ), or oMF ( $n = 126$ ), respectively. These are samples that contain only one of fentanyl, pFF, or oMF. However, these samples may include other analogues, cutting agents, benzodiazepines, or other drugs. Overall, single-fentanyl samples were more likely to contain fentanyl, pFF, or oMF above 5% w/w. The distribution of fentanyl concentrations centered around 10% w/w, indicating fentanyl occurs most frequently at concentrations between 5–15% w/w in single-fentanyl samples. Similar trends have previously been observed for fentanyl.^43,64 The most common concentration ranges for pFF and oMF were 15–20% w/w and 5–10%, respectively. The co-consumption of fentanyls has been shown to produce synergistic effects,¹⁰ so the increased concentrations in these single-fentanyl samples may reflect the need for higher doses to achieve the same effects in the absence of other fentanyls. The frequency of higher concentrations also suggests that samples containing only one of these fentanyls may be more likely to be detected using FT-IR alone.

In contrast, multi-fentanyl samples ( $n = 2638$ ), which contain mixtures of fentanyl, pFF, or oMF in addition to other components, more commonly contained these analogues in lower quantities. Figure S2 shows the w/w co-occurrences for binary mixtures of fentanyl and pFF ( $n = 2183$ ), fentanyl and oMF ( $n = 246$ ), and pFF and oMF ( $n = 57$ ), respectively. Figure S3 shows the co-occurrence of analyte quantities in ternary mixtures of analytes ( $n = 152$ ). Besides oMF in binary mixtures with pFF, in all other types of mixtures, each analyte was most frequently below 5% w/w. The co-consumption of fentanyl and pFF produces greater effects than either one alone,¹⁰ but interaction between oMF and fentanyl or pFF have not been studied. The combination of small amounts of these opioids may be sufficient to produce strong effects while remaining below the detection limit of FT-IR. Beyond detection, an additional challenge at these low quantities is differentiating the analogues. An immunoassay test strip could detect that a fentanyl is present, but could not determine which or how many analogues are in one sample.²²

In general, low concentrations of the three analogues were observed most frequently, which highlights the need for more sensitive detection methods than FT-IR. In addition, the high prevalence of multiple, low-concentration fentanyl analogues in a single sample further underscores the need for sensitive techniques capable of differentiating analogues in mixtures.

Feature Importance

The 10 spectral features that the SERS-based models identified as most important for the classification of each fentanyl are shown in Figure 2. The highlighted features in the fentanyl spectrum shown in Figure 2a clustered around the 745 cm⁻¹ peak, which corresponds to CH $_{3}$ twisting and C–C–C bending modes of the fentanyl amide group and CH $_{2}$ rocking of the piperidine ring (molecular structures displayed in Figure 1).^39,65 For pFF and oMF, the regions with the majority of important features were bands related to vibrational modes of the substituted aniline ring.^39,65 Characteristic features of the pFF spectrum were clustered around a peak at 828 cm⁻¹ that is most notably due to the C–H out-of-plane bending, specifically of the aniline ring.^65,66 As for oMF, the region with the majority of important features was the shoulder of a peak at 1033 cm⁻¹. Because oMF is a new designer analogue, there is no vibrational band assignment in the literature. However, literature peak assignment for fentanyl has identified C=C stretches and C–H in-plane bending of the aromatic rings as producing a peak at a similar wavenumber.^65,67 The methyl substituent on oMF could affect the electron distribution around the aniline ring, leading to shifts of these vibrational modes and uniquely affecting the peak at 1033 cm⁻¹. The identification of unique spectral features shows promise in differentiating fentanyl, pFF, and oMF in complex samples.

Figure 2.

Mean SERS spectra of positive and negative samples, and a single component spectrum with the top 10 most important wavenumbers highlighted by the model: (a) fentanyl, (b) pFF, (c) oMF. Peaks with the majority of highlighted features are labelled. “Single component” refers to samples tested at Substance that were not found to contain any other compounds.

FT-IR spectra were also highlighted with their top 10 features as shown in Figure S4. The highlighted wavenumbers for the fentanyl spectrum were distributed throughout the fingerprint region and the 2500–3000 cm⁻¹ region, none of which identified a single characteristic peak. The majority (six) pFF features highlighted a strong peak at 1506 cm⁻¹, which falls in an area mostly free of bands from cutting agents or other common sample components, making this an ideal peak for classification.^68,69 This peak has previously been identified as relevant for the characterization of pFF using an RF algorithm.⁵³ Half of the oMF features highlighted wavenumbers in the fingerprint region, with the other half highlighting bands in the 2500–3500 region cm⁻¹, none of which identified a clear characteristic peak.⁶⁸ Other than pFF, the majority of important features could not be attributed to one strong peak, potentially complicating the classification of fentanyl and oMF.

Model Performance

F1 scores were used to assess the quality of model training. An F1 score takes the product of recall (sensitivity) and precision multiplied by two and divides it by their sum. In hyperparameter tuning, F1 scores are calculated based on the training data, which are displayed in Table S1. The SERS-based models returned the highest F1 scores of 0.84, 0.94, and 0.80, respectively, while the IR-based scores were 0.78, 0.84, and 0.58. This is likely because the SERS models were provided with more spectral features to leverage.⁷⁰ This is particularly apparent for oMF, which lacked clear distinguishing features in its IR spectrum, and returned a low F1 score of 0.58. Aside from the F1 score of the oMF-IR model, the SERS- and IR-based models yielded relatively high scores otherwise, indicating suitable training for the other target analytes.

SERS also demonstrated improved performance over IR during model validation. Precision-recall curves for test set predictions are shown in Figure S5 and display results across a range of thresholds. Average precision (AP) values from precision-recall curves were used to assess prediction quality (Figure S5). The SERS-based model returned the highest AP values of 0.99, 0.95, and 0.95 for fentanyl, pFF, and oMF, respectively, while IR-based AP scores were 0.88, 0.94, and 0.67. Although the SERS-based models returned improved AP scores, the AP score of the pFF-IR model was also high. This is likely because of the strong, characteristic pFF peak at 1506 cm⁻¹ in the IR spectrum and higher quality training of the pFF-IR model. In contrast, the low AP score of the oMF-IR model may be attributed to its lack of strong characteristic features and the poorly trained model. Overall, these results support the enhanced detection of fentanyl, pFF, and oMF using SERS compared to IR, and the ability of machine learning to predict the presence of these fentanyls using spectral data with clear characteristic features.

The sensitivity, specificity, accuracy, precision, F1 score, and Matthew’s correlation coefficient (MCC) of test set predictions are shown in Table II. The SERS-based models outperformed the IR models in sensitivity, accuracy, F1, and MCC. Although the pFF-IR returned slightly higher specificity and precision than the pFF-SERS model, the specificity and precision of the SERS-based models were high for all components (specificity: 0.94, 0.91, 0.98; precision: 0.96, 0.90, 0.96 for fentanyl, pFF, and oMF, respectively). Additionally, in drug checking contexts, the sensitivity of a technique is more significant than the specificity, because false negative results have more serious implications.⁷¹ Poorer sensitivity of IR models is expected, given that the median w/w percentages of target components were between 3.8–4.3% in the test set.^16,53 The SERS-based models were able to detect target components below the limit of consistent detection of FT-IR, as shown in Table S2. Considering the complexity of the samples being evaluated,^43,44 the ability for the SERS models to detect fentanyl and pFF as low as 0.2% w/w (200 ng/mL) is considerable (Table S2). oMF was also successfully detected at its lowest concentrations in the dataset (0.8% w/w (800 ng/mL). These results also suggest that SERS with RF can distinguish between fentanyl analogues despite structural similarities and detect them below the detection limit of typical vibrational technologies such as FT-IR and Raman.^16,17,25,29

Table II.

Summary of prediction metrics for fentanyl, pFF, and oMF from SERS- and IR-based models for all test set samples, the single-fentanyl subset, and the multi-fentanyl subset of binary and ternary mixtures.

Test set	Data	Analyte	Sensitivity	Specificity	Accuracy	Precision	F1	MCC
All ( $n = 88$ )	SERS	Fentanyl	0.93	0.94	0.93	0.96	0.94	0.86
		pFF	0.80	0.91	0.85	0.90	0.84	0.71
		oMF	0.81	0.98	0.93	0.96	0.88	0.83
	IR	Fentanyl	0.83	0.50	0.71	0.73	0.78	0.36
		pFF	0.68	0.98	0.83	0.97	0.80	0.69
		oMF	0.46	0.89	0.76	0.63	0.53	0.39
Single-fentanyl subset ( $n = 33$ )	SERS	Fentanyl	1.00	0.95	0.97	0.93	0.97	0.94
		pFF	1.00	1.00	1.00	1.00	1.00	1.00
		oMF	1.00	0.97	0.97	0.80	0.89	0.89
	IR	Fentanyl	1.00	0.58	0.76	0.64	0.78	0.61
		pFF	0.93	0.94	0.94	0.93	0.93	0.88
		oMF	1.00	0.93	0.94	0.67	0.80	0.79
Multi-fentanyl subset ( $n = 42$ )	SERS	Fentanyl	0.90	1.00	0.91	1.00	0.95	0.55
		pFF	0.69	0.85	0.74	0.91	0.78	0.50
		oMF	0.77	1.00	0.88	1.00	0.87	0.79
	IR	Fentanyl	0.78	1.00	0.79	1.00	0.87	0.38
		pFF	0.55	1.00	0.69	1.00	0.71	0.53
		oMF	0.36	0.90	0.62	0.80	0.50	0.31

Although the sensitivity of target component detection was improved by using SERS data (SERS: 93%, 80%, 81%; IR: 83%, 68%, 46%), false negative results were investigated to identify methods of further enhancing the sensitivity of the SERS-based models. Single-fentanyl samples (samples with only one of fentanyl, pFF, and oMF, plus any other number of cutting agents, other analogues, or psychoactive substances) appeared to be outperforming binary or ternary multi-fentanyl mixtures of fentanyl, pFF, and oMF (Table S2). Therefore, single-fentanyl samples ( $n = 33$ ) and mixtures ( $n = 42$ ) were evaluated separately.

For single-fentanyl samples, SERS outperformed IR in all metrics, as shown in Table II. The SERS-based models were 100% sensitive for all target analytes, higher than in the overall test set (93%, 80%, 81%) (Table II). A similar trend was observed between the sensitivity of IR model predictions of single-fentanyl samples (100%, 93%, 100%) and the overall test set (83%, 68%, 46%). Although IR models also demonstrated high sensitivity, the specificity of the fentanyl-IR model was low (58%), indicating overprediction. The pFF-IR and oMF-IR models demonstrated both high sensitivity (93%, 100%) and high specificity (94%, 93%), suggesting an enhancement of pFF and oMF identification, but not to the extent of the pFF-SERS and oMF-SERS models. However, the pFF-IR model correctly identified pFF at 1.9% w/w, below the limit of consistent detection by IR with library matching.^16,17 Enhanced detection by SERS and IR models for single-fentanyl samples suggests that these fentanyls are readily detected in the absence of other target analogues and can be reliably detected at lower concentrations, especially using SERS data. Additionally, these results demonstrate that machine learning can improve analyte identification using SERS or IR data, and they support the expectation that SERS offers higher sensitivity for low-concentration fentanyls than IR.

The sensitivity of all models decreased for binary and ternary multi-fentanyl mixtures, as shown in Table II. Predictions by SERS-based models continued to be more sensitive and accurate with higher F1 and MCC scores than IR models. While specificity and precision were high for both SERS and IR models, comparatively lower sensitivities (SERS: 90%, 69%, 77%; IR: 78%, 55%, 36%) and accuracies (SERS: 91%, 74%, 88%; IR: 79%, 69%, 62%) highlighted a discrepancy between single-fentanyl and multi-fentanyl sample predictions. While SERS-based models outperformed IR models, neither approach was able to consistently detect all fentanyls in a mixture. For SERS, false negative results were more frequent for target fentanyls at lower w/w percentage relative to another (Table S2), a trend previously identified for fentanyl.^36,42,57 The SERS models performed best when fentanyls were present at similar concentrations, which is also when identifying each analyte is most relevant, as fentanyl and pFF are known to exert synergistic effects at equal doses.¹⁰ In contrast, the degree of synergism is unknown when these fentanyls are at unequal ratios.⁷² The RF-SERS approach would be valuable in drug checking contexts for characterizing mixtures of fentanyl, pFF, and oMF at comparable levels, or identifying the predominant fentanyl.

The decreased sensitivity of SERS model predictions in mixtures of fentanyl, pFF, and oMF, compared with single-fentanyl samples, could be attributed to the complexity of sample matrices and properties of analyte–nanoparticle interactions. Spectral interferences from other sample components can obscure characteristic peaks, an issue also observed in IR-based analysis, which contributes to the limited detection between 3–10% w/w for FT-IR.^15–17 In multi-fentanyl samples, the increased number of components could lead to overlapping spectral peaks, further reducing the distinctiveness of analyte features. Also, preferential binding to nanoparticles would influence the detection of components over others. Higher-concentration analytes can outcompete trace analytes for binding sites on the nanoparticle surface,^36,57 and differences in the binding affinities of fentanyl, pFF, and oMF may further influence these interactions.^15,73,74 Although understanding these interactions could inform the functionalization of nanoparticles or analytical protocols, a single, reliable substrate and method is preferred for implementing SERS in drug checking settings.¹⁵ Instead, the spectra generated by the colloidal gold nanoparticle method were analyzed to investigate the observed sensitivity differences between multi-fentanyl and single-fentanyl samples.

SERS spectra were overlaid at characteristic regions, 745 cm⁻¹ for fentanyl, 828 cm⁻¹ for pFF, and 1033 cm⁻¹ for oMF (Figure 3). For the fentanyl peak at 745 cm⁻¹, the variability in the surrounding peaks contributed to the difficulty of discerning a clear trend. The intensity of the 745 cm⁻¹ peak generally decreased, but the shape varied with increasing quantities of pFF and oMF. The pFF peak at 828 cm⁻¹ diminished in intensity and shifted to higher wavenumbers as the combined fentanyl and oMF concentration increased. Similarly, the oMF peak at 1033 cm⁻¹ decreased in intensity and shifted to lower wavenumbers with increasing concentrations of other fentanyls. The pFF and oMF peaks were not significantly obstructed by other bands or altered in shape; rather, their intensities and shifts were affected. This effect could explain the diminished sensitivity of the SERS-based model predictions in multi-fentanyl mixtures relative to single-fentanyl samples. With these observations in mind, future machine learning algorithms may be designed to handle variations in characteristic features to improve the detection of fentanyl, pFF, and oMF, specifically in multi-fentanyl samples.

Figure 3.

(a) SERS spectra of all samples with fentanyl with focus on its characteristic peak at 745 cm⁻¹. Colours correspond to the summed w/w percentage of pFF and oMF in each sample. (b) SERS spectra of all samples with pFF with focus on its characteristic peak at 828 cm⁻¹. Colours correspond to the summed w/w percentage of fentanyl and oMF in each sample. (c) SERS spectra of all samples with oMF with focus on its characteristic peak at 1033 cm⁻¹. Colours correspond to the summed w/w percentage of fentanyl and pFF in each sample.

The accessibility of SERS, in terms of cost, portability, and ease-of-use, makes it a strong candidate to fill the gap in trace detection at drug checking sites that lack mass spectrometers.^15,26–28 We have shown that SERS is a more sensitive technique than FT-IR, an instrument widely used across North America.²⁰ Together, FT-IR and SERS can determine the majority of components within an illicit opioid mixture: FT-IR can provide bulk component information,¹⁶ and SERS can provide trace detection. SERS performs particularly well in detecting fentanyl, pFF, or oMF at trace levels when present alone, enhancing detection relative to FT-IR and providing differentiation of fentanyl analogues, an option not offered by fentanyl test strips.²² Given the high potency of fentanyl and its analogues,^7–9,11 these trace quantities remain relevant in point-of-care contexts. Furthermore, we demonstrated that SERS can detect fentanyl, pFF, and oMF in multi-fentanyl mixtures at similar concentrations, an important advantage because fentanyl and pFF are known to exert synergistic effects at equal concentration ratios.¹⁰ Although SERS is less effective at identifying all fentanyls in multi-fentanyl mixtures with unequal concentration ratios, detecting the primary fentanyl analogue provides valuable information. Overall, a SERS–machine learning approach would enhance the current capabilities of drug checking sites that currently rely on FT-IR and test strips.

Conclusion

We analyzed expected-opioid samples from the local illicit drug supply using paper spray mass spectrometry and determined that fentanyl, pFF, and oMF can appear alone or in mixtures of each other, and are most often below 5% w/w. We achieved trace detection of fentanyl, pFF, and oMF in illicit opioid samples using three binary random forest classifiers trained on SERS spectral data. Overall, the SERS performance exceeded that of FT-IR. Characteristic spectral features allowed for effective detection in samples that contained only one of the molecules of interest. However, when the three targets were in the same sample, the performance of both the SERS-based model and the IR-based model declined. In multi-fentanyl mixtures, the characteristic spectral features of a target analyte disappeared if another fentanyl was present at a higher concentration than the target analyte. Without the characteristic features, the model could not consistently detect the compound of interest. This illustrates the capability of SERS for detecting and distinguishing fentanyl and fentanyl analogues at trace quantities when alone in the sample. This also highlights the need for SERS-machine learning research to differentiate fentanyl and analogues at different ratios in illicit opioid mixtures. Future work could include exploring machine learning approaches like data fusion and deep learning, and an investigation into the binding affinities of fentanyl and analogues.

Supplemental Material

sj-pdf-1-asp-10.1177_00037028261457152 - Supplemental material for Detection of Fentanyl Analogues in Illicit Drug Samples Using Surface-Enhanced Raman Spectroscopy and Random Forest Classification

Supplemental material, sj-pdf-1-asp-10.1177_00037028261457152 for Detection of Fentanyl Analogues in Illicit Drug Samples Using Surface-Enhanced Raman Spectroscopy and Random Forest Classification by Ella Newman, Rebecca R. Martens, Lea Gozdzialski, Bruce Wallace, Chris G. Gill and Dennis K. Hore in Applied Spectroscopy

Footnotes

Acknowledgements

We would like to thank people who use drug checking for providing the samples and perspectives essential for this project. We are also grateful for our colleagues at Substance who supplied the confirmatory testing results by collecting and analyzing mass spectra, FT-IR spectra, and immunoassay test strip data. Derek Robinson, Collin Kielty, and Josh Jai offered valuable advice around drug trend reporting, confirmatory testing results, and machine learning methods.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

This project received funds from the Government of Canada’s New Frontiers Research Fund (NFRF, NFRFE-2022-00886), with additional support from the British Columbia Ministry of Health, Island Health Authority, and the Vancouver Foundation (F0120-5607). EN is grateful to NSERC for a CGS-M fellowship. High performance computing support and server resource allocation was provided by the University of Victoria, Westgrid, and the Digital Research Alliance of Canada.

ORCID iDs

Ella Newman

Rebecca R. Martens

Lea Gozdzialski

Bruce Wallace

Chris G. Gill

Dennis K. Hore

Supplemental Material

All supplemental material mentioned in the text accompanies this paper online.

References

United Nations Office on Drugs and Crime (UNODC). “World Drug Report 2025”. https://www.unodc.org/unodc/en/data-and-analysis/world-drug-report-2025.html [accessed 12 June 2026].

Silva

L.T.

Modesto

A.C.F.

Amaral

R.G.

Lopes

F.M.

“Hospitalizations and Deaths Related to Adverse Drug Events Worldwide: Systematic Review of Studies with National Coverage“. Europ. J. Clin. Pharm. 2022. 78: 435-466. 10.1007/s00228-021-03238-2

Knill

Tobias

Matthews

(2022) “A Report on British Columbia’s Unregulated Drug Supply: Drug Checking Trends Across British Colubmia, January–December 2021”. https://drugcheckingbc.ca/wp-content/uploads/sites/4/2022/07/BCCSU_BCs_Annual_Drug_Checking_Report2022.pdf [accessed 12 June 2026].

U.S. Department of Health and Human Service, National Institutes of Health, National Institute on Drug Abuse. “Drug Overdose Deaths: Facts and Figures”. https://nida.nih.gov/research-topics/trends-statistics/overdose-death-rates[accessed 12 June 2026].

Public Health Agency of Canada. “Opioid- and Stimulant-Related Harms in Canada”. https://health-infobase.canada.ca/substance-related-harms/opioids-stimulants/#a9 [accessed 12 June 2026].

St-Denis

. “Fentanyl Analogues in Canada: Trends and Implications”. https://www.ccsa.ca/sites/default/files/2025-03/CCENDU-Newsletter-Issue-4-en.pdf [accessed 12 June 2026].

Suzuki

El-Haddad

“A Review: Fentanyl and Non-Pharmaceutical Fentanyls“. Drug Alcohol Depend. 2017. 171: 107-116. 10.1016/j.drugalcdep.2016.11.033

Varshneya

N.B.

Hassanien

S.H.

Holt

M.C.

Stevens

D.L.

Layle

N.K.

, et al. “Fentanyl Analog Structure-Activity Relationships Demonstrate Determinants of Diverging Potencies for Antinociception and Respiratory Depression“. Pharmacol. Biochem. Behav. 2023. 226: 173572. 10.1016/j.pbb.2023.173572

Canfield

J.R.

Sprague

J.E.

“In Vivo Pharmacokinetic, Pharmacodynamic and Brain Concentration Comparison of Fentanyl and para-Fluorofentanyl in Rats“. Arch. Toxicol. 2025. 99(1): 287-297. 10.1007/s00204-024-03887-z

10.

Canfield

J.R.

Fry

K.K.

Sprague

J.E.

“Enhanced Pharmacokinetic and Pharmacodynamic Effects of Combining p-Fluorofentanyl and Fentanyl“. Drug Alcohol Depend. 2025. 272: 112710. 10.1016/j.drugalcdep.2025.112710

11.

Stang

B.N.

Walton

S.E.

Merrette

Shapiro

Papsun

D.M.

, et al. “ortho-Methylfentanyl Proliferating Across North America As Newest Synthetic Opioid Identified in Fatal Drug Overdoses”. https://www.cfsre.org/images/content/reports/public_alerts/Public_Alert_ortho-Methylfentanyl_120224.pdf [accessed 12 June 2026].

12.

Hassanien

S.H.

Bassman

J.R.

Naccarato

C.M.P.

Twarozynski

J.J.

Traynor

J.R.

, et al. “In Vitro Pharmacology of Fentanyl Analogs at the Human Mu Opioid Receptor and Their Spectroscopic Analysis“. Drug Test. Anal. 2020. 12(8): 1212-1221. 10.1002/dta.2822

13.

Laing

M.K.

Tupper

K.W.

Fairbairn

“Drug Checking as a Potential Strategic Overdose Response in the Fentanyl Era“. Int. J. Drug Policy. 2018. 62: 59-66. 10.1016/j.drugpo.2018.10.001.

14.

Brunt.

“Drug Checking as a Harm Reduction Tool for Recreational Drug Users: Opportunities and Challenges”. https://www.drugsandalcohol.ie/28065/1/EuropeanResponsesGuide2017_BackgroundPaper-Drug-checking-harm-reduction.pdf [accessed 12 June 2026].

15.

Gozdzialski

Wallace

Hore

“Point-of-Care Community Drug Checking Technologies: An Insider Look at the Scientific Principles and Practical Considerations“. Harm Reduct. J. 2023. 20: 39. 10.1186/s12954-023-00764-3

16.

McCrae

Tobias

Grant

Lysyshyn

Laing

, et al. “Assessing the Limit of Detection of Fourier-Transform Infrared Spectroscopy and Immunoassay Strips for Fentanyl in a Real-World Setting“. Drug Alcohol Rev. 2020. 39(1): 98-102. 10.1111/dar.13004

17.

Green

T.C.

Park

J.N.

Gilbert

McKenzie

Struth

, et al. “An Assessment of the Limits of Detection, Sensitivity and Specificity of Three Devices for Public Health-Based Drug Checking of Fentanyl in Street-Acquired Samples“. Int. J. Drug Policy. 2020. 77: 102661. 10.1016/j.drugpo.2020.102661

18.

Tupper

K.W.

McCrae

Garber

Lysyshyn

Wood

“Initial Results of a Drug Checking Pilot Program to Detect Fentanyl Adulteration in a Canadian Setting“. Drug Alcohol Depend. 2018. 190: 242-245. 10.1016/j.drugalcdep.2018.06.020

19.

Peiper

N.C.

Clarke

S.D.

Vincent

L.B.

Ciccarone

Kral

A.H.

, et al. “Fentanyl Test Strips as an Opioid Overdose Prevention Strategy: Findings From a Syringe Services Program in the Southeastern United States“. Int. J. Drug Policy. 2019. 63: 122-128. 10.1016/j.drugpo.2018.08.007

20.

Park

J.N.

Tardif

Thompson

Rosen

J.G.

Arrendondo

, et al. “A Survey of North American Drug Checking Services Operating in 2022“. Int. J. Drug Policy. 2023. 121: 104206. 10.1016/j.drugpo.2023.104206

21.

Rodriguez-Cruz

S.E.

“Evaluating the Sensitivity, Stability, and Cross-Reactivity of Commercial Fentanyl Immunoassay Test Strips“. J. Forensic Sci. 2023. 68(5): 1555-1569. 10.1111/1556-4029.15332

22.

Crepeault

Socias

M.E.

Tobias

Lysyshyn

Custance

, et al. “Examining Fentanyl and Its Analogues in the Unregulated Drug Supply of British Columbia, Canada Using Drug Checking Technologies“. Drug Alcohol Rev. 2022. 42(3): 538–543. 10.1111/dar.13580

23.

Bunaciu

A.A.

Aboul-Enein

H.Y.

Fleschin

“Application of Fourier Transform Infrared Spectrophotometry in Pharmaceutical Drugs Analysis“. Appl. Spectrosc. Rev. 2010. 45(3): 206-219. 10.1080/00387011003601044

24.

Tobias

Lysyshyn

Laing

Nosova

, et al. “Detecting Fentanyl Using Point-of-Care Drug Checking Technologies: A Validation Study“. Drug Alcohol Depend. 2020. 212: 108006. 10.1016/j.drugalcdep.2020.108006

25.

Harper

Powell

Pijl

E.M.

“An Overview of Forensic Drug Testing Methods and Their Suitability for Harm Reduction Point-of-Care Services“. Harm Reduct. J. 2017. 14: 52. 10.1186/s12954-017-0179-5

26.

Cailletauda

De Bleyea

Dumonta

Sacréa

P.-Y.

Netchacovitcha

, et al. “Critical Review of Surface-Enhanced Raman Spectroscopy Applications in the Pharmaceutical Field“. J. Pharm. Biomed. Anal. 2018. 147: 458-472. 10.1016/j.jpba.2017.06.056

27.

Fan

Andrade

G.F.S.

Brolo

A.G.

“A Review on Recent Advances in the Applications of Surface-Enhanced Raman Scattering in Analytical Chemistry“. Anal. Chim. Acta. 2020. 1097: 1-29. 10.1016/j.aca.2019.11.049

28.

Vankeirsbilck

Vercauteren

Baeyens

der Weken

G.V.

Verpoort

, et al. “Applications of Raman Spectroscopy in Pharmaceutical Analysis“. Trends Anal. Chem. 2002. 21(12): 869-877. 10.1016/S0165-9936(02)01208-6

29.

Gozdzialski

Ramsay

Larnder

Wallace

Hore

D.K.

“Fentanyl Detection and Quantification Using Portable Raman Spectroscopy in Community Drug Checking“. J. Raman Spectrosc. 2021. 52(7): 1308-1316. 10.1002/jrs.6133

30.

Ryder

A.G.

“Surface Enhanced Raman Scattering for Narcotic Detection and Applications to Chemical Biology“. Curr. Opin. Chem. Biol. 2005. 9(5): 489-493. 10.1016/j.cbpa.2005.07.001

31.

Gerace

Seganti

Luciano

Lombardo

Di Corcia

, et al. “On-Site Identification of Psychoactive Drugs by Portable Raman Spectroscopy During Drug-Checking Service in Electronic Music Events“. Drug Alcohol Rev. 2019. 38(1): 50-56. 10.1111/dar.12887

32.

Kimani

M.M.

Smith

S.W.

Lanzarotta

Brueggemeyer

J.L.

Batson

J.S.

“Screening Suspect Pharmaceuticals for Illicit Designer Benzodiazepines Using Raman, SERS, and FT-IR Prior to Comprehensive Analysis Using LC-MS“. Forensic Sci. Int. 2022. 338: 111390. 10.1016/j.forsciint.2022.111390

33.

Pérez-Jiménez

A.I.

Lyu

Liu

Ren

“Surface-Enhanced Raman Spectroscopy: Benefits, Trade-Offs and Future Developments“. Chem. Sci. 2020. 11(18): 4563-4577. 10.1039/d0sc00809e

34.

Mirsafavi

Moskovits

Meinhart

“Detection and Classification of Fentanyl and Its Precursors by Surface-Enhanced Raman Spectroscopy“. Analyst. 2020. 145(9): 3440-3446. 10.1039/c9an02568e

35.

Wilson

N.G.

Raveendran

Docoslis

“Portable Identification of Fentanyl Analogues in Drugs Using Surface-Enhanced Raman Scattering“. Sens. Actuators, B. 2021. 330: 129303. 10.1016/j.snb.2020.129303

36.

Smith

Logan

Bazley

Blanchfield

Stokes

, et al. “A Semi-Quantitative Method for the Detection of Fentanyl Using Surface-Enhanced Raman Scattering (SERS) with a Handheld Raman Instrument“. J. Forensic Sci. 2020. 66(2): 505-519. 10.1111/1556-4029.14610

37.

Haddad

Comanescu

M.A.

Green

Kubic

T.A.

Lombardi

J.R.

“Detection and Quantitation of Trace Fentanyl in Heroin by Surface-Enhanced Raman Spectroscopy“. Anal. Chem. 2018. 90(21): 12678-12685. 10.1021/acs.analchem.8b02909

38.

Haddad

Green

Lombardi

J.R.

“Detection of Fentanyl in Binary Mixtures with Cocaine by Use of Surface-Enhanced Raman Spectroscopy“. Spectrosc. Lett. 2019. 52(8): 462-472. 10.1080/00387010.2019.1671871

39.

Chen

Bai

Zhuo

Yao

, et al. “Label-Free Identification of Fentanyl Analogues Using Surface-Enhanced Raman Scattering and Machine Learning Algorithm“. Vib. Spectrosc. 2023. 127: 103567. 10.1016/j.vibspec.2023.103567

40.

Wang

Vendrell-Dones

M.O.

Deriu

Dogruuer

Harrington

P.B.

, et al. “Multivariate Analysis Aided Surfance-Enhanced Raman Spectroscopy (MVA-SERS) Multiplex Quantitative Detection of Trace Fentanyl in Illicit Drug Mixtures Using a Handheld Raman Spectrometer“. Appl. Spectrosc. 2021. 75(10): 1225-1236. 10.1177/00037028211032930

41.

Muneer

Smith

Bazley

M.M.

Cozzolino

Blanchfield

J.T.

“Detection of Low-Level Fentanyl Concentrations in Mixtures of Cocaine, MDMA, Methamphetamine, and Caffeine Via Surface-Enhanced Raman Spectroscopy“. J. Forensic Sci. 2025. 70(1): 73-83. 10.1111/1556-4029.15652

42.

Erkok

S.D.

Kimani

M.M.

Lanzarotta

McCord

“Detecting Fentanyl Analogs in Counterfeit Pharmaceuticals by Surface-Enhanced Raman Spectroscopy Using Handheld Raman Spectrometers“. Forensic Sci. Int. 2026. 378: 112679. 10.1016/j.forsciint.2025.112679

43.

Gonzalez-Nieto

Wallace

Kielty

Gruntman

Robinson

, et al. “Not Just Fentanyl: Understanding the Complexities of the Unregulated Opioid Supply Through Results From a Drug Checking Service in British Columbia, Canada“. Int. J. Drug Policy. 2025. 138: 104751. 10.1016/j.drugpo.2025.104751

44.

Drug Checking.

Substance

“Annual Review 2024”. https://substance.uvic.ca/files/reports/Substance%20Drug%20Checking%202024%20Annual%20Report.pdf [accessed 24 June 2025].

45.

Ott

C.E.

Burns

Sisco

Arroyo

L.E.

“Targeted Fentanyl Screening Utilizing Electrochemical Surface-Enhanced Raman Spectroscopy (EC-SERS) Applied to Authentic Seized Drug Casework Samples“. Forensic Chem. 2023. 34: 100492. 10.1016/j.forc.2023.100492

46.

Lai

Zhang

“Advanced Sample Preparation Techniques for Rapid Surface-Enhanced Raman Spectroscopy Analysis of Complex Samples“. J. Chromatogr. A. 2022. 1675: 463181. 10.1016/j.chroma.2022.463181

47.

dos Santos

D.P.

Sena

M.M.

Almeida

M.R.

Mazali

I.O.

Olivieri

A.C.

, et al. “Unraveling Surface-Enhanced Raman Spectroscopy Results Through Chemometrics and Machine Learning: Principles, Progress, and Trends“. Anal. Bioanal. Chem. 2023. 415: 3945-3966. 10.1007/s00216-023-04620-y

48.

Seifert

“Application of Random Forest Based Approaches to Surface-Enhanced Raman Scattering Data“. Sci. Rep. 2020. 10(1): 5436. 10.1038/s41598-020-62338-8

49.

Martens

R.R.

Gozdzialski

Newman

Gill

Wallace

, et al. “Trace Detection of Adulterants in Illicit Opioid Samples Using Surface-Enhanced Raman Scattering and Random Forest Classification“. Anal. Chem. 2024. 96(30): 12277-12285. 10.1021/acs.analchem.4c01271

50.

Martens

R.R.

Gozdzialski

Newman

Gill

Wallace

, et al. “Optimized Machine Learning Approaches to Combine Surface-Enhanced Raman Scattering and Infared Data for Trace Detection of Xylazine in Illicit Opioids“. Analyst. 2025. 150(4): 700-711. 10.1039/d4an01496k

51.

Cooman

Trejos

Romero

Arroyo

“Implementing Machine Learning for the Identification and Classification of Compound and Mixtures in Portable Raman Instruments“. Chem. Phys. Lett. 2022. 787: 139283. 10.1016/j.cplett.2021.139283

52.

Dong

Weng

Yang

Liu

“Detection and Direct Readout of Drugs in Human Urine Using Dynamic Surface-Enhanced Raman Spectroscopy and Support Vector Machines“. Anal. Chem. 2015. 87(5): 2937-2944. 10.1021/acs.analchem.5b00137

53.

Gozdzialski

Hutchison

Wallace

Gill

Hore

“Toward Automated Infrared Spectral Analysis in Community Drug Checking“. Drug Test. Anal. 2024. 16(10: 83-92. 10.1002/dta.3520

54.

Darie

I.F.

Anton

S.R.

Praisler

“Machine Learning Systems Detecting Illicit Drugs Based on Their ATR-FTIR Spectra“. Inventions. 2023. 8(2): 56. 10.3390/inventions8020056

55.

Breiman

“Random Forests“. Mach. Learn. 2001. 45: 5-32. 10.1023/A:1010933404324

56.

Substance. “Vancouver Island Drug Checking Project”. https://substance.uvic.ca [accessed 7 July 2025].

57.

Gozdzialski

Rowley

Borden

Saatchi

Gill

C.G.

, et al. “Rapid and Accurate Etizolam Detection Using Surface-Enhanced Raman Spectroscopy for Community Drug Checking“. Int. J. Drug Policy. 2022. 102: 103611. 10.1016/j.drugpo.2022.103611

58.

Gozdzialski

Aasen

Larnder

Ramsay

Borden

S.A.

, et al. “Portable Gas Chromatography–Mass Spectrometry in Drug Checking: Detection of Carfentanil and Etizolam in Expected Opioid Samples“. Int. J. Drug Policy. 2021. 97: 103409. 10.1016/j.drugpo.2021.103409

59.

Miskulin

Wallace

Gill

Hore

“A Strategy for the Detection of Benzodiazepine Drugs Using Low-Resolution Paper-Spray Mass Spectrometry for Harm Reduction Drug Checking“. Drug Test. Anal. 2024. 16(10): 1085-1093. 10.1002/dta.3630

60.

Borden

S.A.

Saatchi

Vandergrift

G.W.

Palaty

Lyshyshyn

, et al. “A New Quantitative Drug Checking Technology for Harm Reduction: Pilot Study in Vancouver, Canada Using Paper Spray Mass Spectrometry“. Drug Alcohol Rev. 2022. 41(2): 410-418. 10.1111/dar.13370

61.

Borden

S.A.

Saatchi

Krogh

E.T.

Gill

C.G.

“Rapid and Quantitative Determination of Fentanyls and Pharmaceuticals From Powdered Drug Samples by Paper Spray Mass Spectrometry“. Anal. Sci. Adv. 2020. 1(2): 97-108. 10.1002/ansa.202000083

62.

Wallace

Shkolnikov

Kielty

Robinson

Gozdzialski

, et al. “Is Fentanyl in Everything? Examining the Unexpected Cccurrence of Illicit Opioids in British Columbia’s Drug Supply“. Harm Reduction J. 2025. 22(1): 28. 10.1186/s12954-025-01189-w

63.

Pedregosa

Varoquaux

Gramfort

Michel

Thirion

, et al. “Scikit-learn: Machine Learning in Python”. J. Mach. Learn. Res. 2011. 12(85): 2825–2830.

64.

Tobias

Grant

C.J.

Laing

Arredondo

Lysyshyn

, et al. “Time-Series Analysis of Fentanyl Concentrations in the Unregulated Opioid Drug Supply in a Canadian Setting“. Am. J. Epidemiology. 2022. 191(2): 241-247. 10.1093/aje/kwab129

65.

Wang

Deriu

Mebel

A.M.

McCord

“Surface-Enhanced Raman Spectroscopy, Raman, and Density Functional Theoretical Analyses of Fentanyl and Six Analogs“. J. Raman Spectrosc. 2019. 50(10): 1405-1415. 10.1002/jrs.5656

66.

Erkok

S.D.

Hernandez

Cruz

Mebel

A.M.

McCord

“Differentiating Structurally Similar Fentanyl Analogs by Comparing Density Functional Theory (DFT) Calculations and Surface-Enhanced Raman Spectroscopy (SERS) Results“. Appl. Spectrosc. 2024. 78(7): 667-679. 10.1177/00037028241246010

67.

Leonard

Haddad

Green

Birke

R.L.

Kubik

, et al. “SERS, Raman, and DFT Analyses of Fentanyl and Carfentanil: Towards Detection of Trace Samples“. J. Raman Spectrosc. 2017. 48(10): 1323-1329. 10.1002/jrs.5220

68.

Ferguson

Tupik

S.L.

Haddad

Perr

Gilbert

, et al. “Utility of Gas Chromatography Infrared Spectroscopy (GC-IR) for the Differentiation of Positional Isomers of Fentanyl Related Substances“. Forensic Chem. 2022. 29: 100425. 10.1016/j.forc.2022.100425

69.

Gilbert

Mewis

R.E.

Sutcliffe

O.B.

“Fast & Fluorinated: Development and Validation of a Rapid Benchtop NMR Approach and Other Routine Screening Methods for the Detection and Quantification of Synthesized Fluorofentanyl Derivatives“. Forensic Chem. 2021. 23: 100321. 10.1016/j.forc.2021.100321

70.

Mohammed

Budach

Feuerpfeil

Ihde

Nathansen

, et al. “The Effects of Data Quality on Machine Learning Performance on Tabular Data“. Inf. Syst. 2025. 132: 102549. 10.1016/j.is.2025.102549

71.

Bardwell

Boyd

Tupper

K.W.

Kerr

“‘We Don’t Got that Kind of Time, Man. We’re Trying to Get High!’: Exploring Potential Use of Drug Checking Technologies Among Structurally Vulnerable People Who Use Drugs“. Int. J. Drug Policy. 2019. 71: 125-132. 10.1016/j.drugpo.2019.06.018

72.

Tallarida

R.J.

“Quantitative Methods for Assessing Drug Synergism“. Genes Cancer. 2011. 2(11): 1003-1008. 10.1177/1947601912440575

73.

Xie

Liu

Chen

Liu

, et al. “Key Role of Direct Adsorption on SERS Sensitivity: Synergistic Effect Among Target, Aggregating Agent, and Surface with Au Or Ag Colloid As Surface-Enhanced Raman Spectroscopy Substrate“. J. Phys. Chem. Lett. 2020. 11(3): 1022-1029. 10.1021/acs.jpclett.9b03724

74.

Tay

L.-L.

Poirier

Ghaemi

Hulse

Wang

“Iodide Functionalized Paper-Based SERS Sensors for Improved Detection of Narcotics“. Frontiers Chem. 2021. 9: 680556. 10.3389/fchem.2021.680556

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

2.38 MB