Performance evaluation of handheld Raman spectroscopy for cocaine detection in forensic case samples

Abstract Handheld Raman spectroscopy is an emerging technique for rapid on‐site detection of drugs of abuse. Most devices are developed for on‐scene operation with a user interface that only shows whether cocaine has been detected. Extensive validation studies are unavailable, and so are typically the insight in raw spectral data and the identification criteria. This work evaluates the performance of a commercial handheld Raman spectrometer for cocaine detection based on (i) its performance on 0–100 wt% binary cocaine mixtures, (ii) retrospective comparison of 3,168 case samples from 2015 to 2020 analyzed by both gas chromatography–mass spectrometry (GC–MS) and Raman, (iii) assessment of spectral selectivity, and (iv) comparison of the instrument's on‐screen results with combined partial least square regression (PLS‐R) and discriminant analysis (PLS‐DA) models. The limit of detection was dependent on sample composition and varied between 10 wt% and 40 wt% cocaine. Because the average cocaine content in street samples is well above this limit, a 97.5% true positive rate was observed in case samples. No cocaine false positives were reported, although 12.5% of the negative samples were initially reported as inconclusive by the built‐in software. The spectral assessment showed high selectivity for Raman peaks at 1,712 (cocaine base) and 1,716 cm−1 (cocaine HCl). Combined PLS‐R and PLS‐DA models using these features confirmed and further improved instrument performance. This study scientifically assessed the performance of a commercial Raman spectrometer, providing useful insight on its applicability for both presumptive detection and legally valid evidence of cocaine presence for law enforcement.

substances should be seized, or suspects should be taken into custody. Many current standard methodologies for cocaine detection comprise of colorimetric spot tests based on the formation of a blue colored cobalt(II) thiocyanate complex in presence of cocaine. 2,3 Despite being fast and inexpensive, colorimetric tests have several limitations. First, such tests are destructive and thus require opening the packaging of the (yet) unknown substance. Without adequate precautions, this can impose a health risk when highly potent substances such as fentanyl derivates are encountered. Second, colorimetric tests are known to yield false positive (FP) results on some common pharmaceuticals such as lidocaine, levamisole, and promethazine. 4,5 Third, individual colorimetric tests are only available for a limited set of common drugs meaning that they are only selective towards a limited number of substances. This may lead to false negative (FN) results when the incorrect test is performed or a controlled substance is encountered for which no test is available. 3 Other indicative testing strategies for on-site cocaine detection using handheld devices include electrochemical analysis, 6 (near) infrared spectroscopy, 7-10 and Raman spectroscopy [11][12][13][14] where the latter two offer possibilities to analyze substances directly through packaging material (when this material is sufficiently transparent). Over the last decade, Raman spectroscopy has become a viable tool in forensic chemistry, and its applications have been reviewed multiple times. 9,[13][14][15] In addition to laboratory-grade instrumentation, various commercial handheld spectrometers with a specific focus on the operation by first responders, law enforcement officers and crime scene investigators have come to market. The TruNarc handheld Raman spectrometer was introduced in 2012 as the first commercial instrument designed specifically for forensic drug detection. Most handheld instruments are tailored for on-scene use by untrained personnel using single-button operation and subsequent readback of the (often binary) result on a display. Although efficient and easy to implement, a drawback of this approach is the lack of demonstrable evidential value and validation. Built-in data processing, library search and matching procedures and identification criteria are often proprietary and not fully disclosed by the manufacturer. As a result, the instrument operates as a "black box." Besides, manufacturers generally may not have access to a vast amount of representative case samples for validation, and therefore only limited studies are reported on the performance of the built-in functionality of these devices. [16][17][18] In situations in which the result is only used as an indication, this can be acceptable. However, when judicial actions such as seizure or custody are undertaken, this is unwanted as a FP outcome can have severe adverse consequences on the suspects involved, whereas a FN outcome can lead to serious crime remaining undetected. Additional insight into the performance and selectivity of such handheld devices is also crucial for optimal detection strategies in forensic laboratories.
When multiple (orthogonal) handheld techniques are available, these can provide complementary evidence and ultimately might eliminate the need for confirmatory analysis within the laboratory. In this way, tremendous process efficiency could be achieved in the judicial chain as both time and money are saved by the elimination of transport, administration, laboratory analysis and reporting steps. To that end, the overall turnover time from seizure to conviction could be decreased dramatically.
According to international guidelines provided by the Scientific Working Group on Drugs (SWGDRUG), Raman spectroscopy is considered a Category A technique providing the highest level of selectivity through structural information. 19 In this way, an analytical scheme based on Raman spectroscopy and a colorimetric test may suffice for unambiguous identification suitable as court evidence. However, dedicated requirements are described for handheld Raman spectrometers such as the instrument used in this study. The handheld has "to be assessed and validated for this purpose to ensure that the resolution and spectral range provide sufficient structural information to achieve the selectivity requirement of a Category A technique." 20 Additionally, spectral data need to be reviewable. Currently, in The Netherlands, handheld Raman spectrometers are only used for presumptive testing, and subsequent laboratory testing (e.g., gas chromatography-mass spectrometry [GC-MS] analysis) is required for court evidence. In certain circumstances, however, local policies (e.g., at dance festivals) can lead to a financial settlement if a suspect confesses the possession of a user quantity of illicit drugs after only a presumptive test by a handheld Raman spectrometer. In this way, no further laboratory testing is performed, and the Raman spectrometer is indirectly used for absolute identification purposes.
Nevertheless, one major limitation of Raman spectroscopy in forensic drug detection is its sensitivity. Mixtures-which cocainecontaining case samples often are-can be challenging as the analyzed signal contains features from all substances present, which may complicate database searches. Also, the presence of fluorescent dyes or impurities in a sample, even at low concentration, can obscure Raman signals and prevent identification of the compound of interest. A possible strategy to reduce the influence of fluorescence is the use of higher wavelength excitation lasers; however, this also comes at the cost of a reduced Raman signal. 21 A 785-nm excitation laser was reported as a good compromise between sensitivity and background fluorescence. 9 Surface-enhanced Raman spectroscopy (SERS) is a technique that could significantly increase sensitivity, although the use of dedicated consumables is required and SERS spectra are different from native Raman spectra complicating its applicability in onscene use. 9,22,23 29 Regarding cocaine, De Oliveira Penido and coworkers quantified cocaine base (crack cocaine) mixtures with sodium carbonate and either caffeine or lidocaine from 10 to 100 wt%. 30 Bedward et al. successfully quantified cocaine HCl concealed in food matrices such as cake mix in concentrations above 20 wt% using PLS-R. 31 It must be noted that all these studies regarding cocaine detection were performed on benchtop laboratory instruments. However, for on-scene indicative detection of street samples, these benchtop instruments are not practical, and cheaper and portable spectrometers are needed.
To our knowledge, no earlier work has reported the use of chemometric-based cocaine detection using a handheld Raman spectrometer. In addition, this is the first time the performance, and overall applicability of a handheld Raman spectrometer is assessed on a large set of actual case samples. In this work, we evaluated the performance       However, these consumables were not used as the aim of this study is to assess the performance of the handheld device for nondestructive and "direct scanning" purposes. The NPS substances 3-MEC and 6-APB produced inconclusive results that could be attributed to lacking spectra in the reference library, a known issue for novel drugs and an illustration of the need for the continuous updating of spectral libraries.

| Retrospective analysis
As a next step, all 0-100% binary cocaine mixtures were analyzed, and the results shown on the device's screen were compared to the known sample composition (Table S1). Each of the 80 samples was scanned 10 times. For five of these scans, the sample was repeatedly placed in the instrument without any precautions to mimic practical measurements, and for five other scans, special attention was put in place to make sure samples were neatly aligned in front of the laser for optimal signal. For all samples, the test result shown on the screen was either "cocaine", the actual identity of the present adulterant, or "inconclusive." When the instrument returned the identity of the cutting agent for a cocaine-containing sample, this was considered a FN result. Figure 1a shows the performance of the instrument on binary cocaine mixtures for all tenfold replicates and considering every noncocaine result (including inconclusive results) a FN. Figure S1 shows the comparison of these results with and without these inconclusive

| Spectral selectivity
To further investigate selectivity, the Raman spectral data were examined in more detail. Raw spectra showed major intensity differences both between replicate scans and among different substances.
Although relative intensities of spectral peaks from a single compound were repeatable, large variation was observed in absolute intensity and baseline offset. Both are well-known effects in direct spectroscopic analysis on solid samples caused by particle size differences, light scattering, and spectral interferences. Spectral preprocessing is a common strategy to extract useful information from the raw spectral data and remove nonselective systematic and random signal fluctuations. 24,25,37 Figure 2 shows the (a) raw spectral data, (b) data after SNV preprocessing, (c) subsequent smoothing, and (d) a 1,560-to 1,756-cm −1 selection with baseline correction for both cocaine HCl and cocaine base. Spectra of the cutting agents levamisole, paracetamol, and procaine are shown in the same plots to demonstrate the spectral selectivity. An example of the typical variation in raw spectral data and corresponding spectral consistency after SNV preprocessing can be found in the supplemental information ( Figure S2). Full 300-to 1,800-cm −1 Raman spectra of cocaine HCl, cocaine base and the eight most common cutting agents are shown in Figure S3. Around 1,000 cm −1 , a spectral peak is present in both cocaine types and in levamisole. This spectral peak is absent in the other spectra in and other drug substances did show a prominent Raman signal above 1,700 cm −1 . Common substances that did produce spectral features in this region are procaine, paracetamol, and phenacetin. As shown in Figure 2d, these three substances all yield a spectral peak near or even overlapping with the 1,600-cm −1 peak for cocaine. Procaine is the only substance in this study that had a spectral feature in close proximity-and slightly overlapping with-the 1,700-cm −1 cocaine peak. This high spectral selectivity explains the overall correct performance of the handheld Raman analyzer for high-level cocaine samples. However, in mixtures, which is often the case in actual cocaine case samples, 38   indicating that spectral sensitivity is the limiting factor. indicates that chemometric analysis of the Raman spectral data can lead to a better sensitivity than observed with the analyzer's built-in software. This is in line with PLS-DA scores of the 10-80 wt% cocaine mixtures shown in Figure 5b where a majority of the samples were predicted with a >0.8 score indicating the presence of cocainespecific spectral features. However, such features were almost absent for samples with a sub-0.5 score. The above 50 wt% cocaine samples that yielded a low score and lacked the cocaine-specific peak in the spectrum were found to be the same as the FN or inconclusive results from the handheld spectrometer's built-in software (Figure 1). Especially, in none of the spectra from 10-20 wt% cocaine mixtures with caffein, procaine or paracetamol, the cocaine-specific features were visible. This again indicates that FN or inconclusive results from the handheld spectrometer on binary cocaine mixtures could more likely be attributed to spectral limitations or measurement errors than misidentification by the built-in software.

| Performance assessment
The 45 spectra reported as negative or inconclusive by the handheld Raman analyzer although cocaine was identified by GC-MS (set FN from retrospective analysis) were also assessed by the PLSmodels. In this way, 21 of these spectra retrospectively resulted in a positive result for the presence of cocaine. Raman spectra of these samples showed a peak at either 1,712 or 1,716 cm −1 visible in the normalized ROI area as shown in Figure 7d. These peaks were also visible in this ROI area when the SNV normalization was applied on the entire spectrum, albeit at lower intensity due to strong Raman signals in the spectrum that do not originate from cocaine (Figure 7a-c).

| Discussion
This work demonstrates that the LOD for cocaine detected by a handheld Raman analyzer is highly dependent of sample composition and can vary between 40 wt% for cocaine samples containing procaine or paracetamol to 10 wt% for cocaine samples containing inositol.
Although a worst-case LOD of near 40 wt% seems rather problematic, the composition of typical cocaine-containing street samples justifies its application in a forensic setting. The prior probability to encounter a sub-40 wt% cocaine sample in forensic cases is very low because the reported cocaine content in street samples in Europe is often well above this level with averages up to 70 wt%. 35,36 This is in line with the results from the retrospective analysis described in Section 3.1.2 where only 2.5% of the cocaine-containing case samples analyzed between 2015 and 2020 were missed by the instrument's built-in detection algorithms. In addition, the absence of FP results for cocaine within these 3,168 samples furthermore underlines its applicability as a reliable cocaine detector in forensic casework. The results show that when cocaine is reported by the instrument's built-in software, it is very unlikely that this is a FP result and that the sample does not contain cocaine. In this way, additional spectral assessment and confirmation of the presence of cocaine specific spectral features such as demonstrated in this study may fulfill the SWGDRUG requirements for unambiguous cocaine identification eliminating the need for subsequent GC-MS analysis for certain highly concentrated cocaine samples. On the other hand, when no cocaine or other drug is reported, it is plausible that a controlled substance is still present. It must thus be noted that this study only focuses on cocaine detection. Other controlled substances may produce a less abundant or less selective Raman spectrum leading to different TP and TN percentages in forensic casework. The reported figures are also only valid for light-colored F I G U R E 7 Raman spectra of cocaine-containing case samples that were predicted false negative by the analyzer's built-in software and predicted (true) positive by the combined partial least squares discriminant analysis (PLS-DA) and regression (PLS-R) model. (a) Full-spectrum after standard normal variate (SNV); (b) zoom on the 1,560-to 1,756-cm −1 spectral area containing two cocaine-specific peaks; (c) the region of interest (ROI) area after the preprocessing used for PLS-R model: SNV followed by ROI; (d) spectra after PLS-DA preprocessing: ROI followed by SNV. All spectra overlaid with reference spectra of cocaine HCl (red) and cocaine base (green) solid samples (e.g., powders, chunks, and compressed powders) seized in a drug-suspected forensic setting. Common other drugs of abuse such as heroin and MDMA are often seized as brown powders or in colored tablets respectively, and both these substances yield fluorescence. LODs using direct Raman spectroscopy are therefore expected to be high, and the use of dedicated SERS kits is suggested to detect these compounds. In general, colored samples, samples containing highly fluorescent substances or samples with a higher prior probability to contain a lower concentration of cocaine, can lead to worse figures.

| CONCLUSIONS
The TruNarc handheld Raman analyzer is suitable for reliable cocaine detection in forensic case samples. Detection limits were found to be highly dependent on the type of adulterants present in the sample. cocaine, none of the drugs-of-abuse, cutting agents or pharmaceuticals included in this study yielded a spectral peak at these wavenumbers. A dual-stage PLS-DA and PLS-R model with an emphasis on these two peaks further increased the performance by correctly predicting 21 of the 45 cocaine containing samples that were previously missed by the instrument. Overall, these results demonstrate that reliable on-scene cocaine detection is feasible using the TruNarc handheld Raman analyzer and its built-in software. Additionally, the instrument's spectral resolution and selectivity for cocaine was found adequate for incorporation in analytical schemes ultimately leading to court evidence from handheld techniques.