Nanoparticle surface-enhanced Raman spectroscopy as a noninvasive, label-free tool to monitor hematological malignancy

Hematological malignancies are a heterogeneous group of diseases that include leukemias, lymphomas and multiple myeloma (MM). Together, they account for approximately 10% of all cancers [1,2]. There has been a global increase in mortality for many of these malignancies since 2006 [3] and MM remains incurable [4]. Lymphoma and MM are both lymphoid neoplasms that arise from abnormal lymphocytes. Although distinct diseases, achieving deep levels of response is essential for prolonged progression-free survival in both lymphoma and MM. Many patients achieve a state of complete remission, however, the presence of minimal residual disease (MRD) means patients will ultimately relapse at an undefined rate of recurrence. This uncertainty negatively impacts patient quality of life and the reliability to confirm relapse requires substantive disease activity. Accurately monitoring for and characterizing MRD using innovative techniques could inform clinicians, and potentially improve treatment strategies, to better overall survival, patient reassurance or length of remission [5,6].

Although several methods exist for monitoring treatment response and detection of MRD many of these options are highly complex, relatively insensitive and remain costly [5,7]. The gold standard blood biomarker in MM that quantifies monoclonal immunoglobulin levels by serum protein electrophoresis is readily available but has limited sensitivity. More sensitive methods, such as multiparametric flow cytometry and allele-specific oligonucleotide PCR, are options in MM, but the absence of readily detectable circulating disease or canonical chromosomal translocations in lymphoma preclude their use in lymphoma disease management. Newer methods using next-generation sequencing have yielded promising results for both lymphoma and MM detection of MRD, yet still rely on the bulk presence of disease [5,6,8]. Raman spectroscopy offers a non-invasive, cost-effective approach to detect MRD and monitor disease progression in an unbiased manner that simultaneously reflects the changes in proteins, DNA and RNA captured by the singularly focused methodologies described, with the potential sensitivity to not rely on bulk disease accumulation.

Raman spectroscopy is based on inelastic light scattering and can capture the chemical ‘fingerprints’ of specific biomolecular structures and conformations of macromolecules, including proteins, nucleic acids, lipids, drugs and metabolites. Although conventional Raman signals are weak, SERS enhances Raman light scattering through the addition of a metal (gold or silver) embedded on surfaces or in plasmonic nanoparticles (NPs) [9]. Specifically, the addition of metal AuNPs (nanoSERS) amplifies signal intensity by quantum effects and achieving maximal surface area scattering [9–11].

Studies have shown that Raman and SERS from the blood [12], urine [13], saliva [14] and tears [15] can detect unique signatures for the identification, classification, and diagnosis over a range of diseases including cancer [9], cardiovascular disease [16], Alzheimer's disease [17] and liver disease [18], among others. Within the clinical cancer setting, studies have shown Raman spectroscopy can be applied in diverse applications, including identifying cancerous tissues and establishing tumor borders [19], detecting cell-based changes after therapies [20,21], characterizing cancer-specific exosomes [22,23], and identifying cancer cells from non-cancerous cells [24]. These applications are capable of diagnosing and staging prostate [12,25,26], ovarian [27], breast [28,29], lung [30] and oral cancers [31–33], as well as hematological malignancies [34,35]. The discriminating power of these datasets often uses advanced statistical and machine-learning approaches, most commonly principal component analysis (PCA) alone or combined with a supervised machine learning model, such as a linear discriminant analysis classifier (LDA) or support vector machines (SVM) [25].

These classification models have been effectively applied for diagnosing and staging various cancers when referenced to a healthy control population [14,31,36–38]. However, few studies have shown the ability of SERS to make discriminations in relation to unrelated pathologies or age-related comorbidities. Additionally, detection and classification are only the first steps in clinical management. Monitoring treatment effects and detecting MRD is promising for improving patient outcomes as it provides clinicians with data on whether treatment strategies need to be adjusted. Two previous studies have successfully differentiated between benign and malignant pathologies using nanoSERS technologies [26,27], although only a single study has directly compared distinct pathologies. Yu et al. were able to effectively classify nasopharyngeal cancer from liver cancer using SERS from serum samples [32]. Although analysis of cell-specific changes to different treatment strategies have been done in vitro, in which Raman spectroscopy was shown to discriminate between viable and nonviable cells after photodynamic therapy [20] or chemotherapy [39,40], few studies have monitored changes in blood samples of cancer patients undergoing standard treatments. Manago et al. observed changes in three leukemia samples after chemotherapy [35] and Tsao et al. followed 10 melanoma patients after treatment with immunotherapies or targeted therapies to assess changes in circulating tumor cells using antibody-conjugated gold nanoparticles for nanoSERS analysis [41]. Similarly, an immunolabeled based SERS immunoassay was developed to detect MRD in 13 patients with B-cell hematological malignancies [42]. However, these studies are limited by their small sample size and approaches needed for translation to clinical medicine.

In this study, we combined statistical analysis and machine learning approaches with label-free nanoSERS technology to define a hematological cancer-specific nanoSERS spectral profile compared with two control populations: young–healthy, and aged–sick. Additionally, by assessing cancer patients iteratively within a standard-of-care setting and assigning individualized disease-severity scores, we developed an algorithm that can predict the clinical status based on their nanoSERS spectral profiles.

Materials & methods

Patient samples

Plasma samples and clinical characteristics from patients diagnosed with MM, lymphoma or cardiovascular disease (CVD) and treated at the Saint John Regional Hospital were collected with the approval of the Horizon Health Network Research Ethics Board. Plasma from the healthy control group was obtained from volunteer New Brunswick Community College (NBCC) students under approval from NBCC Research Ethics Review Board. Plasma samples were isolated from whole blood by centrifuging for 1200×g for 5 min. All plasma samples were stored at -80°C and thawed on ice before use. A disease burden scale was designed in collaboration with clinicians. Based on clinical information from medical chart reviews, disease burden scores (Supplementary Table 1) for cancer patients were assigned for each point when plasma samples were collected (Supplementary Figure 1).

Nanoparticle formulation and SERS measurements

AuNPs were synthesized by combining 50 mg/ml polyethylene glycol (PEG) solution (Sigma, Ontario, Canada, Cat. 412325) with 20 mM gold (Au) chloride trihydrate (Sigma) in a 1:100 ratio and vortexed for 3 min. For the complete formation of AuNP, the solution was incubated at room temperature for 3–4 h. PEG solution was made in 1× PBS and dissolved by rotating overnight at 4°C [11]. The AuNPs were characterized alone or in the presence of plasma by UV-Vis spectroscopy (BioTEK Synergy H4, VT, USA; cm^-1 spectral resolution) and transmission electron microscopy (TEM) (JEOL 2011 scanning transmission electron microscope, University of New Brunswick, IN, USA; 50K resolution). Twenty microliters of patient plasma was mixed with 25 μl of AuNPs and incubated at 37°C for 2 min before transferring to a glass slide and allowed to dry. SERS spectra were measured using a Renishaw inVia Raman Spectrometer (Renishaw, University of New Brunswick Planetary and Space Science Centre, Ontario, Canada) with a helium–neon laser beam with 514 excitation wavelength, at 10% laser power and 3 accumulations of spectra ranging from 100 to 2000 cm^-1 after calibration of the instrument.

Spectral processing and data analysis

The spectra were obtained using the OriginLab software package (OriginLab Corporation, MA, USA). For the first portion of the analysis, the resulting raw nanoSERS intensity measurements were used as features for addressing the two classification problems of interest: discriminating cancer subjects from the young–healthy subjects (two-class problem), and discriminating between the cancer, young–healthy, and aged–sick (CVD) cohorts (three-class problem). These approaches sequentially rank features based on criterion values that define the relevance or discriminant power of the feature related to the target classes. Before classifier training, the most relevant features for each problem were identified through filter feature selection, using either an analysis of variance (ANOVA) ranking approach or the ReliefF algorithm. By applying one-way ANOVA testing for the applicable cohorts, the former method can independently rank each feature based on its F-statistic. The latter method, ReliefF, uses the concept of ‘nearest-neighbors’ in the feature space to iteratively score features given their similarity among subjects of the same class and dissimilarity among subjects of differing classes. Unlike the ANOVA method, ReliefF is also able to take interactions between features into consideration. Once ranked, the top features were used to train linear and quadratic discriminant analysis (LDA and QDA) classifiers. Both are supervised learning algorithms that require training exemplars with known class memberships to learn proper associations, enabling the classification of new, ‘unseen’ testing exemplars. With LDA and QDA, class associations are dictated by linear and quadratic decision boundaries, respectively. To evaluate the performance of each classifier on unseen samples, a leave-one-out cross-validation (LOOCV) procedure was implemented, where one sample is used for testing the classifier and all others are used for training. In turn, each sample is used as the testing exemplar and the classifier's performance is averaged across all testing samples to obtain a final estimate. LOOCV accuracy was lastly used to tune the number of features included in each model by adding the ranked features one at a time to the model and re-evaluating its performance. The set of features that provided the highest classification accuracy overall was selected as the final set.

The second portion of the analysis incorporated feature extraction and transformation techniques to increase the information density of the raw spectra. These techniques were used to, first, enable visualization of the class separability afforded by the spectra, and second, to develop a regression model for predicting disease burden scores. In addition to the raw spectra and their first and second derivatives, the augmented feature sets consisted of binary barcoded versions of the spectra and their derivatives. These barcoded sequences were generated by assigning all points exceeding a threshold of 10% the maximum value in the derivative sequence a value of one, and all others zero [43]. PCA was employed as previously proposed by Patel et al. [43] to help visualize the separability of the classes. PCA, a dimensionality reduction technique that maintains the variance of the data [32], was then applied to each of the extracted sequences. The PCA eigenvectors were computed using data from healthy controls and used to compute principal component values for all subjects (controls, cancer and CVD). The features from the raw nanoSERS spectra, derivative sequences, barcode sequences and PC projections were then used to build a linear SVM regression model for predicting the subjects' disease burden scores. The features included in the model were tuned using sequential feature selection (SFS), a wrapper feature selection method that uses model performance (in this case, the mean-square-error [MSE] of the predicted burden scores) to construct a locally optimal subset of features. With SFS, the features that provide the best improvement in MSE are successively added to the feature set until it cannot be further improved. As with the classification models, CV was used to obtain model performance estimates for unseen subjects. Due to the computational expense associated with SFS, 10-fold CV was implemented rather than LOOCV, splitting the data into 10 partitions and using one at a time for testing the model trained on the remaining nine. The final performance measure was computed by averaging from 10 results.

Results

Characterization of nanoparticles and validation of nanoSERS

Transmission electron microscopy (TEM) was used to characterize the AuNPs. The AuNPs were spherical and approximately 50 nm in diameter (Figure 1A). When the AuNPs were mixed with blood plasma, there was an increase in the UV-Vis intensity relative to AuNPs alone (Figure 1B). The UV-Vis spectrum of the AuNPs also showed a new broad plasmon resonance absorption band at the 560 cm^-1 wavelength region (Figure 1B, marked by arrow). This change likely originated from the localized surface plasmons resonance of deposited AuNPs in combination with the biochemical substances in plasma [11,31,33]. The new plasmon resonance absorption band largely overlaps with the Raman excitation laser, which will lead to significant enhancement in the intensity of Raman scattering.

Our nanoSERS approach imaged dried mixtures of plasma and AuNPs. Images (Figure 1C) and normalized spectra (Figure 1D) were obtained from random spots from the dried plasma/AuNP mixtures. To validate our nanoSERS approach, we measured two distinct spots on each of 47 samples taken from our healthy population. Paired analysis at 10 random peaks shows no significant differences between spots (Supplementary Figure 2), thus confirming non-random intra-assay variance (Supplementary Figure 3). The high concordance between two different spots served as proof-of-principle that the nanoSERS approach was robust. Therefore, for subsequent cardiovascular and cancer patient samples, only one spot was imaged to conserve sample and in accordance with clinical sample management and processing expectations.

Clinical demographics

Unique to our study, two reference populations were included (Table 1). The first cohort comprised 47 young–healthy volunteers (certain to be under the age of 65 years but whose date of birth and gender was not included as part of data collection). The second cohort included 25 aged–sick participants that were hospitalized for cardiovascular disease. The ages in this CVD cohort ranged from 52–74 years (mean 65.0 ± 6.4) with 72% male sex. The study cohort included 25 hematological cancer patients, ranging in age from 45 to 80 years (mean 65.2 ± 8.7) and 72% male sex. Of the cancer participants, 60% were diagnosed with MM, while the remaining 40% were diagnosed with lymphoma. Samples from the cancer patients were iteratively collected throughout treatment (Supplementary Figure 1). Disease burden scores (Supplementary Table 1) were assigned to both CVD and cancer patients. The majority of CVD patients had low (40%) or moderate (60%) disease burden scores, while cancer patients ranged from low (12%), to moderate (55%) and moderate-high (29%). Three patients within the cancer cohort died (4%).

Discrimination between clinical cohorts

Two approaches were used to discriminate between the three clinical cohorts to use reference populations to define unique spectral profiles for hematological cancer patients. The dataset includes 192 raw spectra (healthy, n = 91; CVD, n = 25; cancer, n = 76). Each sample consists of 1714 nanoSERS spectra in the range of 100–2000 cm^-1. For the first portion of the analysis, raw data were used as features without applying additional pre-processing and feature extraction methods. The mean nanoSERS spectra of the three cohorts are shown in Figure 2A. After ANOVA ranking or ReliefF-based automatic feature selection, LDA and QDA were used to examine whether the raw data is capable of discriminating cancer patients from aged–sick CVD patients and healthy subjects (three-class problem) or if cancer patients can be distinguished from healthy subjects alone (2-class problem; Figure 2B). Both LDA and QDA were able to discriminate between the cancer patients and healthy subjects with 79.04% accuracy using only the first ranked ANOVA feature. For the more challenging three-class problem, discrimination between cancer, CVD and healthy subjects, the classifiers still performed substantially better than chance (33% for the three-class problem), yielding 69.69% and 71.35% for LDA and QDA, respectively, using the single best ANOVA feature.

Both feature selection approaches selected measurements from the same region (1358.8–1360.9 cm^-1) of the nanoSERS spectra (Figure 2A inset) when discriminating between cancer and healthy subjects. For the three-class problem, the QDA classification accuracies were found using the first ranked feature (at 1103.6 cm^-1 for ANOVA F-value-base and at 1990.3 cm^-1 for Relief (Figure 2A, inserts). The positions of the top n features used to yield the maximum accuracies are shown in Supplementary Table 2.

Second, feature extraction and transformation techniques were applied to the raw spectra to enable visualization of the data's separability. Figure 3A shows the PCA scatter plot of the nanoSERS spectral data of the blood plasma samples of healthy controls, CVD patients and cancer patients using the first PC of the first derivative of the spectra and the first PC of the barcoded spectra (derived from the second derivative of the spectra) [43]. This plot shows a distinct separation between healthy and sick (CVD and cancer) participants and reasonably good differentiation between the three cohorts. Importantly, the healthy control population, representing the most homogenous group, forms a tight cluster on the PCA scatter plot. The spectral regions most responsible for clustering healthy and cancerous samples in the PCA scatter plot are indicated by the PC loading vector shown in Figure 3B.

Monitoring disease progression

Given that nanoSERS spectra changed over treatment courses (Figure 4A & D), we developed a model using nanoSERS to predict disease burden in cancer patients. To develop this prediction model, we used a linear SVM regression model and the top features selected from the raw data, their first and second derivatives, and their PCA scores using a sequential forward selection (SFS) approach. Actual disease burden scores were created using a nonvalidated in-house disease burden scale (Supplementary Table 1) based on clinical parameters. The best model estimated the disease burden with a mean square error (MSE) of 1.6 (based on 15 selected features), compared with assigned disease burden scores (Figure 4B & E).

Representative examples of disease progression are shown for a MM (Figure 4A–C & Supplementary Figure 4A & B) and lymphoma patient (Figure 4D–F & Supplementary Figure 4C & D). In each case, dynamic changes in assigned disease burden scores are reflected by similar changes in predicted scores. For example, the first sample collected for patient MM01 at diagnosis was assigned a score of 8; after completion of an autologous stem cell transplant, the patient remained stable with an assigned score of 4 through a period of 8 months. Similar reductions in the predicted score were observed; however, while the patient remained clinically stable, the predictive score progressively increased. In the second example, LY01, the patient was diagnosed before the initiation of this study, and the first sample was collected during a follow-up visit, where the disease was stable and assigned a disease burden score of 5. In a subsequent 5-month follow-up appointment, the lymphoma had progressed as indicated by an increase in the actual and the predicted disease burden scores so that the patient was started on chemotherapy treatment. The final sample for this patient was collected after completion of this treatment regimen, where the disease was in remission with a concomitant decrease in both predictive and assigned disease burden scores. These examples reflect the prediction model's ability to determine the patient's clinical status.

Discussion

Using advanced statistical and machine-learning approaches, this study describes a unique hematological malignancy nanoSERS spectral profile distinct from two reference populations: young–healthy and aged–sick. Moreover, by incorporating the individual cancer patients' treatments and recovery, our study is one of the first to use a cell-free, label-free nanoSERS approach that integrates patient-specific nanoSERS spectral profiles into a predictive model capable of assessing changes in disease over time.

Raman spectroscopy and nanoSERS have been widely described for their ability to discriminate between cancer and healthy [28,37,44,45] or nonmalignant precursor and early-stage conditions [25,27,33]. However, given the multidimensionality of Raman spectroscopy [9], subtle differences in metabolites detected by Raman spectroscopy may not be disease-specific. For example, the spectral peak of 1585 cm^-1 has been identified as a biomarker to distinguish between gastric cancer and normal tissue [46] and a marker to describe the redox state of cardiomyocytes [47]. Despite this overlap in spectral profiles across diseases, few studies incorporate different malignancies alongside healthy controls to define cancer-specific spectral profiles [32]. For this reason, we analyzed nanoSERS profiles from healthy controls and patients with cardiovascular disease as a distinct noncancer pathology to better define a nanoSERS profile more specific to lymphoma and MM. Indeed, using either LDA or QDA after ANOVA F-valued based first-ranked feature selection, cancer patients could be differentiated from cardiovascular patients and healthy controls with ∼70% accuracy (Figure 2B). Although these accuracies cannot reliably inform diagnosis at this time, with further refinement and scaling of our nanoSERS approach and reference base, diagnostic segregation is plausible.

In clinical practice, accurately detecting MRD using timely and noninvasive techniques, while ensuring high degrees of sensitivity and specificity remains a challenge. Our study serves as proof-in-principle that nanoSERS may offer an approach to monitor disease progression by developing a predictive model that correlates spectral profiles with clinically defined parameters (Figure 4). Previous studies have used Raman spectroscopy or nanoSERS to assess in vitro cytotoxic cell-based changes [20,39,40,48] or antibody-labeled nanoSERS probes were able to monitor a patient's anticancer treatment response [41,42]. Although our study used a nonvalidated tool to assign disease severity, it is clinically based on outcomes and response to therapy. This tool and our model used a cell-free, label-free approach to monitor disease progression, shifting the balance from high specificity towards unbiased determinants, which impact disease progression and should support scaled application and cost projections favourably.

There are several limitations both to this study and within the broader Raman field that restrict progression into clinical practice at this time. Clinical applications require stringent bioanalytical conditions for precise data output, as has been established for clinical applications such as immunohistochemistry [49]. Although we (Supplementary Figure 2) and others [28,50] have attempted to control for internal variability by averaging and comparing spectra within patient datasets, few studies have fully tested the technical boundaries of Raman applications. Numerous factors are well-known to affect the reproducibility of SERS’ reproducibility [51]; yet these conditions have not been optimized across the field. Instead, diverse approaches that incorporate different substrates, such as dried metal surfaces [20,23,38], silver or gold nanoparticles [25,31,33,50] or passive corona accumulation and antibody-conjugated probes [41,42] have been used. Different laser frequencies, detectors, acquisition protocols and sample solvents can result in different peak assignments and affect classification accuracies [14,52]. Sample preparation such as using dried or liquid samples [53] or factors such as temperature, timing and freeze-thawing, have been shown to affect spectral profile [53,54]. Similarly, the type of sample – serum, plasma or isolated extracellular vesicles – can significantly affect spectral classifications [26]. Such diverse approaches, although evidence-building, prevent any type of meta-analysis or development of consensus statements regarding the clinical applicability of nanoSERS for cancer diagnostics and theranostics. Moving forward, the Raman field requires standardization of technical approaches and analytical references that can be relied on across institutions.

In the absence of these global standards, the current expectation is to address the complex heterogeneity of Raman methodologies and spectral profiles by using advanced statistical approaches. Some of the more commonly used data-reduction methods include LDA, PCA, SVM models, and regression models, such as principal components regression or partial least squares models [55]. K-fold cross-validation is often used to validate findings when the sample size is limited [31,36]. To improve accuracy, several studies combine PCA with LDA [31,36]. However, these high accuracies are likely artificial and overfit to the inherent noise since these data reduction methods are limited by sample size, particularly where the number of samples is less than the number of variables/features [28,44,45]. Our study has approximately 100 samples and 1700 features (wavenumbers). To lessen any overfitting of data, we used PCA, LDA and QDA as independent classification strategies (Figures 2 & 3). Even so, our data is likely overfitted due to the low sample size [56] and should be interpreted cautiously until a larger cohort of samples and patients can be recruited.

With low sample sizes and overfit models, the accuracy of classifications may be overestimated and it is possible that the selection of relevant peaks may not represent true disease states. Our study and others [37,44] have shown that the choice of classification method affects the determination of relevant peaks. In our independent PCA, LDA and QDA analyses to classify hematological cancers from cardiovascular disease and healthy controls, none of the selected features overlapped between the three methods (Figures 2B, 3B & Supplementary Table 2). Similarly, there is a lack of consensus in feature selection within a specific cancer type. For example, many of the selected peaks identified in either prostate cancer cell lines or prostate cancer tissue were not consistent across studies [57]. Although similar peak regions were diagnostic for prostate cancer patients compared to nonmalignant prostate conditions [36] or based on Gleason Grade [12] by assessment of serum combined with silver nanoparticles, other studies using alternative SERS approaches failed to identify these same peaks in prostate cancer [26,50]. Again, these results highlight the need for universal protocols and reference standards to ensure reproducibility and accuracy of SERS analysis across clinical centers to be translatable.

To overcome the limitations of machine learning models, increased sample sizes with more controlled populations are needed to explain the heterogeneity and complexity of Raman analysis. Controlled populations would ideally be both age- and sex-matched. While some studies include age and sex-matched cohorts [30,36,44], many Raman studies do not integrate age and/or sex into their analysis [27,28,35,38,50]. In our study, our CVD and cancer cohorts were age- and sex-matched; however, the healthy population was comprised entirely of those under the age of 65 years (Table 1); it remains to be determined if an aged-matched or pooled sample of healthy volunteers would make a better reference standard. Interpatient heterogeneity adds another layer of complexity to Raman analysis. Our study retrospectively assessed a random selection of patients diagnosed and treated at uncontrolled points of time (Supplementary Figure 1) with chemotherapies, targeted therapies, radiation or stem cell transplants, as part of their standard of care. Additionally, this cancer cohort includes both lymphoma and MM patients that, although both lymphoid malignancies [58], are distinct diseases and with larger sample sizes could likely possess distinct nanoSERS spectral profiles. While many of the studies with fewer than 100 total participants [15,27,28,41,44] or as low as three patient samples [35] have reported classification accuracies and selection of relevant peaks despite the small sample size, larger sample sizes are needed to accurately define spectral profiles before clinical utilization. Future studies should consider incorporating nanoSERS analysis as part of large randomized controlled studies that follow participants over defined periods to improve diagnostic, prognostic or theranostic clinical translation to reduce heterogeneity and refine questions.

In addition to its clinical applications, nanoSERS has the power to serve as a translational tool through the ascription of spectral peaks to physiological phenomena such as apoptosis, degradation of DNA and protein abundance [40,41] as well as drug metabolism [59]. With improved standardization and reference samples in the Raman field, this type of analysis could provide mechanistic insight into the development of MRD and/or cancer progression.

Establishing early guidelines for the standardization of references, nanoSERS processes and analysis are necessary. Only through the generation of these consortium statements will the Raman field move into the clinical sphere. Despite the breadth of Raman studies' over the past years, the heterogeneity in samples types, methodologies and analyses has prevented consensus as to how Raman can be used to define real-time biological and/or clinical phenomenon. In addition to further validating the Raman approach by testing its boundaries and sensitivities, we propose that the Raman community engage in meaningful discussions to produce practical guidelines on the use of Raman and nanoSERS for translational medicine.

Conclusion

Our study has shown that a cell-free, label-free nanoSERS approach is capable of monitoring disease progression for patients with hematological malignancies using two reference datasets. The integration of reference-based controls in age and disease parallel populations is an important advancement for confidence to inform clinical decision making. To date, most references were singular (healthy-comparator or intrapatient-remote-tissue reference), which are not discriminating enough for prognosis. A cell-free liquid biopsy approach has an advantage to systemic fingerprinting of a patient's status that could displace conventional approaches of disease monitoring-with potentially lower cost, higher sensitivity, and multiparametric information enrichment. Our study presents another step forward in incorporating Raman spectroscopy into clinical practice through enhanced reference use and machine learning.