A Novel Blood‐Based Screening Test for Colorectal Cancer Using Raman Spectroscopy Authors: Jett Karolewski, Georgia Carroll, Rosemary Carroll, Rodney J. Scott, Stephen Smith, Joel Petit Published: 2026 ABSTRACT ABSTRACTBackgroundLabel‐free vibrational spectroscopic techniques (Raman spectroscopy) combined with machine learning (ML) methodologies have huge potential for the development of screening and diagnostic tests in oncology. Traditionally, screening for colorectal cancer (CRC) has relied on immunochemical faecal occult blood (iFOBT) testing; however, this has shown to have low compliance in the Australian setting. The aim of this study was to assess the efficacy of Raman spectroscopy for the detection of CRC and potential screening for CRC.MethodsPlasma samples were analysed from a total of 370 participants; 117 CRC patients and 253 iFOBT positive control participants. 8 of the CRC patients were from the iFOBT cohort. The control cohort had either advanced adenomas (n = 81), non‐advanced adenomas (n = 76), colitis (n = 2) or no evidence of disease (n = 94). The Raman spectra of these plasma samples were analysed using ML algorithms in separate training (n = 222) and validation (n = 148) cohorts.ResultsA sensitivity of 84% and specificity of 93% was achieved for the detection of CRC in the validation cohort. When analysed in subcategories, 95% of people with CRC would receive a colonoscopy and 42% of FIT positive people could potentially be spared a colonoscopy.ConclusionThis exploratory study demonstrated a high level of accuracy for the detection of CRC using a Raman spectroscopy‐based ML model and has the potential to be used for CRC screening. Introduction In Australia, colorectal cancer (CRC) was the fourth most diagnosed cancer in 2020 and the second most common cause of cancer death in 2022 [1]. Although the incidence of CRC steadily increased in Australia between 1982 and 2001, it has steadily decreased since then, with the largest decrease (approximately 33% between 2001 and 2024) seen in any form of cancer [2]. This reduction in incidence has been attributed to the introduction of the National Bowel Cancer Screening Program (NBCSP) in 2006. Under the NBCSP Australians aged 50–74 are sent free screening iFOBT kits and invited to participate every 2 years. Participation has been shown to result in diagnosis at earlier stages and a lower risk of CRC overall [3, 4, 5, 6, 7]. However, in 2021–2022, only 40% of invitees participated in the NBCSP, and despite public awareness efforts, this percentage has been decreasing in recent years [8]. Comparatively, the United Kingdom, Danish and Dutch national CRC screening programmes achieve participation rates of 70%, 63% and 71%, respectively [9, 10, 11]. Australia's low participation rate impacts the efficacy of the screening programme and is due to many factors that include perceptions associated with the collection of stools. Additionally iFOBT screening programmes suffer from a high false positive rate with an estimated rate of 30%–50% of colonoscopies finding no pathology [8, 10, 11, 12, 13, 14]. These essentially unnecessary colonoscopies burden the healthcare system in terms of cost, morbidity and patient delay. As a result of poor compliance and the need for higher accuracy, the search has been on for blood‐based tests to diagnose CRC. Recently the US Food and Drug Administration has approved the use of a circulating cell‐free DNA test for CRC screening. Sensitivity of this test for diagnosing CRC is reasonable (83.1%) but crucially it is very poor at picking up advanced precancerous lesions (sensitivity 12.5%–13.2%) [15]. Raman spectroscopy (RS) is a chemical analysis technique based upon the interaction of light with chemical bonds. It analyses molecular structure, composition and intermolecular interactions to create a unique ‘molecular fingerprint’ of the individual sample and can be performed on biological tissue and fluid samples [16]. Machine learning (ML) analysis can be applied to Raman spectra in order to provide predictive models for potential screening and diagnostic purposes. While this technology has been applied to various cancers, research on CRC is limited. Small studies suggest promising sensitivity rates ranging from 71% to 100% and specificity ranging from 69% to 100% [17, 18, 19, 20]. In comparison to other forms of testing RS also offers the potential advantage of being minimally invasive, having simple sample collection and preparation, being easy to transport and store while being potentially cost‐effective [21]. In this exploratory study we aimed to assess the accuracy of plasma‐based RS for the detection of CRC. Methods Patient Recruitment and Ethics Participants were recruited from three hospitals in the Hunter New England Local Health District. Participants were recruited from two cohort groups. The first cohort consisted of participants with CRC undergoing surgical resection, while the second cohort consisted of participants undergoing a screening colonoscopy after having an iFOBT positive result. Only those participants who had sufficient plasma and information were included in the final analysis. Written informed consent and blood were collected from all participants in both cohorts prior to their procedures. This study was approved by the Hunter New England Human Research Ethics Committee (2023/ETH00536). Clinical Specimens Blood specimens were collected from participants immediately prior to the procedure being performed. Blood was then processed and stored in separate aliquots at −80°C (full details in Supporting Information 1). A certified pathologist confirmed the subsequent histopathological examinations and CRC was staged using the TNM system as defined by the Union for International Cancer Control (UICC) [22]. RS and ML Technique Plasma samples were analysed using the Hone Lab Black (Hone, Broadmeadow, NSW). The wavenumber range was 200–2500 cm−1 with a resolution of 13 cm−1. The laser wavelength was 1064 nm and the laser power was 450 mW. Square quartz micro cuvettes with 10 mm laser pathlength and 1 mm laser aperture were used with a minimum of 200uL of plasma and analysed immediately after pipetting into the cuvette and placement inside a covered compartment within the RS machine. Analysis was performed in triplicates with an integration time of 10 s. All Raman spectra underwent data pre‐processing. The ML‐model was then developed using a randomised allocation of participants into training and validation cohorts (full details in Supporting Information 1). Statistical Analysis Statistical analysis was performed using Hone Create (v2.1) and SPSS (v28). Participants were analysed in two separate subgrouping methods titled Screening and Principal Diagnosis (PD). The Screening method grouped the cohort simply into those with CRC and those with no CRC. The PD method used the subgroups: CRC, advanced adenoma, non‐advanced adenoma, colitis and no evidence of disease. The Raman spectra of all participants was initially obtained and then randomised into either the training cohort, to develop a predictive model, or the validation cohort, to which the model was applied and the accuracy tested. The randomisation allocation into the training and validation cohorts was 60:40 for the Screening model and 65:35 for the PD model. Intergroup differences in demographics such as age, gender, body mass index (BMI), smoking status, immunosuppression and Charlson comorbidity index (CCI), and histopathological features such as pathological stage, T‐stage, nodal involvement, M‐stage and lymphovascular invasion (LVI) were measured using Pearson's chi‐squared test and Fishers' exact test (significance p < 0.05). The sensitivity and specificity were calculated using cross tabulations. Results Between 2017 and 2021 there were 584 participants recruited for biobank collection from the two cohort groups. Sufficient plasma for RS analysis was obtained from 370 participants, with most exclusions being as a result of insufficient plasma for analysis (Figure 1). Flowchart of participant recruitment. Demographic characteristics of the entire cohort and its subgroupings for each model are shown in Table 1. In the Screening model, the CRC cohort was significantly older, showed a higher proportion of females and NSAID use, had a higher CCI and were more likely to have quit smoking than be currently smoking compared to those with no CRC. However, there was no significant difference found when specifically comparing the training and validation cohorts for all the demographic and histopathological criteria in the Screening model. In the PD model, there was a significant difference between the subgroups in terms of CCI, gender and smoking status. However, except for LVI, there was no significant difference found when specifically comparing the training and validation cohorts for the demographic and histopathological criteria in the PD model (Supporting Information 1). TABLE 1Demographics of participants.No. casesAge (years)Gender, n (%)BMICCINSAID use, n (%)Immunosuppression, n (%)Smoking status, n (%) N MedianWomenMenMedianMedianNoYesNoYesNon‐smokerEx‐smokerSmokerEntire cohort—screeningNo cancer25363129 (51)124 (49)292214 (85)39 (15)245 (97)8 (3)123 (49)72 (29)58 (23)Cancer1177043 (37)74 (63)30588 (75)29 (25)109 (93)8 (7)54 (46)52 (44)11 (9) p a < 0.0010.0110.291< 0.0010.030.106< 0.001Entire cohort—principal diagnosisCancer1177043 (37)74 (63)29588 (75)29 (25)109 (93)8 (7)54 (46)52 (44)11 (9)AA816333 (41)48 (59)28365 (80)16 (20)80 (99)1 (1)36 (44)24 (30)21 (26)NAA766541 (54)35 (46)28267 (88)9 (12)73 (96)3 (4)38 (50)20 (26)18 (24)Inflammation2612 (100)0 (0)2222 (100)0 (0)2 (100)0 (0)1 (50)1 (50)0 (0)NED946353 (56)41 (44)28280 (85)14 (15)90 (96)4 (4)48 (51)27 (29)19 (20) p b 0.9980.010.537< 0.0010.1540.4380.039Abbreviations: AA, advanced adenoma; BMI, body mass index; CCI, Charlson comorbidity index; NAA, non‐advanced adenoma; NED, no evidence of disease. a p‐value based on entire cohort with respect to cancer vs no cancer. b p‐value based on entire cohort with respect to principal diagnosis. Demographics of participants. Abbreviations: AA, advanced adenoma; BMI, body mass index; CCI, Charlson comorbidity index; NAA, non‐advanced adenoma; NED, no evidence of disease. p‐value based on entire cohort with respect to cancer vs no cancer. b p‐value based on entire cohort with respect to principal diagnosis. Training and validation set demographics are shown in Supporting Information 1. Histopathological characteristics for the Screening and PD models are found in Table 2 (full table in Supporting Information 1). TABLE 2Histopathology features of participants.Entire cohortValidation cohort, screeningValidation cohort, principal diagnosis N = 117% N = 43%Sensitivity (%) p a N = 40% p a Pathological stageT‐is65251000.803130.57913328112691112823227133092102534438174076184542200—00T‐stageT‐is65251000.435130.773T19837100410T23127133085923T357492354832152T4141225100512Nodal involvementNo71612661930.97122550.239Yes46391740761845M‐stageLocal1159843100860.531401000.431Met2200.0—00LVINo80682865820.56321530.012Yes37321535931947Abbreviations: LVI, lymphovascular invasion; M‐stage, metastasis stage; T‐is, tumour in situ; T‐stage, tumour stage. a p value based on entire cohort with respect to training and validation cohort. Histopathology features of participants. Abbreviations: LVI, lymphovascular invasion; M‐stage, metastasis stage; T‐is, tumour in situ; T‐stage, tumour stage. p value based on entire cohort with respect to training and validation cohort. The differences between groups in the Raman Spectra produced by the RS machine are shown. Figure 2 shows the derivative of the averaged Raman spectra produced in cancer and non‐cancer patients with Raman shift values included. Figure 3 shows the derivative of the averaged Raman spectra between cancer and non‐cancer patients with mean and standard deviation included. These figures illustrate the difference in spectra seen between the groups, which are of key importance to the predictive model. Derivate graph of averaged Raman spectra with Raman shift values—cancer versus non‐cancer. Derivative graph of averaged Raman spectra with difference and standard deviation—cancer versus non‐cancer. Results for the Screening model training and validation cohorts are represented in Figure 4 and Supporting Information 1. When the Screening model was applied to the validation cohort there was an 86% sensitivity (37/43), 93% specificity (98/105) and an overall accuracy of 91% for the prediction of CRC. The model showed a high positive and negative predictive value: 84% and 94%, respectively. When performing a subgroup analysis on the accuracy of the Screening model there was a high sensitivity and specificity maintained across the spectrum of pathological stage, T‐stage, nodal status, LVI and metastatic disease (Table 2). Screening model validation cohort results. When the PD model was applied to the validation cohort a total of 88% (35/40) of the participants with CRC were accurately identified (Table 3). In total, 42% (11/26) of participants without CRC, 46% (13/28) of participants with NAA and 14% (5/35) of participants with AA were accurately predicted. TABLE 3Results of principal diagnosis model when applied to validation cohort.Actual resultsCancerAANAAInflammationNEDTotalPredicted resultCancer35400039AA3550316NAA0121301237Inflammation000000NED2141001137Total403528026129Abbreviations: AA, advanced adenoma; NAA, non‐advanced adenoma; NED, no evidence of disease. Results of principal diagnosis model when applied to validation cohort. Abbreviations: AA, advanced adenoma; NAA, non‐advanced adenoma; NED, no evidence of disease. Discussion This study provides proof of concept for the use of label‐free RS plasma‐based screening tests for screening CRC. To our knowledge this is the largest study to date utilising RS and ML models to investigate this possibility using screened and CRC patients. We employed a multivariate statistical method integrating the entire Raman Spectra to determine significant spectral locations for the predictive model, resulting in an 86% sensitivity (37/43), 93% specificity (98/105) and an overall accuracy of 91% for the prediction of CRC. The model showed a high positive and negative predictive value: 84% and 94%, respectively. Similar methodologies previously report sensitivities ranging from 71% to 100% and specificities of 69% to 100% [17, 18, 19, 20, 23, 24, 25, 26, 27]. However, only a small proportion of these studies investigated the use of label‐free RS. Additionally, our study has more than twice the number of participants recruited to power the RS model being developed when compared to previous similar studies. This is excluding Jenkins et al. which was the only study identified with a larger cohort size of 405. An important feature of the Screening model is that it demonstrates an accuracy for the diagnosis of CRC of similar or better performance than iFOBT [28, 29, 30]. Several limitations of the currently available multi target stool DNA (mtSDNA) tests and blood‐based molecular tests (e.g., Shield, Epi‐proColon, Colvera) for CRC are the cost and need for specialised equipment. For instance, estimates are that these other methods cost 5‐times to 20‐times the cost of iFOBT [31, 32, 33]. For this reason, iFOBT remains the most cost‐effective method. Comparatively, the application of the Screening model would be relatively inexpensive per test and portable to resource limited laboratories. The PD model is of particular importance to the feasibility of this potential screening test. The methodology tested here shows that the RS that are obtained from plasma can be immediately analysed using both the developed models for classification into likely subgroups. In the case of the PD model, the specific accuracy of the test decreases with respect to CRC and NED. However, if one considers that all tests that return a prediction of either CRC, AA, NAA or inflammation would be recommended to have a colonoscopy, then 95% (38/40) of the participants with CRC would receive a colonoscopy. Furthermore, if all participants that did not have a diagnosis of CRC are grouped together, then the model retained a high specificity at 95% (85/89). The sensitivity of the PD model is low for individual subgroups of adenomas. However, when viewed wholistically with those that would receive colonoscopy the detection rate for AA's would be 60% (21/35). The sensitivity of iFOBT for the detection of AA is highly variable amongst studies but ranges from 25% to 66% and therefore this RS model sits on the high end of that range [28, 29, 30]. Although, a true population‐based study needs to be conducted to ascertain the true sensitivity for adenomas as a screening test. Notably, 42% (11/26) of colonoscopy participants (all iFOBT positive) that had NED were predicted to have NED by the PD model. The question then arises if the combination of RS with iFOBT could reduce the number of negative colonoscopies that result from the screening programmes worldwide. Given this number is around 30%–50%, in Australia this equates to approximately 29 000–48 000 per year [34]. The cost‐estimate of a colonoscopy varies depending on what is included in the calculation as well as whether the procedure is done with Anaesthetist assistance or not. The values often refer specifically to the ‘fixed’ costs (e.g., endoscopists, anaesthetists, nurses, medications, colonoscope maintenance and equipment) and do not include the broader societal costs (e.g., lost productivity from time of work and potential complications) [35, 36]. Considering this, the combination of an RS test with iFOBT could reduce the NBCSP cost by $14.5–$44.5 million annually within the Australian system based on fixed cost alone ($1200–$2200) [37, 38]. Strengths and Limitations This is a retrospective case–control study that utilised stored samples from a previous study rather than specimens specifically collected for the purpose of RS analysis. Although strict protocols were used for the initial project in terms of sample handling, plasma separation and storage, it is possible that these confounders could affect Raman spectra obtained from the stored samples. Furthermore, given the complexity of plasma, RS spectra and ML, our cohort size may still be insufficient. Conclusion This exploratory study has demonstrated that Raman spectroscopy and machine learning have the potential to be used for CRC screening. Larger volume research and general population studies are required for further evaluation. Supporting information Table S1: Demographics of participants (screening model). Table S2: Demographics of participants (principal diagnosis model). Table S3: Histopathology of participants (all models).