Introduction

Both rheumatic musculoskeletal diseases (RMDs) and cancer occur frequently in the general population and these diseases can either arise simultaneously or consecutively in affected patients. Given an RMD prevalence of ~2% and a cancer lifetime risk of ~45%–50% in European countries, in addition to the tumour-promoting effects of chronic inflammation and certain therapeutic regimens, differential diagnosis poses a relevant clinical question.1–3 Additionally, rheumatic symptoms can develop as misleading paraneoplastic manifestations of malignancy.4 5 While the timely diagnosis of cancer is imperative for favourable prognosis, it may pose a considerable challenge in unclear clinical courses in patients with rheumatic diseases.

Both disease entities are associated with a dysregulated systemic immune response. As such, they lead to quantifiable remodelling of metabolism at the cellular, tissue and systemic levels. Several studies, using either 1H nuclear magnetic resonance (NMR) spectroscopy or mass spectrometry to quantify metabolomic changes, have highlighted their clinical potential in cancer management for (1) differentiation between malignant and benign lesions; (2) identification of ‘risk profiles’ to aid early diagnosis; (3) classification of different cancer subtypes and stages; and prediction of (4) treatment response in and, importantly, (5) overall survival in lung and gastrointestinal cancers, for example.6 7 Metabolomic profiling has also been applied to various autoimmune and autoinflammatory RMDs to define biomarkers for (early) diagnosis, refine classification, predict drug response or prognosis8–11 and differentiate selected neoplasms from RMD.12 13 Moreover, we have recently demonstrated a shared ‘autoimmunity’ signature in cancer patients with rheumatic immune-related adverse events (irAE) due to immune checkpoint inhibitor (ICI) therapy or concomitant RMD distinct from both cancer-free patients with RMD as well as irAE-free, ICI-treated patients with various cancers.14

In general, metabolomic studies focus on a single cancer type. However, an in vitro study and a neuronal network analysis of metabolic profiles in a large-scale population study suggest that metabolic profiles depend on the cancer entity.15 16 It has not yet been explored whether a metabolic signature can be used to detect cancer burden in general, particularly if patients have simultaneously have other chronic diseases like RMD. To date, a reliable universal biomarker sensitive and specific for any type of invasive cancer is not yet available for use in daily clinical practice. Thus, here we investigate whether an altered serum metabolome profile can serve as a potential biomarker of cancer in patients with RMD. If this can be demonstrated, it would facilitate early differential diagnosis of non-specific inflammatory symptoms as paraneoplastic versus RMD associated and thus speed up cancer diagnosis and treatment.

Materials and methodsPatient recruitment and sample collection

Patients included in the study suffered from either rheumatoid arthritis (RA) (by 2010 American College of Rheumatology (ACR)/European Alliance of Associations for Rheumatology (EULAR) classification criteria),17 psoriatic arthritis (PsA, by Classification criteria for psoriatic arthritis (CASPAR) criteria),18 other spondyloarthritis (SpA) subtypes (by Assessment of Spondylosrthritis International Society (ASAS) criteria)19 or systemic lupus erythematosus (SLE) (by 2019 EULAR/ACR criteria).20 Patients with PsA and other SpA subtypes were merged into a joint group due to their shared pathogenetic mechanisms. Serum samples were collected within the MalheuR project, a local registry-based study for patients with concomitant rheumatic and malignant diseases, or previously stored samples from intrainstitutional biobanks or other projects were used. The serum samples from RA, PsA/SpA and SLE patients with a history of early-stage, metastatic or cured invasive malignancies or obligatory solid precancerous lesions (mRA, mSpA and mSLE groups) and respective controls (cRA, cSpA and cSLE groups) were analysed in the main cohorts. Control samples were matched by RMD diagnosis and age. All patients were recruited at scheduled rheumatological on-site evaluations and non-fasted serum samples were obtained during routine blood collections. Treatment of RMD and concurrent malignancy followed existing guidelines and recommendations and was not influenced by study participation. A blinded validation cohort (VC) was formed by samples from seven patients with and nine patients without invasive malignancy not included in the main cohort. We tested model performance in study populations of special interest with active cancer or cancer treatment (n=66 and n=41, respectively, not featured in mRA, mSpA or VC cohorts), pulmonary type cancers (n=22) and lymphoid type cancers (n=18). Furthermore, samples from 8 patients with paraneoplasia (PN), 10 patients with facultative solid precancerous lesions and non-invasive (NI) precancerous lesions of the skin or non-melanoma skin cancer and 7 precancer (PC) samples from patients who developed systemic malignancy later after sample collection were included in the analysis. Exclusion criteria were either rheumatic diagnoses other than RA, SpA or SLE or monoclonal gammopathy of undetermined significance as malignant diagnosis without the presence of other cancer types. Patients with NI and non-melanoma skin cancers were only included in the NI cohort and not in any of the other cohorts.

Statistical analysis

Cohorts were described using appropriate measures of the empirical distributions. Differences were considered statistically different for p<0.05 in a descriptive sense. All analyses were performed using R V.>4.00.0 and SPSS (Build V.29.0.0.0 (241), IBM Corporation).

Univariate analysis

Details on the statistical tests used for the univariate analysis between groups can be found in online supplemental file 1.

Multivariate analysisMultivariate diagnostic model

In order to identify the combination of metabolites with the best predictive performance, logistic regression comprising forward selection was performed on subsampling 63.2% of the cohort 1000 times.21 Finally, five variables were selected to be part of the final model. Area under the receiver operating curve (AUC) was selected as the main parameter to describe the goodness of fit of the model.

Cross-validation

To get an unbiased estimated of the AUC fit index, 10-fold cross-validation was performed in triplicate.

Validation and further disease cohorts

To describe the diagnostic performance of the multivariate diagnostic model on additional cohorts, AUC values, accuracies, sensitivities and specificities were reported with respect to various probability cut-offs.

ResultsPatient characteristics

Demographic data and clinical characteristics of the model cohort consisting of RA patients with (mRA, n=56) or without a history of early-stage, metastatic or cured invasive cancer (cRA, n=52) are summarised in table 1. Major differences were the higher age (69.9±9.7 vs 56.1±12.0) and the larger proportion of women (76.8% vs 57.7%) in the mRA cohort. Mean rheumatic disease duration was >10 years in both cohorts (12.9±10.9 vs 10.25±8.08). Breast cancer (39.3%) followed by urogenital cancers (17.9%) and melanoma (12.5%) were the most frequent malignancies. The mean time since cancer diagnosis was 13.9±10.8 years with <1 year in 2 (3.6%), 1 to <5 years in 10 (17.9%) and >5 years in 78.6% of patients with mRA. 75.0% of patients only had only low stages (0–II) of malignancy and 96.4% had recorded complete remission (CR) at last staging. 12.5% of patients had received cancer treatment within six months of sample collection, while another 8.9% of patients were still receiving cancer treatment. While glucocorticoid (GC) use at blood sampling did not differ significantly between the groups, fewer patients with mRA received conventional synthetic disease-modifying antirheumatic drugs (csDMARDs) (66.1% vs 86.5%). The prescription of biological DMARD was comparable but with a more frequent use of tumour necrosis factor inhibitors in the cRA group (21.4% vs 30.8%). Interestingly, targeted synthetic DMARDs or abatacept were prescribed for only three patients with mRA (5.4% each) and not at all for patients with cRA. Altogether, more patients with mRA were in remission from rheumatic disease at sample collection as assessed by Disease Activity Score (includes tenderness and/or swelling of 28 joints, patient global assessment and C reactive protein) (70.4% vs 55.8%).

Table 1

Clinical and demographic characteristics of the study participants in the RA model cohort

A blinded VC was formed from newly recruited RA patients with (n=7) and without (n=9) a history of invasive malignancy (individual details in online supplemental table 1). A second VC included patients with SpA (mSpA, n=33; cSpA, n=75; table 2). As for the model cohort, patients with mSpA were older at blood sampling than controls (63.6±13.7 vs 51.9±11.8). Urogenital cancers (30.3%) and melanoma (21.2%) were more frequent, probably due to the higher proportion of man in the SpA model cohort compared with the RA model cohort. Malignancy characteristics regarding disease duration (10.5±9.4 years; <1 year: 2 (6.1%), 1 to <5 years: 8 (24.2%) and >5 years: 23 (69.7%)), initial stage (63.6% stage 0–II) and treatment (12.1% <6 months at sampling) were comparable to the RA model cohort. However, the proportion of CRs (81.8%) was lower for patients with mSpA compared with patients with mRA. GC treatment was applied more often for patients with mSpA than for patients with cSpA (39.4% vs 26.7%). The use of other DMARDs and clinical remission as determined by the physician did not significantly differ. Of note, a larger proportion of SpA than of patients with RA showed inflammatory disease activity at sample collection.

Table 2

Clinical and demographic characteristics of the study participants in the SpA cohort

Additionally, the metabolic prediction model was tested for cohorts of RA and SpA patients with active cancer cohort and/or treatment (ACC), pulmonary cancer cohort (PCC) and lymphoid cancer cohort (LCC), PN, NI cancer and PC samples from patients with a later cancer diagnosis and a cohort of patients with SLE with individual data listed in online supplemental tables 2-7 and 11.

Univariate analysis and metabolic pathway correlation

We analysed serum lipids and metabolites to identify differences in metabolism between RMDs with and without concurrent malignancy. The serum lipids and 28 low-molecular weight metabolites were identified and quantified using one-dimensional NOESY and Carr-Purcell-Meiboom-Gill (CPMG) 1H NMR spectra as previously described (online supplemental figure 1A).9 Differences in the concentrations of individual metabolites between mRA and cRA groups were determined by univariate analysis. Out of 28 metabolites and 5 lipid ratios, significant differences were found in the serum concentrations of glucose, various amino acids (A, R+K, N, D, H, I, L, P, T and V), various carboxylic acids (acetate, citrate, creatine, isobutyrate, pyruvate and succinate) and the normalised lipid moieties (L4+L5)/L1, L6/L1 and L7/L1 (figure 1A and table 3). These differences remained even after the patients with mRA were stratified according to whether the cancer was still active at the time of blood collection, they had been under therapy for more than 6 months or were in remission and considered cured for over 5 years (online supplemental figure 1B,C). Acetate, histidine and valine rendered area under the curve (AUC) values>0.90 in receiver operating characteristic (ROC) analyses, while 10 other metabolites yielded AUC>0.80 (table 3).

Figure 1

Metabolite concentrations and lipids ratios differ between rheumatoid arthritis (RA) patients with or without a history of invasive malignancy. (A) The volcano plot indicates fold changes and adjusted p values of metabolite concentrations and lipid ratios between the two patient groups. The horizontal dotted line indicates an false discovery rate (FDR)<0.05. (B) Summary of quantitative metabolic pathway enrichment analysis showing the changes between mRA and cRA metabolomes. (C) Correlograms showing Spearman correlation coefficients between the continuous clinical/demographic variables and the metabolites for the mRA and cRA groups. *p<0.05, **p<0.001 and ***p<0.0001. cRA, RA without cancer; mRA, RA with a history of invasive cancer.

Table 3

Univariate analysis results for the metabolites and lipid ratios

Specific metabolic pathways involving pyruvate, carbohydrate and amino acid metabolism were increased in the cRA group in comparison to the mRA group (figure 1B). The influence of clinical, therapeutical or demographic parameters on metabolome profiles was tested using multivariate analysis of variance (MANOVA), but for the majority, significant correlations with metabolite concentrations or lipid ratios were not evident (figure 1C and online supplemental table 8). Since cachexia is considered a cancer syndrome, we tested by partial least squares-differential analysis (PLS-DA) whether differences in body mass index (BMI) could result in different metabolomic profiles. However, PLS-DA could not separate the metabolome of underweight patients from the others (online supplemental figure 1D). Additionally, a PLS-DA with up to five principal components based on the type of cancer of the mRA group did not show a significant separation between patients (maximum accuracy: 39.7%; R2=0.215; Q2=−0.170; online supplemental figures 1E,F).

Multivariate diagnostic model

Despite acceptable ROC values for certain metabolites or normalised lipid moieties, we considered that univariate analysis may be prone to interference in the complex clinical situation of cancer burden and treatment in a long-term-treated inflammatory disease. Therefore, to increase diagnostic accuracy and robustness for clinical application, we applied multivariable, logistic regression modelling on the metabolomic and lipidomic profile together with the clinical and demographic data. The stepwise forward selection subsampling algorithm resulted in the following parameters as predictors in the diagnostic model: the concentrations of acetate, creatine, formate and glycine and the L1/L6 ratio. The coefficient estimates are presented in table 4.

Table 4

Coefficient estimates for the model

To classify patients into the two diagnosis groups, a cut-off value for the predicted probabilities was applied. Using the default cut-off value of 0.5 initially, a subject was classified with a malignancy if the estimated probability is >0.5 and thus cancer diagnosis more likely. ROC analysis of the diagnostic model for patients with RA yielded an excellent AUC=0.995 (figure 2A), and 10-fold cross-validation (performed in triplicate) resulted in an AUC=0.987 with a sensitivity of 0.932 and a specificity of 0.946.

Figure 2

The multivariate diagnostic model performs well on an independent validation cohort (VC). (A) Receiver operating characteristic (ROC) curve for the modelled probability of invasive cancer burden in the model cohort. (B) ROC curve for the modelled probability of invasive cancer burden in the blinded VC. (C) Heatmap comparing the metabolite expression levels between the model (rheumatoid arthritis (RA) without cancer (cRA) and RA with a history of invasive cancer (mRA)) and the validation (validation for RA without cancer (cVC) and validation for RA with a history of invasive cancer (mVC)) cohorts. (D) Dot plots comparing the concentrations of the metabolites and the L1/L6 ratio used in the diagnostic model between the model (cRA and mRA) and the validation (cVC and mVC) cohorts.

For validation of the model, a blinded cohort of seven new RA patients with a history of invasive malignancy and nine new patients without cancer was collected (VC, online supplemental table 1). Applying the diagnostic model, 14 of 16 patients (87.5%) were correctly classified (online supplemental table 9) and ROC analysis rendered an AUC of 0.937 with a sensitivity of 0.889 and a specificity of 0.857 (figure 2B). Overall, most metabolite expressions were comparable between the model and VCs for the corresponding categories (figure 2C). Importantly, the concentrations of the metabolites and the L1/L6 ratio used in the diagnostic model were comparable between the model and the VCs for the corresponding categories (figure 2D). Moreover, except for the difference in glycine concentration between female and male patients with mRA, clinical and demographic variables did not seem to influence the concentrations of the selected metabolites (online supplemental figure 2A-F).

Transfer to other diseases

To assess if the diagnostic model can be transferred and universally applied to other RMDs, we first tested it on an SpA cohort of 55 malignancy (mSpA) and 75 control (cSpA) patients. ROC analysis of the diagnostic model yielded an AUC=0.912 (figure 3A). Using a default cut-off value of 0.5, 104 of 130 (80.0%) SpA samples were correctly assigned, with a sensitivity of 0.964 and a specificity of 0.680. There were some differences for a number of the metabolites between the RA model cohort and the SpA cohort (figure 3B). Importantly, the concentrations of the metabolites and the L1/L6 ratio used in the diagnostic model were comparable between the RA model and the SpA cohorts for the corresponding categories (online supplemental figure 3). As for the RA model cohort, a MANOVA confounder analysis did not suffer any significant influence from the clinical and demographic variables in the metabolome differences observed between the cSpA and mSpA groups, especially for those metabolites used in the model (online supplemental table 10). Also, differences between the groups were detectable irrespective of the SpA disease activity (figure 3C). Additionally, we explored the effects of cancer treatment type as a possible confounder of the metabolic signature in patients with RA and SpA. Importantly, we have not observed significant differences between patients treated by any type of antineoplastic drug or radiation compared with those who only underwent surgery or were untreated for their cancer at the time of sample collection (figure 3B).

Figure 3

The high performance of the multivariate diagnostic model persists in the SpA cohort with higher probability cut-offs. (A) Receiver operating characteristic (ROC) curve for the modelled probability of cancer burden in the SpA cohort. (B) Heatmap comparing the metabolite expression between the model (cRA and mRA) and the SpA (cSPA and mSPA) cohorts. Both cancer cohorts were stratified according to whether patients were treated with any type of antineoplastic drug or radiation (T) compared with those who only underwent a surgery or were untreated for their cancer at the time of sample collection (NT). (C) Heatmap comparing the metabolite expression in the SpA cohort (cSPA and mSPA) subdivided into rheumatic disease activity: high (H) or low (L) active and remission (R). Predicted probabilities of control (CNT) and malignancy (MAL) patients in the merged RA/SpA cohort summarised in a box plot graph (D) and a frequency plot (E). RA, rheumatoid arthritis; SpA, spondyloarthritis.

Patients with active cancer, pulmonary or lymphoid cancer types are subpopulations of special interest for daily rheumatology practice. In the ACC (online supplemental table 2), the diagnostic model accurately assigned cancer diagnosis in 14/15 untreated patients with active cancer (the one patient not recognised correctly was already included in the blinded VC in online supplemental table 1), all 24 patients with active cancer with previous (eg, due to progression or a history of another cancer) or current cancer treatment and all 43 patients with current cancer treatment irrespective of remission status at sample collection. Furthermore, all 22 patients in PCC (online supplemental table 3) and 17/18 patients in the LCC (online supplemental table 4) were correctly classified as cancer by the model.

We next assessed if the diagnostic model was also suitable for oncological settings other than verified early-stage, metastatic or cured invasive cancer. The diagnostic model accurately assigned cancer diagnosis for all patients (8/8) with PN (online supplemental table 5), out of which 5 were not previously treated for their cancer at sample collection. In contrast, only 50% (5/10) of patients with NI or in situ precancerous lesions and non-melanoma skin cancers (NI, online supplemental table 6) were correctly classified, though two melanoma in situ patients with a high risk of progression to an invasive cancer stage were correctly classified. Finally, the model classified all (7/7) patients (PC, online supplemental table 7) as cancer free who were diagnosed with an invasive malignancy only long after (5–56 months) sample collection. Consequently, while use of the diagnostic model can correctly identify cases of systemic cancer forms already present in a patient—irrespective of whether or not they have been diagnosed at sample collection – it is not able to identify patients at risk of later invasive cancer development.

To achieve higher sensitivity or specificity, we tested cut-off values other than 0.5 for the model. Preferably, the diagnosis of a malignancy should only be assigned if the probability of it is very high (cut-off≥0.9) or, conversely, malignancy should ideally only be ruled out if its probability is very low (cut-off≤0.1). Samples from a merged RA and SpA cohort consisting of patients from the cSpA, mSpA, VC, ACC, PCC, LCC, PN, NI and PC cohorts, but not from the model RA cohort initially used for model formulation, were used for these considerations. The model rendered an AUC=0.927 for the merged cohort. For a 0.5 cut-off, 84.0% of samples were classified correctly with an excellent sensitivity of 0.958 but with a lower specificity of 0.717. Using a 0.9 cut-off, the accuracy of malignancy diagnosis was increased to 86.1% with a sensitivity of 0.916 and a specificity of 0.804. A 0.1 cut-off had only a low accuracy of 78.1%, with a comparably high sensitivity of 0.968 but with a lower specificity of 0.587 for cancer diagnosis (table 5).

Table 5

Comparison of the model performance for the merged RA/SpA cohort with different cut-off values 0.9, 0.5 and 0.1

Two aspects become apparent from the distribution of the model value read outs for control and malignancy patients (figure 3D, E): the superiority of a higher cut-off for group separation is indicated by more than 75% of the predicted probabilities for control patients located from 0 up to approximately 0.8. Three malignancy patients, including one incorrectly classified patient in the blinded VC, were not detected by any cut-off value>0.1 and one patient with lymphoma was not detected by cut-off values>0.5. A clinical re-evaluation of these cases shows that all three failed diagnoses (two with renal carcinoma and one with melanoma) had low-grade tumours in the tumour, node, metastases classification and two had their cancer diagnosis and permanent treatment by excision more than 10 years before sample collection. The stage II lymphoma patient with a low model readout (0.26) was diagnosed 8 years earlier and in complete therapy-free remission for more than 5 years at sample collection. The model was, nevertheless, able to classify other patients with low-grade renal carcinoma and/or melanoma correctly (n=6) even if diagnosis had been made several (4.9–25.8) years previously.

Finally, we tested the model for SLE as a rheumatic disease with a fundamentally different pathogenic mechanism. Using samples from 10 patients with mSLE and 32 patients with cSLE, the model rendered an AUC=0.656 with only 38.1% accurately classified by the model, though with a high specificity (0.800) but with a very poor sensitivity (0.250) at a cut-off of 0.9 (online supplemental tables 11 and 12 and online supplemental figure 4A). Since RA and SLE are clinically distinct, it was not surprising that the concentrations of the metabolites and lipid ratios differed between the cRA and cSLE groups and between the mRA and mSLE groups (online supplemental figure 4B).

Discussion

A timely diagnosis of cancer is imperative for successful therapy. In patients with an ongoing autoimmune condition, cancer diagnosis can be delayed due to the similarity of symptoms. In this study, we discovered a set of biomarkers that predict the presence of malignant disease in arthritis or PN patients with high sensitivity. This limited-invasive assay can expedite up cancer diagnosis and might be especially applicable for patients with a high risk for cancer development such as lymphoma in RA.22 To our knowledge, the coincidence of cancer and other chronic diseases has not been included in the scope of metabolome research thus far, despite the proven value of NMR spectroscopic and mass spectrometric analyses to identify biomarker candidates separately in both disease entities.6–9 12 13 16 Here, we show for the first time that despite the systemic remodelling driven by both conditions, a common serum metabolic signature of cancer can be detected and differentiated from cancer-free profiles in patients simultaneously suffering from chronic autoimmune arthritis.

Metabolic pathway enrichment analysis revealed differences between patients with cRA and mRA in several pathways associated with nutrient oxidation and interconversion of metabolic intermediates to supply biosynthetic demands. Among such metabolic pathways, those associated with amino acids (eg, ammonia recycling, the urea cycle and multiple other amino acid metabolic pathways) and carbohydrates (eg, glycolysis and amino sugar metabolism) appear to have been particularly enriched in patients with cRA. Several serum metabolome analyses have reported a systemic glycolytic signature in RA.9 However, similar metabolomic studies on patients with cancer did not identify glycolysis as a significantly enriched metabolic pathway even though a wide variety of tumour cells are markedly glycolytic.16 23 24 It is likely that the coexistence of the two pathologies in patients with mRA results in a significant reduction in glucose plasma levels but without a concomitant elevation of plasma lactate levels. This may be explained by the finding that in tumours, lactate produced by cells in the markedly hypoxic environments feeds the oxidative metabolism of cells under normoxia.25 Additionally, the high biosynthetic rates associated with tumour growth may contribute to a more pronounced depletion of amino acid plasma pools in patients with mRA compared with patients with cRA. In such circumstances, the tricarboxylic acid (TCA) cycle in tumours assumes a marked anabolic character allowing the interconversion of TCA cycle intermediates to feed important biosynthetic requirements.26 Recent studies identified positive roles for arginine in preventing inflammation and osteoclastogenesis in arthritis and measured higher serum levels of arginine and ornithine in patients with RA.27 28 Unfortunately, one limitation of using NMR metabolomics is that the resonances used to identify arginine in 1H NMR spectra overlap with those from lysine.29 Therefore, we could not specifically quantify the serum concentration of arginine.

Although significant differences were detectable for 20 metabolites and lipid ratios by univariate analysis, we consider that the use of single parameters as biomarkers in such a complex clinical situation may be extremely prone to interference by demographic and disease-associated and treatment-associated factors.30 Hence, metabolomic and lipidomic profiles of the model cohort were subjected to multivariate logistic regression analysis to achieve robustness and high diagnostic accuracy in clinical settings. Counterintuitively, none of the top-ranked univariate parameters and no clinical variables were included in the five variables selected for the final diagnostic model (age and gender were only selected in the sixth and seventh positions, respectively). ROC analysis of the model based on the concentrations of four metabolites and one lipid ratio combined with respective coefficient estimates yielded an excellent AUC. Importantly, the high diagnostic performance persisted in the blinded VC of newly recruited patients with RA, in the SpA cohort and in study populations of special interest with active and even untreated cancer, active cancer treatment, pulmonary and lymphoid type cancers with even higher probability cut-offs indicating that the metabolite patterns could indeed serve as a diagnostic tool in arthritides. The model did not identify three early-stage, low-grade tumours and was not reliable for NI facultative precancerous lesions and non-melanoma skin cancers. The limited invasiveness at sampling can be a potential reason for this. However, these cancer types pose a less serious threat compared with the ones that the model picked up with high sensitivity. Most notably, in the NI cohort, two melanoma in situ cases associated with a high inherent risk of later-course metastatic spread were correctly classified by the model.

Even in the high-risk population studied by Buergel et al,16 the metabolomic state did not provide predictive information for the six cancer types analysed. We hypothesise that significant alterations in an individual’s systemic metabolome only occurs when cancerogenesis trespasses from in situ to the systemic level and does not present a common prominent risk signature ahead of this timepoint. Once cancerogenesis has reached the systemic level by its invasiveness, our diagnostic model can aid differential diagnosis, particularly for non-specific symptoms like fever, night sweat, weight loss or suspected PN. The atypical manifestation of the latter often mislead physicians and the resulting delay in cancer diagnosis and treatment is associated with an unfavourable prognosis.4 5 The potential to facilitate a rapid and correct diagnosis of malignancy in rheumatic paraneoplastic syndromes or the presentation of new symptoms suspicious of concomitant cancer in patients with RA and SpA is a major success of our model.

Collectively, the applicability of the current model encompasses around two thirds of RMD cases.1 Given the poor AUC performance and markedly different metabolomic profile of the SLE cohort, the current model is unsuitable for universal application in more systemic RMD entities. Unfortunately, the limited number of mSLE samples in our study did not allow us to clarify whether a different diagnostic model performs better at recognising cancer burden in SLE.

A major strength of our model is its insensitivity to common demographic and clinical confounders such as age, BMI, disease activity of both cancer and RMD as well as treatment type and status, making it reliably applicable in a very heterogenous study population. Furthermore, its coverage of a large spectrum of invasive cancer types at different stages also adds to its robustness in heterogenous clinical cohorts.

Before translation into clinical practice, our diagnostic model still requires validation in a larger multinational and multiethnic cohort to engender reliability for genetically heterogenous populations. Thereafter, this limited-invasive assay has considerable potential of high clinical value to facilitate timely diagnosis of cancer in paraneoplastic rheumatic syndromes as well as become a valuable active surveillance tool in RA and SpA patients with a high risk of developing cancer.