External Validation of Plasma Glycosaminoglycans as Biomarkers to Improve Lung Cancer Risk Stratification

Abstract

Background:

Lung cancer screening excludes individuals not considered at an increased risk for lung cancer, as predicted by risk models like the Liverpool Lung Project version 3 (LLPv3). In this study, we sought to validate whether plasma glycosaminoglycan profiles (GAGomes) could predict lung cancer independent of LLPv3 and other prespecified comorbidities.

Methods:

In this retrospective cohort-based case–control study, we included patients who were suspected of having lung cancer at baseline and were either diagnosed with lung cancer (cases) or remained cancer-free for 5 years after baseline (controls). Plasma GAGomes were measured at baseline and used to compute a prespecified GAGome score to discriminate lung cancer from controls. We then applied multivariable Bayesian logistic regression to evaluate the likelihood that 7 LLPv3 predictors or 14 comorbidities had an effect on the GAGome score. We tested the independence of the GAGome score from LLPv3-predicted 5-year risk using the likelihood ratio test and assessed whether it improved lung cancer risk prediction in a set equivalent to an LLPv3-predicted 5-year risk of ≥1.51%.

Results:

We included 653 lung cancer and 653 controls. The AUC of the GAGome score was 0.63 (95% confidence interval, 0.62–63). None of the LLPv3 predictors or comorbidities were compatible with a significant effect on the score. The GAGome score was independent of LLPv3 (P < 0.001) and improved its sensitivity (72% vs. 69%) and specificity (61% vs. 59%).

Conclusions:

Plasma GAGomes identified additional lung cancer cases beyond those predicted by LLPv3 alone.

Impact:

GAGomes could improve risk-stratified lung cancer if validated in a screening population.

Introduction

Lung cancer screening reduces lung cancer–specific mortality (1), but unlike other cancer types, eligibility is based on risk factors beyond age, like smoking history. Risk-stratified screening simplifies the program, reducing false positives and potentially limiting overdiagnosis and overtreatment (2). Various risk models have been developed to identify high-risk individuals eligible for lung cancer screening. These range from simple criteria like those proposed by the US Preventive Services Task Force (3) to more complex models tailored to specific regions, such as the Liverpool Lung Project version 3 (LLPv3) in the United Kingdom (4). These models typically combine demographic, lifestyle, and medical or family history variables.

Biomarkers are not routinely used for risk prediction, though they could, in principle, identify individuals with increased risk, even when conventional risk models classify them as “low-risk.” Therefore, biomarkers could provide access to screening in the primary care settings for those otherwise excluded (e.g., due to age). However, several challenges hinder the integration of biomarkers into risk prediction, primarily the need for robust performance validation and the risk of confounding from other comorbidities. Many lung cancer biomarkers have been proposed, but most have not been validated without modifications in fully independent external populations (5). For example, a 7-protein panel (6) or a 4-miRNA panel (7) both showed high performance in validation cohorts; yet the first study recalibrated the panel prior to validation, and in the second study, the validation cohort was not fully independent of the discovery cohort—factors likely contributing to optimistic performance. Another example is a 13-protein panel that reported a validated AUC = 0.60 but in which the only fully independent validation cohort consisted of 37 samples (8). Among lung cancer biomarkers with external validation, a prominent example is a score developed based on a blood-based panel consisting of the precursor form of surfactant protein B, cancer antigen 125, carcinoembryonic antigen (CEA), and cytokeratin-19 fragment (CYFRA; ref. 9), which was later validated in a large independent cohort (10). However, comorbidities such as chronic obstructive pulmonary disease (COPD) can cause elevated CEA (11) or CYFRA (12) levels, underscoring the need to assess biomarker-based risk predictions in patients with conditions that could contribute to lung cancer risk or act as confounding factors, leading to false positives.

We previously found that changes in plasma, glycosaminoglycan profiles (GAGomes) were associated with a future cancer diagnosis, including lung cancer (13). Glycosaminoglycans (GAG) are long unbranched heteropolysaccharides with complex sulfation patterns that promote cancer-related pathways (14). In an earlier study, we developed a GAGome score that differentiated lung cancer from cancer-free patients who were initially suspected of having lung cancer (15).

In this study, we sought to validate the same GAGome score to distinguish lung cancer as well as other non-lung cancers from cancer-free controls in a large, external cohort from the LLP. All samples were collected during investigations for lung cancer following symptomatic referral, in which confounding comorbidities and differential diagnoses are challenging. We selected a symptomatic cohort rather than a screening cohort for two reasons that are often overlooked in biomarker studies (5). First, this design allowed us to assess whether the GAGome score was confounded by specific comorbidities that are too rare in screening populations (e.g., rheumatoid arthritis; ref. 16) or differed in association with rare forms of lung cancer (e.g., small cell lung cancer). The rarity of these conditions would require extremely large screening cohorts. Second, it allowed us to test the utility of GAGome scores outside the population currently eligible for screening. In this cohort-based case–control design, we could determine whether adding the GAGome score to reestimate LLPv3’s 5-year predicted risk improved the accuracy of classifying patients as “high-risk” for lung cancer (defined as a 5-year risk ≥ 1.51%, equivalent to the 2.5% cutoff used in LLPv2 for screening eligibility in the UK NHS Targeted Lung Health Check; refs. 4, 17).

Materials and Methods

Study design

This retrospective cohort-based case–control study aimed to include all ≥35-year-old adults (any sex, any smoking history, and any comorbidity) initially symptomatic for lung cancer and consenting to be part of the LLP Biobank, Liverpool, United Kingdom. Patients either had a lung cancer diagnosis within 1 year after the baseline visit or underwent surgery for a diagnosed lung cancer (lung cancer cases) or were cancer-free at the baseline visit but were later diagnosed with cancer in the next 5 years [“other cancer” (OC) cases] or were found to be cancer-free for the 5 years after the baseline visit (controls). The “OC” cases were included per protocol as an additional group of cases because changes in the plasma GAGome have been previously associated with incident any-type cancer (13), potentially elevating the GAGome score already at the baseline visit. Cancer was defined as lung cancer [International Classification of Diseases for Oncology, Third Edition codes = C33–34 (lung cancer) or C45 (mesothelioma)] or OC [International Classification of Diseases for Oncology, Third Edition codes = any C00–96 except lung cancer and C44-4A (nonmelanoma skin cancer)]. The baseline visit was defined at the biospecimen collection visit.

Eligible participants were initially identified among all ≥35-year-old patients with available EDTA-plasma collected after January 1, 2004 in the LLP Biobank with ≥5 years of follow-up after the baseline visit. Key inclusion criteria were not receiving treatment for cancer at the baseline visit and no history of cancer in the 5 years before the baseline visit unless curatively treated. Exclusion criteria were cancer diagnosis based on self-reported information only, <35 years at the baseline visit, and <5 years of follow-up. All OC cases and controls in the LLP Biobank were included, whereas lung cancer cases were a random sample.

The reference standard to determine presence of cancer was a diagnosis recorded in the electronic health medical record and assessed based on clinical, pathologic, multidisciplinary team findings or registry information from the date of the baseline visit to the date of follow-up end (December 2023; July 2023 for registry data) or the date of death.

The study was approved by the Swedish Ethical Review Authority (no. 2023-05407-01). Patient written informed consent was originally obtained by the University of Liverpool, United Kingdom, for participation in the LLP Biobank with ethical approval obtained from the Liverpool Central Research Ethics Committee (ref. 97/141) in accordance with the Declaration of Helsinki. The study was reported in compliance with the STARD guidelines, as applicable (Supplementary Table S1). It was preregistered as a substudy named LEVANTIS-0019 Hospital Cohort within the LEVANTIS-0087A study (clinicaltrial.gov; identifier, NCT05235009) sponsored by Elypta AB. The sponsor participated in the study design, execution, analysis, and review of the final manuscript. The study protocol was not amended after data unblinding (Supplementary Materials and Methods).

GAGome score

EDTA-plasma was collected between January 2004 and December 2019 at room temperature in either scintillation vials (Monovette EDTA, product no. 01.1605.001, Sarstedt) or microcentrifuge tubes (K2EDTA separator, product no. 455040, Greiner Bio-One), kept at 4°C until centrifugation at 1,800 to 2,200 rpm for 10 to 15 minutes at room temperature within 24 hours and stored at < −75°C until analysis. GAGome measurements were performed in singleton by a single blinded central laboratory using kits from a single lot (MIRAM GAGome Kit, product no FRUOV2, Elypta AB), wherein the absolute concentration (in µg/mL) of 17 chondroitin sulfate, heparan sulfate, and hyaluronic acid disaccharides were quantified using an ultra-high-performance liquid chromatography with triple-quadrupole mass spectrometry system (Waters Acquity I-Class Plus Xevo TQ-S micro, Waters Corporation) following the kit instructions for use. The analysis protocol was previously described (18, 19). Outlier samples were identified and excluded per protocol. The plasma GAGome was then used as input to compute the previously published plasma LC GAGome score (simply referred to as the GAGome score; ref. 15), with no changes to the underlying logistic regression model. The assessors of the GAGome score were blinded to the reference standard assessment, and vice versa.

Statistical analysis

Receiving operating characteristic analysis

The discrimination of cases versus controls by the GAGome score was evaluated by the area under the ROC curve (AUC). Internal validation was performed using bootstrap resampling (N_boots = 1,000) to correct for optimism by subtracting the change in statistic in the test versus training set from the observed statistic. The so-corrected statistic was reported together with the empirical 95% confidence intervals (CI). The observed AUC was evaluated in the following subsets: age (<65 vs. ≥65 years), sex (female vs. male), comorbidity (see below), histopathologic cancer type, and tumor–node–metastasis (TNM) stage (aggregated by stage I vs. II vs. III vs. IV depending on type).

Bayesian multivariable linear mixed-effect regression model

We used a Bayesian linear mixed-effect regression model to test the probability of an association between the GAGome score (response variable) and each of the following variables (explanatory variables) while adjusting for the GAGome analysis batch (group variable): each outcome [lung cancer (binary) and OC (binary)]; each LLPv3 predictor [age (continuous, in years, standardized, no missing data), sex (binary, no missing data); personal history of cancer (binary, no missing data); personal history of pneumonia, emphysema, bronchitis, tuberculosis, or COPD (binary, no missing data); personal history of exposure to asbestos (binary, 92 missing); family history of lung cancer (categorical: no (reference); yes, early onset; yes, late onset; no missing data); and smoking duration (continuous, in years, standardized; 31 missing)] and 14 preselected comorbidities [COPD, emphysema, bronchitis, asthma, hypertension, hyperlipidemia, bronchiectasis, chronic lung disease, gout, inflammatory polyarthropathy, noninfectious enteritis of the intestine, phlebitis, rheumatoid arthritis, and type 2 diabetes (binary)]. Missing data were imputed as evidence of absence (i.e., recoded as “no” or “never”). The response variable was transformed to log-odds [i.e., $\log \mathrm{}\left(\frac{GAGome\mathrm{\hspace{0.17em}}score}{100\mathrm{-}GAGome\mathrm{\hspace{0.17em}}score}\right)$] and modeled with a skew normal distribution with heterogenous variance with weakly informed priors. We defined a “region of practical equivalence” for the regression coefficient posterior distribution between −0.52 and +0.50 (equivalent to a GAGome score change that implies an odds-ratio (OR) for cancer between 0.95–1.05) and then followed the SEXIT framework (20) to estimate the posterior probability of existence and of significance for each association. Prespecified posterior probabilities >90% were considered statistically significant. We also defined a “region of large effect” for the regression coefficient posterior distribution when outside −1.7 to +1.4 (equivalent to a GAGome score change that implies an OR for cancer <0.85 or >1.15) and estimated the posterior probability for each association of being large, in which a prespecified posterior probability >90% was considered a large effect, >60% and ≤90% moderate, >15% and ≤60% small, and ≤15% negligible.

Adding the GAGome score to the LLPv3 risk model

We assessed the effect of adding the GAGome score to the LLPv3-predicted 5-year risk (defined as the output of the underlying logistic regression model) to improve the accuracy of screening eligibility. First, we tested the independence of the GAGome score from the LLPv3-predicted 5-year risk using the likelihood ratio test, in which a logistic regression model using LLPv3 as continuous explanatory variable was fitted to a binary response variable (lung cancer case vs. control) and compared with a saturated model in which the GAGome score was added as a second continuous explanatory variable. P values <0.05 were considered significant. Next, we applied the LLPv3 criteria (5-year risk ≥ 1.51%) to calculate the number of “high-risk” patients according to the LLPv3 model. Then, we selected a cutoff for the saturated model (LLPv3 plus GAGome score) so that the number of patients above this cutoff score was the same as the number of “high-risk” patients according to the LLPv3 model. The resulting patient subsets had thus equal sizes, with one subset at “high-risk” according to LLPv3 and the other one according to LLPv3 plus GAGome score. The subset with the largest proportion of lung cancer cases would identify the most efficient model between LLPv3 and LLPv3 plus GAGome score to decide screening eligibility (21). We compared the sensitivity and specificity after internal validation, performed as described above.

Power calculation

The study was initially powered to achieve a precision of 10% and 0.5% in the 95% CI for sensitivity and specificity, respectively, assuming 22.5% sensitivity and 99.0% specificity for the GAGome score, leading to N > 764 for each arm (22). After inclusion, we realized that AUC was a more appropriate statistic to power the study. We estimated that the power for the included sample sizes to detect an AUC > 0.60 for α < 0.05 was >99% (23). All statistical analyses were performed in R (v4.3.0).

Data availability

The data used in this study is not publicly available because it uses clinical records protected by patient confidentiality. Requests for access to deidentified data can be directed to the corresponding author. Data that can be shared can be released via a material transfer agreement.

Results

Baseline characteristics

We sourced 1,440 patients initially suspected of lung cancer in the LLP Biobank. After exclusions (Supplementary Fig. S1), the study population consisted of 1,387 patients, of which 653 remained cancer-free for ≥5 years of follow-up (controls), 653 had lung cancer, and 81 had OCs. Baseline characteristics were typical of patients initially suspected of lung cancer (Table 1; see Supplementary Table S2 for baseline characteristics of OCs), as well as the distribution of lung cancer subtypes and TNM stage at diagnosis (Table 2).

Table 1.

Baseline characteristics for lung cancer cases versus controls (for OCs, see Supplementary Table S1). Continuous variables were summarized as median (IQR in brackets).

	Lung cancer (N = 653)	Control (N = 653)	P value
Demographics, lifestyle, and medical history (including LLPv3 risk factors)
Age (years)	69 (63, 74)	65 (57, 72)	<0.001
Female	286 (44%)	271 (42%)	0.43
Smoking duration (years)	43 (33, 51)	30 (0, 42)	<0.001
Personal history of cancer	97 (15%)	68 (10%)	0.02
Personal history of LLPv3 lung comorbidities	305 (47%)	341 (52%)	0.053
Personal history of asbestos exposure	105 (16%)	53 (8%)	<0.001
Family history of lung cancer
Early onset (before age 60)	44 (6.7%)	0 (0%)	<0.001
Late onset (on or after age 60)	103 (16%)	67 (10%)
Comorbidities
Chronic lung disease	285 (44%)	222 (34%)	<0.001
COPD	263 (40%)	278 (43%)	0.43
Hypertension	264 (40%)	158 (24%)	<0.001
Emphysema	149 (23%)	99 (15%)	<0.001
Hyperlipidemia	111 (17%)	67 (10%)	<0.001
Type 2 diabetes	99 (15%)	51 (7.8%)	<0.001
Inflammatory polyarthropathy	56 (8.6%)	35 (5.4%)	0.03
Asthma	43 (6.6%)	70 (11%)	0.011
Rheumatoid arthritis	23 (3.5%)	11 (1.7%)	0.056
Bronchiectasis	14 (2.1%)	36 (5.5%)	0.0025
Gout	5 (0.8%)	4 (0.6%)	1
Noninfectious enteritis of intestine	6 (0.9%)	9 (1.4%)	0.6
Bronchitis	4 (0.6%)	16 (2.5%)	0.013
Phlebitis	3 (0.5%)	2 (0.3%)	1
Risk scores
Follow-up time (years)	2.8 (0.93, 6.8)	9.2 (6.5, 14)	<0.001
Time between baseline and diagnosis (days)	11 (4, 15)	NA
LLPv3 5-year predicted risk (%)	2.8 (1.2, 5.5)	0.91 (0.26, 2.9)	<0.001
GAGome score (points, 0–100)	9.2 (0.7, 73)	1.5 (0.088, 27)	<0.001

Abbreviations: COPD, Chronic obstructive pulmonary disease; LLPv3, Liverpool Lung Project version 3; NA, not applicable.

Table 2.

Lung cancer subtypes per TNM stage at diagnosis among lung cancer cases (N = 653). TNM stage in patients with missing data was not imputed.

	Stage I (N = 176)	Stage II (N = 157)	Stage III (N = 150)	Stage IV (N = 82)	Not staged (N = 88)	All lung cancer (N = 653)
Lung adenocarcinoma	84 (48%)	72 (46%)	67 (45%)	38 (46%)	19 (22%)	280 (43%)
Lung squamous cell carcinoma	85 (48%)	80 (51%)	67 (45%)	23 (28%)	11 (12%)	266 (41%)
Small cell lung cancer	7 (4.0%)	4 (2.5%)	16 (11%)	19 (23%)	30 (34%)	76 (12%)
Lung cancer (other)	0 (0%)	1 (0.6%)	0 (0%)	1 (1.2%)	12 (14%)	14 (2.1%)
Mesothelioma	0 (0%)	0 (0%)	0 (0%)	1 (1.2%)	16 (18%)	17 (2.6%)

GAGome score analysis in lung cancer cases versus controls

The GAGome score had an adjusted AUC = 0.63 (95% CI, 0.62–0.63; Supplementary Fig. S2; Supplementary Table S3) to discriminate lung cancer cases versus controls. For comparison, LLPv3 had an adjusted AUC = 0.70 (95% CI, 0.70–0.70). In multivariable analysis, none of the LLPv3 predictors and none of the 14 preselected comorbidities had a significant effect on the GAGome score (Fig. 1; Supplementary Table S4). Consistently, the AUC was similar across comorbidity subsets, ranging 0.66 in patients with no comorbidities (N = 372) to 0.60 in patients with at least one comorbidity (N = 934). Conversely, a lung cancer diagnosis had a significant large effect on the GAGome score (Fig. 1; Supplementary Table S4).

Figure 1.

Association between the GAGome score (in log-odds scale) and different variables (outcome, comorbidities, demographics, lifestyle, and medical history) using a multivariable Bayesian linear regression mixed-effect model adjusted for GAGome analysis batch across the whole cohort (all cases and controls). A posterior distribution for the regression coefficient between −0.52 and +0.50 (in gray, dashed lines) was defined as the “region of practical equivalence” (i.e., not statistically significant), and below −1.65 and above +1.42 (dotted lines) as a large effect. Abbreviations: COPD, Chronic obstructive pulmonary disease; OC, Other cancer; SD, Standard deviation.

The AUC was similar across demographics, across lung cancer subtypes [from 0.60 in lung adenocarcinomas (N = 280) to 0.65 in small cell lung cancer (N = 76); Supplementary Table S3]. We observed a monotonic increase in AUC by stage, ranging 0.57 in stage I (N = 176) to 0.71 in stage IV (N = 82; Supplementary Table S3). We observed equivalent discrimination by LLPv3 risk group with AUC = 0.62 in either patients with LLPv3 5-year risk <1.51% (N = 583) or ≥1.51% (N = 723; Supplementary Table S3).

GAGome score and LLPv3 predictions to improve accuracy of risk stratification

The GAGome score was an independent predictor of lung cancer cases versus controls when added to the LLPv3 model, which used only the LLPv3-predicted 5-year risk as the explanatory variable (likelihood ratio χ² = 177 vs. 149; P < 0.0001).

There were 723 (55%) patients with LLPv3 5-year predicted risk ≥1.51%, of which 453 were cases, leading to an adjusted 69% sensitivity (95% CI, 63%–76%) and 59% specificity (95% CI, 52%–66%; Table 3). By selecting a cutoff for the saturated model (LLPv3 5-predicted risk plus GAGome score) so that the number of test positives was equal to 723, the saturated model had an adjusted 72% sensitivity (95% CI, 63–78) and 61% specificity (95% CI, 54%–67%; Table 3).

Table 3.

Assessment of the effect on risk stratification accuracy when the GAGome score is added to the LLPv3 5-year predicted risk in a risk set with the same number of test positives as the number of patients with LLPv3 ≥ 1.51% (N = 723).

	LLPv3 ≥ 1.51%	LLPv3 plus GAGome score
N test–positive	723a	723b
N test–negative	583	583
True positives	453	467
False negatives	200	186
False positives	270	256
True negatives	383	397
Sensitivity (95% CI)	69% (63%–76%)	72% (63%–78%)
Specificity (95% CI)	59% (52%–66%)	61% (54%–67%)

^aDefined as the number of subjects with LLPv3 >1.51%.

^bDefined with a cutoff such as the number of subjects with a score higher than such cutoff is identical to the number of subjects with LLPv3 >1.51%.

Specifically, the saturated model detected 42 cases that had LLPv3 5-year predicted risk <1.51% but missed 28 cases with LLPv3 5-year predicted risk ≥1.51%. Conversely, LLPv3 5-year predicted risk ≥1.51% incorrectly classified 51 controls as “high-risk” compared with 37 controls misclassified by the saturated model (Table 3).

GAGome score analysis in OC cases versus controls

Because changes in the plasma GAGome were previously associated with OCs, we assessed the performance of the GAGome score in predicting future OC cases (N = 81; see Supplementary Table S2 for baseline characteristics) versus controls (N = 653). The median time between the baseline visit and OC diagnosis was 1.4 years (IQR = 0.26–2.4).

The GAGome score had an adjusted AUC = 0.63 (95% CI, 0.62–0.63; Supplementary Fig. S3) to discriminate OC versus controls. For comparison, LLPv3 had an adjusted AUC = 0.49 (95% CI, 0.48–0.49). In multivariable analysis, an OC diagnosis had a significant large effect on the GAGome score equivalent to that observed for a lung cancer diagnosis (Fig. 1). The AUC was similar across cancer types, ranging from 0.52 in bladder cancer (N = 8) to 0.69 in renal cell carcinoma (N = 5; Supplementary Table S5).

Discussion

In this study, we investigated a previously developed plasma GAGome score (15) in an external population from a lung cancer diagnostic clinic, in which patients were suspected of having lung cancer. The GAGome score was validated, and it identified patients who had lung cancer with modest discrimination (AUC = 0.63). Importantly, an increase of the GAGome score could not be explained by any other risk factors or comorbidities tested in this study. Indeed, a lung cancer diagnosis of any subtype was the most probable reason for an elevation in the GAGome score. This likely reflects a metabolic process altering circulating GAGs, which is specific to malignancy but only active in a subset of patients, as our previous studies suggested (refs. 13, 15, 24–27). Mechanistically, this could reflect a state of sustained inflammation in these patients, which is known to remodel GAGs in the tumor microenvironment and thus predispose patients to a higher risk of developing cancer (28, 29).

Whereas other noninvasive biomarkers have reported higher AUCs in patients initially suspected of having lung cancer, most studies did not validate these AUCs in fully independent populations larger than 100 samples, without modifying the model specifications. To our knowledge, only CYFRA 21-1 and CEA-based protein panels have been validated without modifications in populations as large as the present study (N > 1,000), with reported AUCs ranging from 0.65 (30) to 0.79 (10, 31). In this context, the plasma GAGome score reported consistent discrimination in both the current study and the original study in which it was developed. Nevertheless, further validation in lung cancer screening cohort is needed to justify the clinical use of GAGomes for prediction of lung cancer risk. We believe that initiatives like the INTEGRAL program will be crucial to standardizing the assessment of validated lung cancer biomarkers for improving risk predictions in such populations (9).

In terms of clinical utility, GAGome scores identified 42 additional lung cancer cases that fell below the LLPv3 risk threshold and would have been missed by lung cancer screening. The improvement in performance due to the GAGome score was limited but significant, with sensitivity and specificity increasing by 3% and 2% points, respectively. We believe that the population in this study likely provided worst-case estimates for sensitivity and specificity rather than best-case estimates. LLPv3 itself reported a much lower AUC (0.68) in this study compared with previous studies conducted in lung cancer screening cohorts (AUC = 0.81;ref. 4). This is expected because LLPv3 performs better in screening settings in which most individuals are healthy and at low risk, unlike the symptomatic cohort in the present study. In prior studies, LLPv3 sensitivity at the 5-year risk ≥1.51% was 65%, similar to the current study, but the specificity was 82%, more than 20% points higher than in this study.

One of the limitations of our study is that the cases and controls were not representative of asymptomatic, presumed healthy subjects eligible for lung cancer screening. The controls in our study were ill and suspected of having lung cancer at the time of sampling, leading to a higher prevalence of comorbidities. However, this design gave us more power to identify potential confounding factors that would be rare in a screening population. Because the GAGome score had its lowest AUC in patients with at least one comorbidity (who comprised 71% of our cohort), it is likely that the true specificity was underestimated. In a lung cancer screening population, in which virtually all subjects are healthy and without comorbidities, the specificity could be higher. Another limitation was that the GAGome AUC was lower when cases diagnosed at stage I or more than 6 months after the baseline visit. These cases are more common in screening populations, so the sensitivity observed here might be overestimated.

The benefits of adopting biomarkers for risk-stratified screening could go beyond improving the accuracy of risk predictions. Low adherence is a major barrier to making lung cancer screening cost-effective. Risk models can improve cost-effectiveness (32). Biomarker could further enhance it by increasing the accuracy of risk models and encouraging nonadherent individuals to participate in screening, e.g., through primary care implementation (33). Whereas the health economic value of biomarkers should be balanced against other initiatives aimed at increasing adherence, their combination with improved risk prediction accuracy makes them an appealing option.

These findings suggest that plasma GAGomes change in a subset of patients who have or will develop lung cancer or OCs. These changes are unlikely to be driven by comorbidities or other risk factors, making GAGome scores promising biomarkers for identifying additional individuals currently considered at “low risk” of lung cancer. Further studies in screening populations are needed to determine if the increase—albeit limited—in sensitivity and specificity attributable to GAGome scores have sufficient health economic value to justify their use in improving risk-stratified screening.

Supplementary Material

Figure S1

Title: Patient Flow. Diagram to report flow of patients through the study.

Figure S2.

Title: GAGome scores in LC vs. controls. GAGome scores (in log-odds) in LC vs. controls by A) group and by subsets of B) TNM stage at diagnosis, C) LLPv3 5-year predicted risk, and D) LC subtype at diagnosis. The black horizontal segment indicates the group median, and the vertical segment indicates the group interquartile range.

Figure S3

Title: GAGome scores in OC vs. controls. GAGome scores (in log-odds) in OC vs. controls(A), by subsets of LLPv3 risk (B), and primary disease at diagnosis (C). The black horizontal segment indicates the group median, and the vertical segment indicates the group interquartile range.

Table S1

Checklist for the STARD 2015 reporting guideline for diagnostic accuracy studies.

Table S2

Baseline characteristics for OC. Continuous variables were summarized as median (interquartile range in brackets). Key: LLPv3 - Liverpool Lung Project risk score version 3; COPD - Chronic obstructive pulmonary disease. *7 missing data were omitted.

Table S3

Observed AUC of the GAGome score and LLPv3-predicted 5-year risk across subsets in LC (N = 653) versus control patients (N = 653).

Table S4

Posterior probability of direction, of significance, or of being large for different variables on the GAGome score based on multivariable Bayesian linear regression. Probabilities >90% are highlighted in bold. Key: HDI – Highest Density Interval.

Table S5

Observed AUC of the GAGome score and LLPv3 5-year predicted risk across subsets in the OC versus control population. Key: NE – not estimable.

Supp Materials and Methods

Study Protocol

Authors’ Disclosures

M.P.A. Davies reports grants from Elypta AB during the conduct of the study and outside the submitted work. J.K. Field reports grants from Elypta AB during the conduct of the study. F. Gatto reports personal fees and nonfinancial support from Elypta AB during the conduct of the study and personal fees from BioInnovation Institute Fonden outside the submitted work.

Authors’ Contributions

M.P.A. Davies: Conceptualization, resources, data curation, project administration, writing–review and editing. J.K. Field: Conceptualization, resources, supervision, writing–review and editing. F. Gatto: Conceptualization, formal analysis, supervision, funding acquisition, methodology, writing–original draft, project administration.