Blood-Based Biomarker Panel for Personalized Lung Cancer Risk Assessment

Abstract

PURPOSE

To investigate whether a panel of circulating protein biomarkers would improve risk assessment for lung cancer screening in combination with a risk model on the basis of participant characteristics.

METHODS

A blinded validation study was performed using prostate lung colorectal ovarian (PLCO) Cancer Screening Trial data and biospecimens to evaluate the performance of a four-marker protein panel (4MP) consisting of the precursor form of surfactant protein B, cancer antigen 125, carcinoembryonic antigen, and cytokeratin-19 fragment in combination with a lung cancer risk prediction model (PLCO_m2012) compared with current US Preventive Services Task Force (USPSTF) screening criteria. The 4MP was assayed in 1,299 sera collected preceding lung cancer diagnosis and 8,709 noncase sera.

RESULTS

The 4MP alone yielded an area under the receiver operating characteristic curve of 0.79 (95% CI, 0.77 to 0.82) for case sera collected within 1-year preceding diagnosis and 0.74 (95% CI, 0.72 to 0.76) among the entire specimen set. The combined 4MP + PLCO_m2012 model yielded an area under the receiver operating characteristic curve of 0.85 (95% CI, 0.82 to 0.88) for case sera collected within 1 year preceding diagnosis. The benefit of the 4MP in the combined model resulted from improvement in sensitivity at high specificity. Compared with the USPSTF2021 criteria, the combined 4MP + PLCO_m2012 model exhibited statistically significant improvements in sensitivity and specificity. Among PLCO participants with ≥ 10 smoking pack-years, the 4MP + PLCO_m2012 model would have identified for annual screening 9.2% more lung cancer cases and would have reduced referral by 13.7% among noncases compared with USPSTF2021 criteria.

CONCLUSION

A blood-based biomarker panel in combination with PLCO_m2012 significantly improves lung cancer risk assessment for lung cancer screening.

INTRODUCTION

The National Lung Cancer Screening Trial (NLST) demonstrated that screening high-risk individuals with low-dose computed tomography (LDCT) reduces mortality because of lung cancer by 20%.¹ Similar findings have been reported from the NELSON trial.² In 2013, the US Preventive Services Task Force (USPSTF) recommended LDCT screening for individuals who are between 55 and 80 years old, have ≥ 30 pack-years (PYs) of smoking, and are either currently smoking or formerly smoked with ≤ 15 years since quitting.³ A concern has been that a majority of clinically diagnosed lung cancer cases would not meet USPSTF2013 criteria.⁴ Recently, the USPSTF expanded eligibility for LDCT screening. The USPSTF2021 recommendation is annual screening for lung cancer with LDCT for adults age 50-80 years who have a ≥ 20 PY smoking history and currently smoke or have quit within the past 15 years.⁵ The new USPSTF2021 criteria were given a B recommendation with recognition of needs for research to improve uptake of LDCT screening and to develop biomarkers to more accurately identify persons who would benefit from screening.⁵

CONTEXT

• Key Objective

• Can a combination of a blood-based four-marker protein panel with the prostate lung colorectal ovarian (PLCO)_m2012 lung cancer prediction model better identify individuals for lung cancer screening compared with current US Preventive Services Task Force (USPSTF) criteria?

• Knowledge Generated

• Using prediagnostic case and noncase sera from the PLCO Cancer Screening Trial data, a combined four-marker protein panel + PLCO_m2012 model improved sensitivity by 11.9% and 9.9% and specificity by 12.9% and 6.9% compared with USPSTF2013 and the recent USPSTF2021 criteria, respectively.

• Relevance

• A blood-based biomarker panel-PLCO_m2012 combination significantly improves lung cancer risk assessment compared with USPSTF criteria.

Lung cancer risk prediction models have the potential to identify individuals who would benefit from LDCT screening.³ A model on the basis of individual's characteristics was developed using the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Screening Trial control group (PLCO_m2012) that predicts a 6-year risk of lung cancer among individuals who currently smoke or who formerly smoked.⁶ The model exhibited improved sensitivity and positive predictive value without loss of specificity compared with the NLST criteria.⁶ Currently, a modified version of the PLCO_m2012 model is being implemented into lung cancer screening programs in Ontario, Canada, and the United Kingdom.⁷

Additional means to personalize risk profiles include the use of biomarkers. In previous studies, we identified the precursor form of surfactant protein B (Pro-SFTPB) as predictive of lung cancer risk.^8,9 More recently, we demonstrated in a proof-of-principle study that a four-marker protein panel (4MP) consisting of Pro-SFTPB, cancer antigen 125, carcinoembryonic antigen, and cytokeratin-19 fragment has the potential to identify individuals at risk for developing lung cancer better than Pro-SFTPB alone.¹⁰

We aimed to determine whether the 4MP biomarker panel improves lung cancer risk prediction together with the PLCO_m2012 model risk estimates and the extent to which the combined 4MP + PLCO_m2012 data improve selection of individuals for lung cancer screening beyond the USPSTF2013 and USPSTF2021 criteria.

METHODS

PLCO Specimen Set

Detailed information regarding the PLCO Cancer Screening Trial is provided in the Data Supplement (online only). Briefly, the PLCO Cancer Screening Trial was a randomized multicenter trial, which aimed to evaluate the impact of screening for prostate, lung, colorectal, and ovarian cancer on disease-specific mortality.¹¹ A biorepository was created for blood specimens that were annually collected from consented, intervention group participants.¹² There were 42,450 individuals who have ever smoked in the intervention arm; 85% of the participants in the intervention arm had at least one collection.

All histologically confirmed lung cancers from the ever-smoked participants in the intervention arm who were diagnosed within 6 years of study entry (n = 552 case participants) with prediagnostic sera available were selected for the current study. Noncase participants who have ever smoked were randomly selected in a 4:1 ratio with cases (n = 2,193 noncases). Noncases were followed for an additional 13 years during which they remained cancer-free. For each selected participant, all sera within 6 years of study entry, or up to the time of diagnosis for lung cancer cases, were included in the study specimen set (n = 10,008 specimen total; Table 1; Data Supplement).

TABLE 1.

Patient and Lung Cancer Characteristics for the Entire Specimen Set

Testing of the 4MP in the PLCO Specimen Set

Details regarding immunoassays are provided in the Data Supplement. Levels of ProSFTPB, cancer antigen 125, carcinoembryonic antigen, and cytokeratin-19 fragment were determined using bead-based immunoassays as previously described.¹⁰ Biomarker scores for the combined 4MP were calculated on the basis of the previously developed logistic regression model.¹⁰

Risk Model on the Basis of Participant Characteristics

The PLCO_m2012 model was implemented as previously published.⁶ The model predicts 6-year risk of lung cancer diagnosis; this duration was chosen to optimize application and testing in the NLST, which had a median follow-up of 6 years.⁶ Predictive variables in the PLCO_m2012 model were based on baseline questionnaire information and include age, race or ethnic group, education, body mass index, chronic obstructive pulmonary disease, personal history of cancer, family history of lung cancer and smoking status (current v former), intensity, duration, and quit time.⁶

Statistical Analysis

Initial assessment of the discrimination performance of the 4MP score¹⁰ for distinguishing individuals who went on to develop lung cancer from those who did not was done using the entire study specimen set (Table 1). Discrimination estimates included time-dependent area under the receiver operating characteristic curve (AUC), and sensitivity and specificity measurements. Analyses were stratified by different time intervals of blood collection preceding time of diagnosis ranging from 0-6 months and 0-1 year, with annual increments from 0 to 6 years.

To validate a combination rule of 4MP + PLCO_m2012, the entire study specimen set (n = 10,008 samples) from 10 participating PLCO study centers was divided into a development set and a validation set (Data Supplement). Given the limited number of cases with < 10 PY smoking history in the study specimen set, we focused analyses on those participants with ≥ 10 PY and stratified them into low-, medium-, and high-risk groups defined by pack-years and years since quitting (Data Supplement). We also focused on using the 4MP to inform referral to LDCT on the basis of 1-year risk assessment. Thus, the model combining 4MP + PLCO_m2012 was fixed using the development set that consisted of specimens collected from 145 cases and 760 noncases from participants enrolled at five of the 10 PLCO study centers. The validation set comprised the specimens collected from 200 cases and 991 noncases from participants at the other five participating PLCO study centers (Data Supplement). The five PLCO centers used for validation had a similarly broad distribution in the United States as the five PLCO centers used to develop the combined 4MP + PLCO_m2012 model.

We used PLCO_m2012 risk thresholds of ≥ 1.7% 6 years and ≥ 1% 6 years, which have been shown to result in similar numbers of screening-eligible individuals as the USPSTF2013 and USPSTF2021 screening criteria, respectively.^13,14 Assuming uniform incidence, 1.7% 6-year risk and 1% 6-year risk corresponds to 0.285% and 0.167% 1-year risk.

The combined model of the 4MP + PLCO_m2012 for predicting lung cancer within 1 year was developed by fitting a logistic regression with the 4MP score and the linear predictor of the PLCO_m2012⁶ as two separate predictors (Data Supplement). We applied the intercept adjustment of Prentice and Pyke for logistic regression under case-control sampling to the combined model.¹⁵ The model is thus calibrated to the 6-year risk observed among participants in the intervention arm with ≥ 10 PY (intervention arm of the PLCO Cancer Screening Cohort [ESIA] 10+).

We fitted a logistic regression on the combined model development set with the 4MP score and risk-strata indicators as predictors and made strata-specific intercept adjustments for logistic regression under stratified case-control sampling (Data Supplement).^15,16 This calibrated version of the 4MP considers risk factor information (PYs and years since quitting) used to stratify an individual's risk. No adjustments were made to the published PLCO_m2012. Analyses involving PLCO_m2012 included only individuals with complete PLCO_m2012 predictor data (98% of cases and controls). We assessed overall and strata-level calibration of 4MP and 4MP + PLCO_m2012 by comparing average predicted risk and observed incidence (Data Supplement).

The three risk models (4MP, PLCO_m2012, and 4MP + PLCO_m2012) paired with the two risk thresholds defined six prespecified tests for validation. The overall sensitivity and specificity of the tests were estimated as weighted averages of the observed strata-specific sensitivities and specificities, weighted to reflect the distribution across risk strata in ESIA10+.

We additionally performed an extrapolation to the US population among individuals who smoked ≥ 10 PY (ESUS10+) by using the number of persons and average PLCO_m2012 predicted risk for the defined low-, medium-, and high-risk strata from data collected by the 2015 National Health Interview Survey to calibrate the models and weight performances across risk strata (see the Data Supplement for additional details).

To account for outcome-dependent sampling and multiple specimens per participant, 95% CI were calculated as percentiles of the sampling distribution estimated from 1,000 resamples using a bootstrap method for clustered case-control data.¹⁷ Analyses were performed using R software, version 3.6.1 (R Project for Statistical Computing).¹⁸

RESULTS

Performance of the 4MP in the PLCO Specimen Set

The test set for assessing the performance of the 4MP alone consisted of 552 cases for whom 1,299 serial sera were collected within 6 years of diagnosis and 2,193 noncases for whom 8,709 serial sera were collected (Table 1). Testing the 4MP with a previously fixed combination rule¹⁰ yielded an AUC of 0.79 (95% CI, 0.77 to 0.82) when considering all case sera collected within 1 year before diagnosis (Fig 1; Data Supplement). The AUC of the 4MP for distinguishing all case sera collected 1-6 years before diagnosis from all noncase sera was 0.72 (95% CI, 0.69 to 0.75). Stratification of cases into early-stage (I + II) and late-stage (III + IV) lung cancer diagnosed within one year after blood draw compared with all noncase sera resulted in AUCs of 0.75 (95% CI, 0.72 to 0.78) and 0.83 (95% CI, 0.79 to 0.87), respectively (Data Supplement). The performance of the 4MP for distinguishing all case sera collected within 1 year of lung cancer diagnosis stratified into non–small-cell lung cancer or small-cell lung cancer compared with all noncase sera were 0.80 (95% CI, 0.77 to 0.83) and 0.77 (95% CI, 0.69 to 0.83), respectively (Data Supplement).

FIG 1.

Time-dependent classifier performance of the 4MP for distinguishing all cases and cases stratified by stage from noncases in the entire prostate lung colorectal ovarian specimen set. A fourth-order fitted spline curve was used to illustrate the trajectory of the classifier performance of the 4MP in relation to time to diagnosis from blood draw. 4MP, four-marker protein panel; AUC, area under the receiver operating characteristic curve; DX, diagnosis.

Performance of the Combined 4MP Plus PLCO_m2012 Model Among PLCO Participants Who Smoked ≥ 10 PY

We next assessed whether the 4MP would improve upon the PLCO_m2012 risk model for identifying individuals with a smoking history of ≥ 10 PY who would benefit from LDCT screening (Data Supplement). The combined 4MP + PLCO_m2012 model yielded improved discrimination performance compared with PLCO_m2012 alone for distinguishing case sera collected within 1 year of lung cancer diagnosis from noncase sera (AUC: 0.85 [95% CI, 0.82 to 0.88] v 0.80 [95% CI, 0.77 to 0.83]; difference of 0.05 [95% CI for the difference, 0.03 to 0.07]; Fig 2; Data Supplement). The benefit of the 4MP in the 4MP + PLCO_m2012 model compared with PLCO_m2012 alone resulted from improvement in sensitivity while maintaining high specificity (Fig 2).

FIG 2.

Discrimination performance in the validation set for the combined 4MP + PLCO_m2012 model and PLCO_m2012 alone for distinguishing lung cancer cases versus noncases. Case sera were collected within 1 year of diagnosis. Participants smoked ≥ 10 pack-years. 4MP, four-marker protein panel; AUC, area under the receiver operating characteristic curve; PLCO, prostate lung colorectal ovarian.

To assess for potential clinical benefits, we compared the sensitivity and specificity of the combined 4MP + PLCO_m2012 model with those of the USPSTF2013 and USPSTF2021 eligibility criteria. In comparison with USPSTF2013 criteria consisting of ≥ 1.7% 6-year risk threshold, the combined 4MP + PLCO_m2012 model resulted in the improved sensitivity of 83.5% versus 71.6% (95% CI for the difference of 11.9%, 7.0 to 17.2) and specificity of 69.3% versus 56.4% (95% CI for the difference of 12.9%, 10.6 to 15.2; Table 2; Data Supplement). Within the ever smoker intervention arm 10+ PY group, the improved performance would have resulted in referral to screening of 12.6% more lung cancer cases among the 119 cases who would otherwise receive a lung cancer diagnosis within a year and nonreferral of 4,156 (29.6%) of the 14,061 noncases (Table 2; Data Supplement). On the basis of the combined 4MP + PLCO_m2012 model, of the 119 participants expected to receive a lung cancer diagnosis within 1 year (Data Supplement), 100 would be criteria-positive (Table 2). The number of criteria-positive noncases that would need to be referred to screening to refer one participant who would otherwise receive a lung cancer diagnosis within 1 year would be 99.

TABLE 2.

Accuracy Performances in the Validation Set for the 4MP, PLCO_m2012, and the Combined Model of 4MP Plus PLCO_m2012 at Fixed Thresholds of ≥ 1.7% and ≥ 1% 6-Year Risk, to be Comparable With USPSTF2013 and USPSTF2021 Criteria in ESIA10+

At a ≥ 1.0% 6-year risk threshold corresponding to the USPSTF2021 criteria, the combined 4MP + PLCO_m2012 model exhibited overall improved sensitivity (88.4% v 78.5%, difference of 9.9% [95% CI for the difference: 5.3 to 14.4]) and improved specificity (56.2% v 49.3%, difference of 6.9% [95% CI for the difference, 4.6 to 9.4]; Table 2). If applied within the ever smoker intervention arm 10+ PY group, the combined 4MP + PLCO_m2012 model would have resulted in 9.2% more lung cancer cases among the 119 cases who would otherwise receive a lung cancer diagnosis within a year and 2,244 (13.7%) noncases among 16,356 otherwise referred for annual screening (Table 2; Data Supplement). On the basis of the ≥ 1.0% 6-year risk threshold, 105 of the 119 participants who received a lung cancer diagnosis within 1 year were criteria-positive with the combined 4MP + PLCO_m2012 model (Table 2). The number of criteria-positive noncases that would need to be referred to screening to refer one participant who would otherwise receive a lung cancer diagnosis within 1 year would be 134.

DISCUSSION

Current LDCT screening criteria are based on dichotomization of continuous data and do not include some known risk factors for lung cancer. A substantial number of individuals ultimately diagnosed with lung cancer fall outside of current enrollment criteria.^1,19,20 Additional issues related to LDCT pertain to perceived potential harms.²¹ It follows that there is a need for improved criteria for LDCT screening as recognized in the recent USPSTF recommendations.⁵ The PLCO_m2012 lung cancer risk prediction model offers improved sensitivity without loss of specificity compared with NLST-based criteria.⁶ PLCO_m2012 has been extensively externally validated.^6,13,22–28 Here, we demonstrate that a combined model of the 4MP with PLCO_m2012 yields further improved performance compared with PLCO_m2012 alone and improved sensitivity and specificity compared with the USPSTF2013 and USPSTF2021 criteria. In the US population, if the 4MP + PLCO_m2012 model was applied to individuals with ≥ 10 PY smoking history and age 55-74 years, an estimated 12,095 (15.1%) more lung cancer cases among the 79,924 cases who would otherwise receive a lung cancer diagnosis within a year would be offered screening and 986,668 (15.2%) of the 6,489,951 noncases would not need screening compared with USPSTF2013 criteria (Data Supplement). Similar results are predicted compared with the USPSTF2021 criteria. This is potentially a very large public health benefit as it better identifies individuals at high risk of lung cancer and expands upon the number of individuals who would be considered eligible for lung cancer screening, thereby addressing some of the limitations of current screening eligibility criteria.⁵ Moreover, it is important to note that countries outside of the United States have not adopted the USPSTF2021 criteria and most have not adopted lung cancer screening. The European Union is currently testing an age-smoking eligibility approach that more closely resembles the USPSTF2013 eligibility criteria.^29,30 Therefore, our findings have relevance on a global scale.

We envision that testing of the 4MP would be useful for adults who are currently eligible for LDCT screening on the basis of USPSTF criteria and expanded to additionally include individuals who have ≥ 10 PY smoking history. Such a test could be performed as part of routine blood testing for wellness checks. An individual identified to be at high risk for lung cancer on the basis of the 4MP + PLCO_m2012 model or the 4MP alone, if information required for the PLCO_m2012 is not available, would then trigger referral to LDCT screening. For individuals eligible for LDCT screening with a low risk of lung cancer, the 4MP + PLCO_m2012 or the 4MP alone may also identify a substantial group of individuals who may not need annual LDCT screening.

In the NLST study, much of the mortality advantage from LDCT screening came from incident cancers that were detected at earlier stages and that were associated with better outcomes.^19,31 The combined 4MP + PLCO_m2012 model exhibited high sensitivity for early-stage lung cancer diagnosed within one year of blood draw with an AUC of 0.82 (95% CI, 0.78 to 0.86) in the validation set. Thus, implementation of the 4MP + PLCO_m2012 model for screening selection would be expected to lead to mortality reduction.

The uptake of LDCT screening remains low despite USPSTF recommendations.^32,33 Surveys assessing attitudes toward screening have consistently revealed that patient and provider uncertainty with overdiagnosis, inconclusive results, and false-positive screens represent major barriers to wider uptake.^34,35 Individualized risk assessment with the addition of a minimally invasive blood test has the potential to be more informative to base a decision to enroll in a lung cancer screening program.

We note some limitations to our study. In PLCO, age between 55 and 74 years was an entry criterion.³⁶ Therefore, the predictive performance of the combined model outside of this age range could not be determined. Similarly, the predictive performance of the combined 4MP + PLCO_m2012 model for participants who smoked < 10 PY was not assessed. In the NLST study, individuals with PLCO_m2012 < 0.64% 6-year risk had no mortality benefit from LDCT screening.³⁷ Nevertheless, a primary question remains as to whether a biomarker test can identify subsets of individuals who smoked < 10 PY and yet have ≥ 1.0% individualized 6-year risk who may benefit from LDCT screening. Validation of the 4MP + PLCO_m2012 model in a more generalizable population of individuals who smoked is needed. The performance of the 4MP + PLCO_m2012 among different ethnic and racial groups was not assessed because of small strata.

Indeterminate lung nodules engender a significant diagnostic conundrum. We have previously demonstrated that the 4MP contributes to the assessment of indeterminate pulmonary nodules with improved performance compared with the nodule size alone in predicting likelihood of cancer.³⁸ This observation corroborates the association between 4MP and lung cancer. The 4MP may potentially yield additional information that can be incorporated into nodule malignancy probability models for assessment of LDCT-positive screens.

Another important consideration will be cost-effectiveness analyses of using the combined 4MP + PLCO_m2012 model for risk-based referral to lung cancer screening programs in addition to the improvement in performance against the increase in costs when considering all four protein markers instead of two to three proteins. Screening according to USPSTF2013 has been shown to be cost-effective.^39,40 It is expected that a simple 4MP test would be less costly than LDCT. Modeling analyses suggest that if this is the case, a combined biomarker and LDCT screening strategy would likely be cost-effective.⁴¹ Given the substantially improved performance (sensitivity and specificity) of the combined 4MP + PLCO_m2012 versus the USPSTF2013 and USPSTF2021 criteria, and the simplicity of the test, we anticipate its implementation to be also cost-effective.⁴¹ Nevertheless, specific cost-effectiveness studies would be needed.

In conclusion, the 4MP + PLCO_m2012 model yielded superior predictive performance and sensitivity and specificity for ruling individuals into LDCT screening compared with USPSTF2013 or USPSTF2021 eligibility criteria and with the PLCO_m2012 model alone. These findings have important implications for improving lung cancer screening programs and reducing the burden of lung cancer through personalized risk assessment.

AUTHOR CONTRIBUTIONS

Conception and design: Johannes F. Fahrmann, Ehsan Irajizad, Nikul Patel, Jennifer B. Dennison, Kim-Anh Do, Edwin Ostrin, Margaret R. Spitz, Rafael Meza, Martin C. Tammemägi, Ziding Feng, Samir Hanash

Financial support: Martin C. Tammemägi, Ziding Feng

Administrative support: Ziding Feng, Samir Hanash

Collection and assembly of data: Johannes F. Fahrmann, Ehsan Irajizad, Nikul Patel, Eunice Murage, Martin C. Tammemägi, Ziding Feng, Samir Hanash

Data analysis and interpretation: Johannes F. Fahrmann, Tracey Marsh, Ehsan Irajizad, Nikul Patel, Jody Vykoukal, Jennifer B. Dennison, Kim-Anh Do, Edwin Ostrin, Stephen Lam, Sanjay Shete, Rafael Meza, Martin C. Tammemägi, Ziding Feng, Samir Hanash

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

Blood-Based Biomarker Panel for Personalized Lung Cancer Risk Assessment

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/jco/authors/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).