An effective and affordable blood test for lung cancer early detection using four protein markers and artificial intelligence

Abstract

Background

Lung cancer constitutes the leading cause of cancer mortality globally. This study assessed LungCanSeek, a novel blood-based protein test for lung cancer early detection.

Methods

This retrospective study enrolled 1,814 participants (1,095 lung cancer, 719 non-cancer) from three different cohorts. Blood samples were analyzed for four protein tumor markers (PTMs) using Roche cobas. Artificial intelligence (AI) algorithms were developed for lung cancer detection and subtype classification: lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), and small cell lung cancer (SCLC). A two-step approach was modeled, using LungCanSeek for initial screening, followed by low-dose computed tomography (LDCT) for LungCanSeek’s positive cases.

Results

LungCanSeek achieved 83.5% sensitivity, 90.3% specificity, and 86.2% accuracy overall. Sensitivities of LUAD, LUSC, and SCLC were 83.3%, 81.4%, and 91.9%. Sensitivity increased with clinical stage in non-small cell lung cancer (NSCLC): 59.5% (I), 69.8% (II), 86.5% (III), and 91.3% (IV). Sensitivities of limited-stage and extensive-stage SCLC were 91.3% and 93.0%, respectively. The subtype classification accuracy was 77.4%. Simulation model analysis showed that the two-step approach reduced 10.3-fold false positives and 2.5-fold cost compared to LDCT for lung cancer screening in high-risk population.

Conclusions

LungCanSeek is a non-invasive and cost-effective test for lung cancer early detection. The two-step approach offers a cost-effective strategy for population-wide lung cancer screening.

Highlight box.

Key findings

• LungCanSeek, an artificial intelligence (AI)-integrated 4-protein blood test, demonstrated promising performance for lung cancer early detection with 83.5% sensitivity and 90.3% specificity.

What is known and what is new?

• Low-dose computed tomography’s (LDCT) high false-positive rate and reliance on specialized infrastructure and radiologists limit its effectiveness in lung cancer prevention.

• This study demonstrates that LungCanSeek is a novel blood-based test for lung cancer early detection with superior and robust performance.

What is the implication, and what should change now?

• LungCanSeek offers an affordable and practical solution for lung cancer early detection.

• LungCanSeek can be used in combination with LDCT to expand screening accessibility, making it a cost-effective strategy for population-wide lung cancer screening.

Introduction

Lung cancer is the leading cause of cancer mortality worldwide (1), comprising two primary types: small cell lung cancer (SCLC), representing approximately 15% of cases, and non-small cell lung cancer (NSCLC), constituting roughly 85% of cases (2). Among NSCLC, the most common subtypes are lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) (3-5). Despite advancements in the treatment for lung cancer, the overall 5-year survival rate remains below 20%, largely due to the diagnosis at a late stage (6). Early detection followed by radical surgery significantly improves outcomes, achieving 5-year survival rates ranging from 70% to 90% for NSCLC patients (7). Unfortunately, over 75% of NSCLC cases are diagnosed at an advanced stage. Moreover, SCLC is characterized by aggressive behavior and early metastasis, with over 60% of patients presenting with extensive-stage disease (8). The overall 5-year survival rate for SCLC is notably low at around 5% (9). Therefore, early detection strategies are pivotal in improving lung cancer patients’ clinical outcomes and survival rates.

Large clinical trials have shown that lung cancer screening using low-dose computed tomography (LDCT) reduces mortality among high-risk individuals (10-12). However, challenges like the high false-positive rate of 96.4% in the positive results (10), high cost and infrastructure requirements of LDCT, along with the need for skilled radiologists, limit its widespread adoption, particularly in low- and middle-income countries (LMICs). The advancements in liquid biopsy techniques, including the analyses of circulating nucleic acids (13,14), proteins (15), and tumor cells (16,17), offer promising avenues for non-invasive lung cancer screening. These methods are advantageous due to their non-invasive nature and easy blood sample collection. However, using next-generation sequencing (NGS) to detect cancer-derived DNA features for cancer early detection (13,14,18,19) remains impractical for LMICs due to the cost and infrastructure limitations. Therefore, even in high-income countries, late-stage diagnosis and disparities in healthcare resources underscore the need for robust and affordable screening methods. Developing such methods is crucial for enabling widespread early detection of lung cancer, especially in LMICs.

The clinical utility of blood-based protein markers for lung cancer diagnosis has been extensively validated (20-23). Compared to the diagnostic methods such as bronchoscopy and imaging, protein marker-based tests offer advantages such as non-invasiveness, simplicity, and cost-effectiveness (24). Previous studies have demonstrated that combining multiple protein tumor markers (PTMs) enhances cancer detection sensitivity compared to a single PTM (25,26). However, conventional clinical practice evaluates each marker using independent thresholds, which can lead to false-positive accumulation (27), resulting in unnecessary diagnostic workups and increasing financial and psychological burdens. In our previous study (28), we employed artificial intelligence (AI) to integrate a seven-marker PTM panel for multi-cancer early detection and successfully reduced the false-positive rate from 43.1% (conventional method) to 7.1% (AI-driven model). Here, we extend this AI-based approach to develop a protein assay specifically for the early detection of lung cancer.

This retrospective study constructs three independent cohorts to evaluate the performance and robustness of LungCanSeek, a protein test designed for lung cancer early detection, which utilizes the AI method to integrate the levels of four selected PTMs [carcinoembryonic antigen (CEA), cytokeratin 19 fragment antigen 21-1 (CYFRA 21-1), pro-gastrin releasing peptide (ProGRP), squamous cell carcinoma antigen (SCCA)] and clinical information (age and gender) to identify lung cancer patients and predict histological subtypes. Meanwhile, we propose a two-step lung cancer screening approach—using LungCanSeek for initial screening, followed by LDCT for LungCanSeek’s positive cases. The two-step approach not only significantly reduces the unnecessary biopsies caused by LDCT’s high false positives but also cuts the screening cost down substantially, showing great potential in population-level lung cancer screening. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-317/rc).

Methods

Participants

From November 2019 to March 2024, 1,814 participants (1,095 with lung cancer and 719 without cancer) were recruited from Sun Yat-sen Memorial Hospital, SeekIn, and Henan Cancer Hospital for this study. Based on their sources, these cases were divided into three cohorts. A total of 383 lung cancer patients and 332 non-cancer individuals from Sun Yat-sen Memorial Hospital constituted the training cohort. SeekIn recruited 120 lung cancer patients and 178 non-cancer individuals for independent validation cohort 1. Additionally, 592 lung cancer patients were recruited from Henan Cancer Hospital, a specialized cancer hospital where obtaining non-cancer samples was challenging, and 209 non-cancer individuals were gathered from Shenyou Bio, which is a wholly owned subsidiary of SeekIn and located in the same city as Henan Cancer Hospital, to form independent validation cohort 2. In this study, lung cancer patients were defined as those with pathologically confirmed lung cancer who had not received any treatment prior to blood sampling. Those with a history of previous malignancy were excluded. NSCLC was staged according to the American Joint Committee on Cancer (AJCC) Staging Manual (8th edition) (29), and SCLC was staged using the Veterans Administration (VA) Lung Group Classification Scheme, distinguishing between limited and extensive stages (30). Non-cancer individuals were defined as those without a known cancer diagnosis or a prior history of malignancy who underwent routine physical examinations. No additional cancer-related diagnostic testing was performed. Validation cohort 1 (SeekIn) in this study was derived from our previously published studies (28,31). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Medical Ethics Committees of Henan Cancer Hospital (No. 2023-KY-0153) and Sun Yat-sen Memorial Hospital (No. SYSKY-2023-435-02). Informed consent was obtained from all individual participants. Shenyou Bio was also informed and agreed the study. All the data collected were anonymized to protect participant privacy.

Quantification of PTMs

Peripheral blood samples were collected using a 10 mL serum collection tube (BD Bioscience, San Jose, USA) from Sun Yat-sen Memorial Hospital and Henan Cancer Hospital. Serum was separated via centrifugation at 1,300 ×g for 10 minutes at 4 ℃. Samples from SeekIn utilized a 10 mL Cell-Free DNA BCT tube (Streck, La Vista, USA), with plasma separation achieved by centrifugation at 1,600 ×g for 10 minutes at 4 ℃. 500 µL of serum or plasma was used to quantify the levels of four selected PTMs, including CEA, CYFRA 21-1, ProGRP, and SCCA, using the Roche cobas e411/e601 (Roche Diagnostics GmbH, Mannheim, Germany).

The conventional clinical method for lung cancer detection

The conventional clinical method to determine cancer utilizes a single threshold based on predetermined reference ranges for each PTM, as recommended by the manufacturer. The manufacturer-suggested cut-off values for the PTMs were as follows: 5.0 ng/mL for CEA, 3.3 ng/mL for CYFRA 21-1, 59.5 pg/mL for ProGRP, and 2.7 ng/mL for SCCA.

Construction of the models of lung cancer detection and subtype classification

In this study, quantitative results, which were corrected by modified Z-score from the four selected protein markers and two clinical characteristics (age and gender) served as input features for an AI algorithm, following the modeling approach detailed in our previous publication (28). The process comprised two parts. The first is a lung cancer detection model to ascertain the presence of lung cancer in individuals. The Generalized Linear Models (GLM) algorithm was chosen to establish the lung cancer detection model due to its high interpretability, flexibility, and low risk of overfitting (28). This process with 10-fold cross-validation was repeated 30 times. The average prediction value from the GLM model was defined as the probability of lung cancer (PLC) index. PLC value at 90.1% specificity in the training cohort was selected as the cut-off value. When the PLC was greater than the cut-off, it indicated that lung cancer signals were detected. Otherwise, no lung cancer signal was detected. The second part is a subtype classification model. For the prediction of subtype, the true positive patients of three cohorts were used to develop the lung cancer subtype classification model through the Random Forest method (32). This process employed 10-fold cross-validation to compute the average prediction probability of patients belonging to each subtype. The subtype with the highest prediction probability was considered the potential lung cancer subtype.

Statistical analysis

We employed t-distributed stochastic neighbor embedding (t-SNE) cluster analysis, a generalized nonlinear dimensionality reduction algorithm (33), to evaluate distinct expression patterns of PTMs between lung cancer patients and non-cancer individuals, as well as among three different lung cancer subtypes.

Statistical analyses were conducted via R statistical software (version 4.2.0, RRID: SCR_001905). The receiver operating characteristic (ROC) curve’s performance in distinguishing lung cancer cases from non-cancer individuals was evaluated with the pROC package (version 1.18.0, RRID: SCR_024286). Expression levels of protein markers between lung cancer patients and non-cancer individuals were compared using the Wilcoxon rank-sum test. Sensitivity, specificity, and predictive values were calculated using the epiR software package (version 2.0.60, RRID: SCR_021673). Differences in the area under the curve (AUC) were assessed using the DeLong test.

Results

Participants’ demographic characteristics

A total of 1,814 individuals (1,095 with lung cancer and 719 without cancer) were enrolled in this case-control study. The demographics and clinical characteristics of all participants are summarized in Table 1. The lung cancer group consisted of three subtypes: LUAD (547 cases, 52.4%), LUSC (236 cases, 21.6%), and SCLC (173 cases, 15.8%). Among patients with NSCLC, the distribution across stages was as follows: stage I (14.9%), stage II (4.8%), stage III (15.5%), and stage IV (33.6%). For SCLC patients, 4.2% were diagnosed with limited-stage disease, while 9.1% presented with extensive-stage disease. There was a difference in age distribution between the cancer and non-cancer groups, with a higher proportion of cancer patients in the elderly age group (Chi-squared test, P<0.001). Both the cancer and non-cancer groups exhibited a similar higher proportion of males (67.7% and 68.8%, respectively) compared to females.

Table 1. Clinical and demographic characteristics of participants.

Characteristics	Training cohort (SYSMH)			Validation cohort 1 (SeekIn)			Validation cohort 2 (HNCH)			Total
	Cancer (n=383)	Non-cancer (n=332)		Cancer (n=120)	Non-cancer (n=178)		Cancer (n=592)	Non-cancer (n=209)		Cancer (n=1,095)	Non-cancer (n=719)
Age, years	62.4±9.9	52.8±11.6		60.8±10.3	49.9±11.9		63.6±11.0	51.7±12.6		62.9±10.6	51.8±12.0
≤55	76 (19.8)	193 (58.1)		29 (24.2)	124 (69.7)		142 (24.0)	146 (69.9)		247 (22.6)	463 (64.4)
>55	307 (80.2)	139 (41.9)		91 (75.8)	54 (30.3)		450 (76.0)	63 (30.1)		848 (77.4)	256 (35.6)
Gender
Female	106 (27.7)	32 (9.6)		45 (37.5)	91 (51.1)		203 (34.3)	101 (48.3)		354 (32.3)	224 (31.2)
Male	277 (72.3)	300 (90.4)		75 (62.5)	87 (48.9)		389 (65.7)	108 (51.7)		741 (67.7)	495 (68.8)
Subtype		–			–			–			–
LUAD	240 (62.7)			45 (37.5)			289 (48.8)			574 (52.4)
LUSC	87 (22.7)			31 (25.8)			118 (19.9)			236 (21.6)
SCLC	37 (9.7)			27 (22.5)			109 (18.4)			173 (15.8)
Unknown†	19 (4.9)			17 (14.2)			76 (12.8)			112 (10.2)
Stage		–			–			–			–
I	141 (36.8)			2 (1.7)			20 (3.4)			163 (14.9)
II	47 (12.3)			0 (0.0)			6 (1.0)			53 (4.8)
III	105 (27.4)			4 (3.3)			61 (10.3)			170 (15.5)
IV	53 (13.8)			24 (20.0)			291 (49.2)			368 (33.6)
LS	19 (5.0)			5 (4.2)			22 (3.7)			46 (4.2)
ES	18 (4.7)			7 (5.8)			75 (12.7)			100 (9.1)
Unknown	0			78 (65.0)			117 (19.8)			195 (17.8)

Data are presented as mean ± standard deviation or n (%). ^†, lung adenocarcinoma mixed with squamous lung cancer was classified as unknown type. ES, extensive-stage; HNCH, Henan Cancer Hospital; LS, limited-stage; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; SCLC, small cell lung cancer; SYSMH, Sun Yat-sen Memorial Hospital, Sun Yat-sen University.

The performance of the conventional clinical method for lung cancer detection

In our previous study on multi-cancer early detection (28), CEA, CYFRA 21-1, neuron-specific enolase (NSE) were found to be significantly elevated in lung cancer patients compared to non-cancer individuals. However, NSE was excluded during the early stage of assay development due to its high susceptibility to hemolysis-induced false positives, which we also observed experimentally and have been documented in a previous study (34). In contrast, CEA and CYFRA 21-1 were retained as validated markers for lung cancer detection (28). In addition, ProGRP and SCCA were included based on clinical evidence supporting their utility in SCLC and LUSC detection, respectively (22,35,36). Therefore, we finally chose CEA, CYFRA 21-1, ProGRP, and SCCA to build a panel for lung cancer early detection.

The levels of the four selected PTMs for the training cohort are shown in Figure 1. Compared to non-cancer individuals, the levels of CEA and CYFRA 21-1 were significantly higher in all subtypes of lung cancer patients (Wilcoxon rank-sum test, P<0.001; Figure 1A,1B), underscoring their utility in lung cancer detection across all subtypes. Meanwhile, ProGRP levels in LUAD and SCLC and SCCA levels in LUSC were also significantly elevated relative to non-cancer individuals (Wilcoxon rank-sum test, P<0.001; Figure 1C,1D), suggesting their potential for cancer detection and possible subtyping. These findings were further supported by t-SNE clustering analysis (Figure S1), demonstrating distinct patterns that differentiate lung cancer patients from non-cancer individuals, thus affirming the effectiveness of these PTMs in lung cancer detection.

Figure 1.

The levels of each PTM in different lung cancer subtypes and non-cancer individuals (A-D). The black horizontal lines indicate the cut-off values recommended by the manufacturer. ns, no significant; ****, P<0.0001. CEA, carcinoembryonic antigen; CYFRA 21-1, cytokeratin 19 fragment antigen 21-1; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; ProGRP, pro-gastrin releasing peptide; PTM, protein tumor marker; SCCA, squamous cell carcinoma antigen; SCLC, small cell lung cancer.

We investigated the performance of these four PTMs across all lung cancer subtypes in the training cohort using the conventional clinical method, as depicted in Table S1. Each PTM exhibited specificity greater than 86.1%, while their individual sensitivity for subtype detection varied widely, ranging from 5.4% to 72.4% (median 38.3%), except for ProGRP in SCLC, which achieved 94.6% sensitivity. We combined these PTMs to achieve higher sensitivities for all subtypes. However, as the number of PTMs increased, sensitivity improved from 13.6% to 90.1%, albeit at the expense of reduced specificity from 97.0% to 75.3% and increased false-positive rate from 3.0% to 24.7% (Figure S2). Therefore, an AI approach (LungCanSeek) was employed to mitigate the issue of accumulating high false positives while maintaining sensitivity.

The performance of LungCanSeek for lung cancer detection

LungCanSeek’s lung cancer detection model was developed using the training cohort from Sun Yat-sen Memorial Hospital, achieving an AUC of 0.912. Furthermore, it achieved AUCs of 0.874 and 0.960 in two independent validation cohorts respectively (Figure 2A). Notably, there was no significant difference in AUC between the training cohort and validation cohort 1 (DeLong test, P=0.11), indicating the robustness of LungCanSeek. Validation cohort 2 exhibited superior performance in AUC compared to the training cohort (DeLong test, P<0.001), likely due to the higher prevalence of advanced lung cancer cases in validation cohort 2 compared to the training cohort (stage IV NSCLC and extensive-stage SCLC at 61.9% and 18.5% respectively). Despite the significant difference in the proportion of early-stage (I–II) and late-stage (III–IV) cases between training cohort and validation cohort 2 (Chi-squared test, P<0.001), LungCanSeek maintained comparable AUCs for early-stage (I–II) cases across both cohorts (DeLong test, P=0.64; Figure S3A), as well as for late-stage (III–IV) cases (DeLong test, P=0.46; Figure S3B). Although the three cohorts were not gender-matched or age-matched, LungCanSeek demonstrated consistent performance across gender and age subgroups within each cohort (DeLong test, P>0.05; Figure S4). LungCanSeek demonstrated sufficient specificity (~90.0%) and varying sensitivities across the three cohorts: 75.2% in the training cohort [95% confidence interval (CI): 70.6–79.4%], 67.5% in validation cohort 1 (95% CI: 58.3–75.8%), and 92.1% in validation cohort 2 (95% CI: 89.6–94.1%) (Table S2). In a total of 1,814 participants (1,095 with lung cancer and 719 without cancer) across three cohorts, LungCanSeek demonstrated an overall sensitivity of 83.5% (95% CI: 81.1–85.6%) at a specificity of 90.3% (95% CI: 87.9–92.3%), resulting in 86.2% accuracy (Table 2). However, the conventional clinical method showed slightly higher sensitivity (88.2%, 95% CI: 86.2–90.1%) but significantly lower specificity (80.9%, 95% CI: 77.9–83.8%) across all the cases (Table 2). The sensitivities observed in different lung cancer subtypes were 83.3% (95% CI: 80.0–86.2%) in LUAD, 81.4% (95% CI: 75.8–86.1%) in LUSC, and 91.9% (95% CI: 86.8–95.5%) in SCLC (Figure 2B). In NSCLC, sensitivities increased with clinical stage progression (I, II, III, and IV), with rates of 59.5% (95% CI: 51.6–67.1%), 69.8% (95% CI: 55.7–81.7%), 86.5% (95% CI: 80.4–91.2%), and 91.3% (95% CI: 87.9–94.0%) respectively (Figure 2C). For limited-stage and extensive-stage SCLC, sensitivities were 91.3% (95% CI: 79.2–97.6%) and 93.0% (95% CI: 86.1–97.1%), respectively (Figure 2D). Overall, LungCanSeek demonstrated a sensitivity of 67.2% in early-stage lung cancer patients (stage I and II NSCLC, and limited-stage SCLC), highlighting its value in lung cancer early detection.

Figure 2.

The performance of LungCanSeek test. (A) The ROC curve evaluates the performance of LungCanSeek in the training and independent validation cohorts with the AUC depicted for each cohort. (B) LungCanSeek’s sensitivities in three lung cancer subtypes at approximately 90.0% specificity. The y-axis represents sensitivities by subtype classes, while the x-axis denotes different lung cancer subtypes with subtype classes ordered based on decreasing sensitivity. The bars indicate the 95% confidence interval, and the numbers in parentheses represent the sample size for each subtype class. (C) LungCanSeek’s sensitivity in each clinical stage of NSCLC at approximately 90.0% specificity. The y-axis shows sensitivities based on four different stages of NSCLC, with bars indicating the 95% confidence interval. Numbers in parentheses indicate the sample size for each clinical stage of NSCLC. (D) LungCanSeek’s sensitivity in each clinical stage of SCLC at approximately 90.0% specificity. The y-axis represents sensitivities based on two different stages of SCLC, with bars indicating the 95% confidence interval. Numbers in parentheses indicate the sample size for each clinical stage of SCLC. AUC, area under the curve; ES, extensive-stage; LS, limited-stage; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; NSCLC, non-small cell lung cancer; ROC, receiver operating characteristic; SCLC, small cell lung cancer.

Table 2. Performance of the conventional clinical method and LungCanSeek across all cases.

Performance	Conventional clinical method			LungCanSeek
	Cancer	Non-cancer		Cancer	Non-cancer
Predicted cancer	966†	137†		914	70
Predicted non-cancer	129	582		181	649
Sensitivity (95% CI), %	88.2 (86.2–90.1)			83.5 (81.1–85.6)
Specificity (95% CI), %	80.9 (77.9–83.8)			90.3 (87.9–92.3)

^†, subjects with at least one of the markers included in the panel showing values above the cut-off point were considered to be positive. CI, confidence interval.

Classification of lung cancer subtypes

Different subtypes of lung cancer exhibit distinct clinical characteristics that require tailored treatment approaches. Figure 3A showed significant differences in PTM expression patterns among three lung cancer subtypes, particularly in distinguishing SCLC. Using these subtype-specific PTM profiles, a supervised AI algorithm accurately predicted lung cancer subtypes in 829 patients with true positive tests across the three cohorts. Overall subtype classification accuracy was 77.4%, with specific accuracies of 74.8% for NSCLC (86.6% for LUAD and 54.7% for LUSC) and 77.4% for SCLC (Figure 3B), underscoring the method’s feasibility for subtype prediction without the risks and challenges associated with tumor biopsy.

Figure 3.

The performance of lung cancer subtype classification. (A) Dimensionality reduction of PTMs shows significant differences between different subtypes of lung cancer. The red dots represent LUAD patients. The blue dots represent LUSC patients. The purple dots represent SCLC patients. (B) Alluvial diagram illustrates the accuracy of lung cancer subtype classification. It shows the agreement between actual (first column) and predicted (second column) subtype classifications per sample using LungCanSeek. The numbers on the right side display the classification accuracy for each subtype. LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; PTM, protein tumor marker; SCLC, small cell lung cancer; t-SNE, t-distributed stochastic neighbor embedding.

Two-step approach for lung cancer early detection in the high-risk population

To address the high false-positive rate of LDCT, which was reported as 96.4% (10), a two-step lung cancer screening approach was proposed, utilizing LungCanSeek for the initial screening, followed by LDCT for the individuals who tested positive in the initial screening (Figure S5).

Among the estimated 15 million high-risk individuals eligible for lung cancer screening in the United States in 2024 (37), we assumed a screening adherence rate of 60%, consistent with the uptake rate observed for colorectal cancer screening in 2021 (38). Accordingly, we modeled a screening scenario involving 9 million high-risk individuals, with an assumed lung cancer incidence rate of 2.5% in this population (39). The modeled real-world sensitivity of LungCanSeek was adjusted downward proportionally based on the sensitivity differences observed between the prospective study and the retrospective study of the EarlyCDT-Lung test (15,40), from 83.5% to 65.4%. The modeled real-world performance of the two-step approach was calculated based on LungCanSeek’s modeled real-world performance combined with LDCT’s performance, with a sensitivity of 93.1% at 76.5% specificity in National Lung Screening Trial (NLST) study (39), resulting in a sensitivity of 60.9% at 97.7% specificity (Figure 4). The two-step approach reduced false positives by 10.3-fold to just 200,026 compared to 2,062,125 with LDCT (Figure 4). Although LDCT identified more lung cancer cases (209,475) than the two-step approach (136,997, missing 34.6% of LDCT-detected cases), the two-step approach demonstrated a superior positive predictive value (PPV) of 40.6% which was 4.4-fold higher than that of LDCT (9.2%). Additionally, the negative predictive value (NPV) of the two-step approach (99.0%) was slightly lower than that of LDCT (99.8%). Meanwhile, LDCT ($277 per test) (41) incurred substantially higher costs, totaling $2.493 billion. In contrast, the two-step approach achieved a cost reduction of 2.5-fold, amounting to a total cost of $996.5 million and a cost of $111 per individual screened. The cost of per cancer case detected using LDCT was $11,901 compared to $7,274 using the two-step approach, which represented a 1.6-fold difference (Figure 4). Hence, the two-step approach not only significantly reduced false positives but also cut the screening cost down substantially, making it a cost-effective strategy for population-wide lung cancer screening.

Figure 4.

Modeling lung cancer screening. The modeled performance and cost among LDCT, LungCanSeek, and the two-step approach in a screening of 9 million high-risk adults. CPCPI, cost of per cancer patient identified; CPIS, cost of per individual screened; LDCT, low-dose computed tomography; NNS, number needed to screen; NPV, negative predictive value; PPV, positive predictive value.

Discussion

Early detection is crucial for improving prognosis by enabling timely treatment before cancer spreads, regardless of cancer type. The US-based NLST study has demonstrated that early detection through LDCT scans can reduce lung cancer mortality by 20% (10), underscoring the critical importance of early detection efforts. However, 96.4% of the LDCT positive results are false positives (10), and the majority of false-positive findings may drive expensive and invasive diagnostic procedures, such as biopsy, and persistent anxiety for the individuals. In addition, challenges such as the requirements for specialized imaging facilities, skilled radiologists, and radiation exposure continue to limit the wide adoption of LDCT, even in high-income countries (42,43). Therefore, there is a strong case for a simple blood test as an early detection method for lung cancer.

Herein, we report an efficient and robust lung cancer early detection test named LungCanSeek, which utilizes the quantifications of four selected PTMs (CEA, CYFRA 21-1, ProGRP, and SCCA) and clinical information (age and gender) empowered by AI. In this multicenter study involving 1,814 participants across three cohorts, LungCanSeek demonstrated an impressive performance with 83.5% sensitivity at 90.3% specificity. Notably, LungCanSeek maintained sensitivity above 81% across all three subtypes of lung cancer. The sensitivity of SCLC (91.9%) was the highest among all the subtypes, which is significant given the approximately 5% 5-year survival rate for SCLC patients (9). It proved effective not only for the late-stage patients (stage III and IV NSCLC plus extensive-stage SCLC) but also for the early-stage patients (stage I and II NSCLC plus limited-stage SCLC), achieving a sensitivity of 67.2%, crucial for lung cancer early detection. Meanwhile, two independent validation cohorts from different geographical locations and laboratories, with different sample types (plasma and serum), were utilized in this study, underscoring the reliability and generalizability of LungCanSeek. Moreover, LungCanSeek’s reliance on just four widely available protein markers simplifies testing, leveraging reagents and instruments already in clinical use without specialized skills or complex infrastructure. As a novel non-invasive blood assay, LungCanSeek is characterized by ease of use and affordability, with an estimated reagent cost of approximately $15 per test. The cost-effectiveness and operational simplicity enable broader population coverage within the same budget, reduce the per-patient screening cost, and decrease the burden of unnecessary invasive procedures, making it easier to implement in large-scale populations, even in LMICs.

Compared with other blood-based lung cancer early detection tests, such as OncImmune’s EarlyCDT-Lung, utilizing a panel of seven autoantibodies with a 41.0% sensitivity at 91.0% specificity (15), and DELFI’s FirstLook-Lung based on the fragment patterns from cell-free DNA with a 84.0% sensitivity at 50.9% specificity (37), LungCanSeek demonstrated superior overall performance with 83.5% sensitivity at 90.3% specificity. Notably, LungCanSeek also achieved 77.4% accuracy in classifying lung cancer into three histological subtypes, further supporting its potential clinical utility. The subtype prediction capability of LungCanSeek may serve as a supportive tool in clinical decision-making, particularly in situations where tissue samples are unavailable or inadequate. This capability is beneficial for directing diagnostic workups. It also holds potential significance for patients whose subtype remains unidentified.

The screenings for colon and prostate cancers are two steps for avoiding financial burdens on the individuals and public healthcare system. Prostate cancer screening starts with a blood test of prostate-specific antigen (PSA) and then uses magnetic resonance imaging (MRI) to rule out the false positives (44), Colon cancer screening starts with a stool-based fecal immunochemical test (FIT) and then uses colonoscopy to confirm the positive findings (45). The two-step approach is much more cost-effective than using the costly and resource-demanding MRI and coloscopy as the primary screening, which has been implemented worldwide, especially in countries with universal health coverage. In lung cancer screening, we similarly propose a two-step approach, utilizing LungCanSeek for initial screening, followed by LDCT for LungCanSeek’s positive cases. Despite there being a population eligible for lung cancer screening of 15 million in the United States (37), only about 5% received recommended screening with a LDCT (10), primarily due to its complex procedure, radiation exposure (11), and high false-positive rate (10). LungCanSeek, as an easy-to-perform, non-invasive, and high-specificity blood test, is an excellent way to increase adherence with lung cancer screening, achieving a screening rate similar to that of colorectal cancer screening at around 60% (38). By modeling the two-step approach and LDCT in a lung cancer screening of 15 million high-risk adults with a 60% screening rate, the two-step approach reduced false positives by more than 10-fold and cost by 2.5-fold compared to LDCT. The two-step approach not only yields substantial economic benefits but also effectively reduces LDCT’s high false positives, avoiding unnecessary biopsies. This strategy can significantly reduce the psychological and physiological burdens on individuals and prevent the waste of healthcare resources, which is particularly valuable for lung cancer screening, given its global incidence rate of only 0.03% (1).

We recognized some limitations in this study. Despite the combined sensitivity for early-stage cases across all cohorts being 67.2%, which remains limited, compared to LDCT. Additionally, the relatively low proportion of early-stage cases in both validation cohorts may have impacted the stability of sensitivity estimates. Improving the sensitivity of the test in early-stage patients would prevent the cancer from progressing in asymptomatic patients, especially when LungCanSeek is used as a primary screening. Meanwhile, the evaluation of the two-step screening approach in this study is based on simulation modeling using retrospective data and published parameters. While the results provide preliminary insights into the potential benefits of this strategy, prospective validation is essential to confirm its clinical utility and cost-effectiveness.

Conclusions

In summary, this retrospective study demonstrates that LungCanSeek is a novel blood-based test for lung cancer early detection with superior and robust performance. It also provides accurate subtype prediction that may guide patients’ clinical management. LungCanSeek offers an affordable and practical solution for lung cancer early detection, particularly in LMICs. The two-step approach, which combines LungCanSeek with LDCT, not only significantly reduces the burdens, caused by LDCT’s high false positives, but also helps cut down a significant amount of screening cost, making it a cost-effective strategy for population-wide lung cancer screening.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Medical Ethics Committees of Henan Cancer Hospital (No. 2023-KY-0153) and Sun Yat-sen Memorial Hospital (No. SYSKY-2023-435-02). Informed consent was obtained from all individual participants. Shenyou Bio was also informed and agreed the study.

Data Sharing Statement

Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-317/dss

tlcr-14-09-3444-dss.pdf ^{(68.9KB, pdf)}

DOI: 10.21037/tlcr-2025-317