Lung cancer (LC) is the most prevalent and deadliest form of human cancer worldwide, accounting for 1.8 million deaths annually and 17.8% of all cancer-related fatalities [1]. While in some cases the diagnosis LC is straightforward, it remains challenging in others, particularly when imaging studies reveal indeterminate nodules [2,3]. Several studies have suggested that low-dose computed tomography (LDCT) screening can reduce mortality by LC; however, the results remain controversial [3–8]. Additionally, LDCT is a high-demand resource, leading to delays in early access, and the associated costs, along with the high incidence of indeterminate nodules, make it crucial to prioritize the use of this study in high-risk patients [4,5,9]. Furthermore, while advanced imaging such as PET-CT offers high diagnostic value, it is an expensive and often less accessible technique in many hospitals, limiting its use as a first-line triage tool.

Circulating tumour biomarkers (TB) are valuable diagnostic tools, particularly when used in combination to enhance sensitivity and specificity [10–13]. However, the optimal TB combination for maximizing accurate diagnosis of LC remains uncertain [10,12–19]. Previous studies by our group have demonstrated that a combination of six TB [carcinoembryonic antigen (CEA), cytokeratin fragment 21-1 (CYFRA 21-1), cancer antigen 15-3 (CA 15-3), squamous cell carcinoma antigen (SCC-Ag), progastrin-releasing peptide (ProGRP), and neuron-specific enolase (NSE)] correlates with the presence of LC and its major histological subtypes, i.e., non-small-cell LC (NSCLC) – adenocarcinoma (ADC) and squamous cell carcinoma (SCC) – and small-cell LC (SCLC) [11,20]. This combined TB model showed significantly greater diagnostic accuracy than a clinical model based solely on tumour size, age, and smoking status [20]. Although some studies indicate the potential benefit of integrating TB serum concentrations with LDCT for optimizing the diagnosis of LC [21], this approach has yet to be fully established. LDCT is a high-demand resource, leading to delays in early access, and the associated costs, along with the high incidence of indeterminate nodules, make it crucial to prioritize the use of this study in high-risk patients [4,5,9].

It is important to acknowledge that TB markers can yield false positives in certain pathophysiological conditions that require differentiation from LC. Incorporating clinical and laboratory variables that identify these conditions can enhance diagnostic accuracy.

To address this challenge, we developed expert software named CLAUDIA (Cancerous Lung Algorithm Useful for DIAgnosis). Using clinical variables, computed tomography (CT) data, and TB concentrations from 5000 patients, CLAUDIA calculates the risk of lung cancer (LC) and suggests a histological classification.

The study aimed to assess and validate the clinical utility of a TB-based software tool for decision-making in rapid-diagnosis pulmonary units in 12 hospitals in Spain and Portugal.

The diagnosis of LC was confirmed in 1392 (69.4%) patients, while 613 (30.6%) were found to have benign disease. Among the LC cases, 266 were SCLC (19.1%) and 1126 NSCLC (80.9%); adenocarcinomas were the most frequent subtype (n=635; 56.4%), followed by SCC (n=319; 28.3%) and cancers of indeterminate lineage (n=172; 15.3%) (Fig. 1).

In the suspected LC cases, the main finding leading to the diagnostic workup was the presence of radiographic nodules in 49.8% of cases, which were slightly more frequent in non-cancer than in LC patients (56.9% vs. 46.7%). Dyspnoea, haemoptysis, thoracic pain, and persistent cough were present in around 10%; thoracic pain and persistent cough were more frequent in LC, while haemoptysis was more common in non-cancer cases. Constitutional syndrome components occurred in 5% of LC cases versus 1.5% in non-cancer patients (Table 1).

Table 2 compares the clinical and imaging findings and TB concentrations between LC and non-LC patients. Significant differences were observed in gender, smoking habits, cigarette consumption per year, lung nodules (especially those >3cm), and TB concentrations, all of which were more prevalent in LC patients (p<0.01–0.001). Smaller nodules were more commonly found in non-LC patients. Within the LC subgroup analysis, NSCLC patients were generally older and had different TB profiles: CEA, CYFRA 21-1, SCC-Ag, and CA 15-3 levels were higher in NSCLC, whereas NSE and ProGRP were elevated in SCLC patients (p=0.001). All TB values were higher in NSCLC than in non-LC patients, except for SCC-Ag, which had similar values in SCLC and non-LC patients.

Table 3 summarizes the diagnostic sensitivity, specificity, NPV, and PPV of TB for predicting the risk of LC, both individually and in combination. The individual diagnostic sensitivity ranged from 19.4% (NSE) to 59.8% (CYFRA 21-1), while specificity was notably higher, varying from 89.6% (CYFRA 21-1) to 99.3% (NSE). This variation is observed because LC is not a single disease but consists of multiple histological subtypes with distinct behaviours, expression patterns and treatment responses. For example, NSE is predominantly elevated in SCLC, which accounts for approximately 20% of LC cases. Consequently, the sensitivity of NSE is low across all LC subtypes but remains highly specific for SCLC when elevated. By incorporating a comprehensive tumour marker panel, the software accounts for these variations, enabling accurate classification of LC subtypes based on expression patterns. Combined TB assessments, whether defined as a ≥1 abnormal TB marker or through software-based analysis, substantially improved diagnostic sensitivity (89.4% and 87.7%, respectively). The software demonstrated a higher specificity (75.5%) compared to ≥1 abnormal TB (63.9%) and exhibited the best NPV (73.03%) and PPV (89.06%).

The NRI analysis of the software led to an overall reclassification rate of 9.8% compared to the elevated ≥1 TB rule. Among 113 reclassified cases, 104 were changed from positive to negative (7.09%), while 9 shifted from negative to positive. Consequently, the software improved NRI by approximately 10% in distinguishing LC from non-LC cases.

The NB difference between the elevated ≥1 TB approach and the software was 0.07%, meaning that for every 14 cases analysed, one additional true cancer case was identified without increasing the false positives.

The prevalence and size of lung nodules were significantly greater in LC patients (Table 4). Across all nodule size categories (<1cm, 1–3cm, and >3cm), TB levels were significantly higher in LC patients (Fig. 3B). Fig. 2A illustrates the relationship between TB levels and the histological subtypes of LC. NSE and ProGRP showed greater sensitivity in detecting SCLC, whereas CYFRA 21-1 and CEA demonstrated higher diagnostic utility in NSCLC. Of note, NSE and ProGRP concentrations were comparatively lower in NSCLC. Furthermore, CYFRA 21-1 concentrations were markedly elevated in SCC, while CEA levels were significantly increased in ADC.

Fig. 2.

(A) The sensitivity of the tumour biomarkers in relation to histology. (B) Serum tumour biomarker sensitivity is subdivided according to tumour stage in NSCLC.

Fig. 2B reveals that, within NSCLC, the sensitivity of TB tends to rise proportionally with increasing tumour burden or more extensive dissemination.

Fig. 3A presents concordant findings for SCLC, further supporting the diagnostic relevance of NSE and ProGRP in this histological subtype.

Fig. 3.

(A) Serum tumour biomarker sensitivity is subdivided according to tumour stage in SCLC. (B) Probability of lung cancer according to serum tumour biomarker levels and nodule size (CT scan). ADC: adenocarcinoma; SCC: squamous cell carcinoma; uNSCLC: unspecific non-small-cell lung cancer; SCLC: small-cell lung cancer. CA 15.3: carbohydrate antigen 15.3; CEA: carcinoembryonic antigen; CYFRA 21-1: cytokeratin-19 fragment, NSE: neuron-specific enolase; ProGRP: progastrin-releasing peptide; SCC-Ag: squamous cell carcinoma-associated antigen. For further explanations, see the text. OR: odd ratio; cm: centimetre. TB: tumour biomarker; T1: initial analysis; T2: second analysis 3–4 weeks later; LC: patients with final diagnosis of lung cancer; No LC: patients with final diagnosis of no lung cancer.

Fig. 3B depicts the estimated probability of malignancy stratified according to nodule size and TB positivity, demonstrating that both increased nodule dimensions and TB positivity are associated with a significantly higher likelihood of cancer.

Table 5 summarizes the concordance between algorithm-predicted classifications and the definitive histological diagnoses. The algorithm achieved high concordance for SCLC (90.0%), NSCLC (91.2%), and its major subtypes, including ADC (79.4%) and SCC (63.4%). However, the performance was less robust in cases categorized as unspecified NSCLC (uNSCLC), as well as in the moderate-risk groups, which yielded concordance rates ranging from 8.7% to 38.0%. These findings demonstrate that the algorithm accurately distinguishes among the most common tumour types, while less well-defined or indeterminate categories remain a challenge. The heterogeneity analysis of the diagnostic accuracy across the 12 participating centres revealed mixed results. We found homogeneity in specificity between centers (p=0.081; I2=39.0%), indicating the algorithm's stable capacity to correctly identify non-malignancy. Conversely, significant statistical heterogeneity was detected for sensitivity (p=0.005; I2=59.6%). We tested the hypothesis that this variability was driven by differences in case-mix. When centers were classified into two clinical subgroups based on the percentage of early-stage patients (those with >25% and those with <20%) see Table S3, homogeneity was successfully demonstrated in both sensitivity and specificity within each subgroup, confirming that the observed heterogeneity was explained by differences in patient populations (Table S2).

Recent bibliometric analyses provide insights into the growing interest in TBs for the diagnosis of LC. A review of 990 publications from 2000 to 2022 highlights the sustained research focus on TBs and underscores the unmet need for expert software to analyse these biomarkers [29]. Our study addresses this gap by validating a six-TB panel (CEA, CA 15-3, SCC, CYFRA 21-1, NSE, and ProGRP) in conjunction with expert software. Previous research has demonstrated that this panel delivers high sensitivity (88.5%) and specificity (82%) in single determinations, with further improvements through serial testing [20]. In our multicentre study, the panel exhibited a similar sensitivity (87.7%), confirming its robustness and applicability in larger, diverse patient populations. Notably, this present study represents the first large-scale multicentre evaluation of these TBs with the software, as most previous investigations have been limited to smaller, single centre cohorts.

Our findings demonstrate that the CLAUDIA software achieves strong diagnostic performance (sensitivity 89%, specificity 73%), which is highly competitive with established methods described in the diagnostic literature, such as LDCT [4,5,7,30,31] and nucleic acid-based liquid biopsies [9,32]. Since our cohort has a high disease prevalence due to its focus on patients already undergoing diagnostic workup for highly suspicious clinical and radiological findings, the algorithm's performance requires further validation in the context of a low-prevalence screening setting before its utility can be fully established as a supplement to current screening procedures.

LC comprises various histological subtypes associated with distinct TB expression patterns. For instance, SCLC predominantly expresses NSE and ProGRP, while NSCLC subtypes exhibit higher levels of other TBs (Figs. 2A, 2B, 3A). This diversity underscores the need for a comprehensive TB panel capable of detecting all LC subtypes. As shown in Figs. 2B, 3A and Table 3, there is an association between the TB expression pattern and the histological subtype, which enables the software to suggest a particular histology based on the TB concentrations measured. Furthermore, tumour stage plays a crucial role in diagnosis as early-stage LC – characterized by smaller nodules – poses a greater diagnostic challenge than advanced disease. Previous studies, including our own, have demonstrated that abnormal TB levels significantly increase cancer risk across all nodule sizes. Specifically, patients with nodules smaller than one centimetre exhibit a fivefold increased risk if TB levels are abnormal, while those with nodules larger than three centimetres and abnormal TB levels have a >95% risk of LC [11,13,15,20,33]. Fig. 3B illustrates the strong association between nodule size and TB positivity in the probability of presenting cancer.

While essential, static imaging techniques such as CT scans have limitations in assessing the dynamic nature of tumour growth and behaviour. This challenge is particularly evident in aggressive tumours, which can progress rapidly over short timeframes. The software addresses this issue by integrating static imaging data (e.g., nodule size, shape, and presence) with dynamic biomarker data from TB concentrations. This combination of the two types of data significantly enhances diagnostic accuracy and facilitates earlier detection of LC.

Unlike artificial intelligence models [34], our approach utilises predefined cut-off values for different pathologies based on empirical data. It applies expert-driven rules to personalize these cut-offs, leveraging a robust, validated database of over 5000 patients, including healthy individuals, those with benign pathologies, and LC patients [11,13,20]. By proactively addressing potential sources of error, the software ensures a high level of reliability. It minimizes false positives, even in complex cases involving confounding factors such as pleural effusion or other non-cancerous conditions. In cases of diagnostic uncertainty, repeating measurements after 3–4 weeks significantly improves diagnostic accuracy, reaching a high specificity as already shown in previous studies by our group [20].

In our study, the software demonstrated notable improvements in diagnostic performance compared to TBs only. Specificity increased from 63.9% to 75.5%, PPV from 84.9% to 89.1%, and sensitivity was maintained at 87.7%. The NRI analysis revealed a 10% improvement in the accuracy of classification compared to using TB alone, which, according to some authors, could have a greater clinical impact than a 10% increase in the area under the curve in some analyses such as the risk-prediction [35], highlighting the clinical relevance of incorporating the expert software into diagnostic workflows. Our results demonstrate that the algorithm performs reliably for the primary histologic subtypes of LC, particularly SCLC and NSCLC, reaching concordance rates of 90.0% and 91.2%, respectively. Subtype-level discrimination was also strong for ADC and SCC, supporting the potential utility of the algorithm in routine diagnostic workflows. Nevertheless, accuracy substantially decreased in uNSCLC and in specimens classified as moderate risk, indicating that indeterminate or overlapping features may reduce algorithm precision in these categories. This limitation underscores the need for further refinement of the algorithm, potentially through training with a larger and more diverse dataset or by integrating additional diagnostic modalities. Ultimately, improving algorithmic performance in these challenging cases is critical for its broader implementation in clinical practice. These analytical features render this working model a valuable asset for Rapid Diagnosis Units. It facilitates swift stratification of patients based on their risk of developing LC, enabling the expeditious identification of patients who should be given high priority for treatment. Moreover, the ability to suggest the histological subtype of a tumour may provide critical guidance in determining the suitability of a patient for surgical intervention. In cases in which SCLC is indicated, concordance with the definitive histopathological diagnosis exceeds 90%. Early identification of the potential histological subtype also facilitates more targeted diagnostic strategies, such as obtaining sufficient tissue for next-generation sequencing, particularly in patients with tumours located in anatomically challenging regions.

The specificity achieved by the expert software in this study (75.5%) was slightly lower than the 82% reported in previous studies [20]. This discrepancy may be attributed to the multicentre nature of the trial, which involved routine clinical conditions across 12 hospitals and laboratories. Nevertheless, our findings in previous studies reinforce the potential of serial TB determinations in reducing false positives, emphasizing the importance of dynamic TB assessments in real-world clinical settings [20,36,37].

Overall, our findings highlight the potential of the six-TB panel and expert software as valuable complements to LDCT in early LC detection. The strong correlation between TB results and nodule size suggests that patients with small nodules and abnormal TB levels should be prioritized as high-risk groups. Compared to established screening methods for other cancers, such as faecal occult blood testing for colorectal cancer or prostate-specific antigen testing for prostate cancer, our TB panel and software demonstrate superior diagnostic performance, with advantages that include being non-invasive, cost-effective, widely available, easy to repeat in a short time and capable of achieving high specificity when combined with serial testing.

We conclude that this study confirms the effectiveness of a six-serum TB panel (CEA, CA 15-3, SCC-Ag, CYFRA 21-1, NSE, and ProGRP) combined with the CLAUDIA expert software for diagnosing LC. The CLAUDIA expert software further enhances diagnostic accuracy, achieving the highest sensitivity-to-specificity ratio (87.7% and 75.5%), as well as the best PPV (89%), and NPV (73.03%) among currently available diagnostic tests for LC. Moreover, the correlation between TB results and nodule size, along with a sensitivity of 70% in early stages, suggests that the TB lung panel and CLAUDIA expert software could serve as valuable complementary tools to LDCT scans for early detection of LC. The aim of the CLAUDIA algorithm was not to replace established imaging modalities but to serve as a simple, fast (providing results in 4h), standardized, and affordable tool that supports rapid clinical decision-making. The algorithm provides results within 4h and is designed to be applicable in any healthcare setting, regardless of local imaging resources.

We acknowledge that the observed prevalence of lung cancer in our study cohort (69.4%) is high. This was expected, as the population comprised patients referred to Rapid Diagnostic Units due to an existing clinical or radiological suspicion of malignancy, resulting in a naturally high pre-test probability. This high-risk context is precisely where clinicians need the most support in stratifying indeterminate cases and act promptly.

Importantly, CLAUDIA goes beyond simple binary risk stratification by predicting the most probable histological subtype in high-risk cases (Table 5). This crucial additional information allows clinicians to anticipate subsequent diagnostic steps or complementary molecular testing, significantly improving patient management.

The most relevant finding in our meta-analysis is the significant heterogeneity observed in sensitivity, which contrasts with the stable performance observed in specificity. Our subsequent subgroup analysis, classifying centres by the proportion of early-stage patients, effectively resolved this heterogeneity. This confirms that the variation in sensitivity is predominantly a function of the pre-test probability (i.e., the disease stage distribution) rather than fundamental differences in the core performance of the CLAUDIA algorithm or variations in laboratory methodology. This dependency on case-mix is a known phenomenon in diagnostic tests, and the restoration of homogeneity within defined subgroups strengthens the argument for the test's consistency when applied to similar patient populations.

Further research is needed to validate these findings and assess their clinical significance.

R. Molina, A. Barco and J. Trapé: study concept and design, statistical analysis, and manuscript drafting. They had full data and took responsibility for the integrity of the data analysis which they performed.

All authors collaborated in the recruitment of patients, data curation and submission of results for analysis.

The authors declare that the material has not been partial or totally produced with the help of any artificial intelligence software or tool.

Software CLAUDIA was funded in part by Roche Diagbostics SL.

The authors declare not to have any conflicts of interest that may be considered to influence directly or indirectly the content of the manuscript.