Multimodal modeling based on DNA methylation analysis in bronchoalveolar lavage fluid for early lung cancer detection Authors: Jincheng Liu, Chengyi You, Li Bai, Yangfan Lv, Wei Zeng, Bin Wang, Wen Zhang, Zansheng Huang, Dongfan Ye, Yuhang Guo, Ping Wang, Peihua Zhou, Heng You, Tianxiu Wu, Chuangye Wang, Bin Qing, Zhi Xu Published: 2026 ABSTRACT AbstractLung cancer diagnosis poses a significant clinical challenge, with emphasis on enhancing the positivity rate and accuracy of early detection. The use of bronchoalveolar lavage fluid (BALF) for detecting the methylation of ras-association domain family member 1 A (RASSF1A) and short stature homeobox 2 (SHOX2) genes has emerged as a novel molecular diagnostic technique for lung cancer. Nonetheless, this method’s positivity rate can vary due to factors such as BALF quality, and its diagnostic consistency is uncertain. It was a prospective diagnostic study with randomized sampling. In this study, 310 patients with lung lesions detected by computed tomography (CT) imaging were enrolled, and they were randomized 1:1 into pre-biopsy BALF group and post-biopsy BALF group. RASSF1A and SHOX2 methylation in BALF were detected, and CT images and tumor markers of patients were collected to develop a multimodal model based on BALF methylation for predicting malignant lung lesions. An internal validation set was employed to gauge the model’s effectiveness. The findings revealed a statistically significant increase in gene methylation positivity rate and pathological cytology rates in the post-biopsy BALF group compared to the pre-biopsy BALF group (P < 0.05). The model demonstrated an area under the curve (AUC) of 0.985 for predicting malignant lung masses and 0.903 for lung nodules in the training set. When tested on the validation set, the AUC for predicting malignant lung masses and lung nodules was 0.930 and 0.811, respectively. The multimodal prediction model constructed based on RASSF1A and SHOX2 methylation of post-biopsy BALF demonstrates a high predictive value for identifying malignant lung lesions.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-50119-8.Keywords: Bronchoalveolar lavage fluid, DNA methylation, Lung lesions, RASSF1A gene, SHOX2 geneSubject terms: Biomarkers, Cancer, Computational biology and bioinformatics, Diseases, Oncology Introduction Among all types of malignant tumors, lung cancer holds the highest rates of incidence and mortality globally1. According to the TNM staging system, the 5-year survival rate for patients with stage I non-small cell lung cancer (NSCLC) ranges from 68% to 92%, whereas for stage IV, it is less than 10%2,3. Consequently, increasing the rate of early detection is crucial for improving the 5-year survival rate for lung cancer patients. Currently, low-dose computed tomography (LDCT) has been extensively used for lung cancer screening, which can reduce the mortality rate among high-risk, asymptomatic lung cancer patients by 31%4. However, LDCT, with its false-positivity rate ranging from approximately 12.9% to 25.9%, is susceptible to misdiagnosis and over-treatment. Furthermore, long-term follow-up can lead to a higher cumulative radiation dose for patients5,6. The determination of whether lung lesions are benign or malignant relies on pathologic diagnosis, which is considered the gold standard. However, this process is limited by the accuracy of the sample collection and the pathologist’s personal experience7. Crucial for regulating gene transcription and expression, DNA methylation involves modifying the methylation status of Cytosine-phosphate-Guanine(CpG) islands present in gene promoters8. It has been established that heightened DNA methylation of oncogenes is intricately linked to tumorigenesis and often occurs during the initial phases of tumor development9. Specifically, researches indicate a strong correlation between lung cancer and the excessive methylation of the RASSF1A and SHOX2 genes10. Moreover, RASSF1A and SHOX2 have the advantage of high specificity and are widely recognized compared with other lung cancer-related genes. Compared to blood and other specimens, BALF offers the advantages of high tissue specificity, convenient sampling, and minimal invasiveness, making it an ideal specimen for methylation detection11. Studies reporting on the detection of RASSF1A and SHOX2 methylation in BALF for lung cancer diagnosis have shown overall sensitivities ranging from 71.5% to 81.0%, with specificities between 90% and 97.4%12,13. However, the sensitivity needs further improvement, and there is a paucity of research on the detection of two-gene methylation in lung nodules using BALF14,15. Bronchoalveolar lavage cytology is a widely used yet less sensitive method for diagnosing lung cancer16,17. Traditional bronchoalveolar lavage techniques may produce limited pathological cells, which can diminish the diagnostic effectiveness. To enhance the detection rate, it is crucial to refine the lavage method. Typically, BALF sampling is performed before bronchoscopy biopsy to avoid contamination with red blood cells, which can compromise the quality of cytomorphological analysis18,19. However, this sequence might lead to a decreased number of cancer cells in the BALF, subsequently resulting in lower DNA methylation detection levels20. To enhance the diagnostic precision of BALF in identifying lung cancer, it is crucial to modify the method of collecting BALF to maximize the yield of tumor cells and DNA within the sample. In this research, we compared the methylation levels of BALF and the positivity rate of pathological cytology obtained post-biopsy with those collected pre-biopsy. We constructed a multimodal prediction model for lung lesions based on BALF methylation, lung CT imaging features, and tumor markers. The objective was to verify the clinical value of this model in identifying lung lesions. Materials and methods Patient inclusion and exclusion criteria This study was a single-center, prospective diagnostic accuracy study with randomized allocation to BALF sampling methods. A total of 340 patients with lung lesions who attended the Second Affiliated Hospital of Army Medical University and underwent radial ultrasound-guided bronchoscopy from September 2022 to April 2024 were enrolled. The follow-up cut-off date was set as August 1, 2024. After excluding 10 cases with missing baseline data, 12 cases without histopathologic results, and 8 cases lost to follow-up, 310 patients were included in the final analysis. After exclusion, there were no missing values among the included samples. Histopathology diagnosed 217 cases as malignant and 93 as benign. Additionally, 137 patients with lung lesions who underwent bronchial radial ultrasound-guided examination from May 2024 to August 2024 were included as a validation set for internal validation, with 84 diagnosed as malignant and 53 as benign by histopathology. The study was approved by the Ethics Committee of the Second Affiliated Hospital of Army Medical University, and informed consent was obtained from participants or their family members. Inclusion criteria: (1) The screening result of chest CT indicated a pulmonary nodule or a lung mass lesion; (2) All patients received radial ultrasound-guided bronchoscopy and lung biopsy; (3) Patients’ age was over 18 years old; (4) Patients had been admitted to the hospital to complete blood routine, tumor markers, chest CT, bronchoscopy, and other examinations; (5) Informed consents were signed by each patient or their family member; (6) Patients had histopathology and cytology results. Exclusion criteria: (1) Individuals with a prior history of malignancies in other systems; (2) Those with severe cardiovascular disorders; (3) Incomplete clinical data; (4) Patients had no pathologic results. Follow-up: All patients were followed up every 2–3 months by reviewing their medical records through the hospital outpatient system or via telephone. Grouping of patients Patients were randomly assigned to the pre-biopsy or post-biopsy BALF groups using a computer-generated random number table (SPSS 26.0). To ensure allocation concealment, the randomization sequence was generated and securely maintained in sealed, opaque envelopes by an independent assistant who was not involved in the clinical procedures. Each envelope was opened only after informed consent had been obtained and patient enrollment had been confirmed. The card inside the envelope would inform the bronchoscopy operator whether the patient belongs to the pre-biopsy BALF group or the post-biopsy BALF group. Randomization was applied to the sampling procedure (timing of BALF collection relative to biopsy), not to therapeutic interventions. This design allows for unbiased comparison of diagnostic yield between the two collection methods. Furthermore, the pre-biopsy BALF group and the post-biopsy BALF group were each subdivided into the lung mass group and the lung nodule group based on the diameter size of the lesions (Fig. 1). This study was blinded to the patients and statistical analyses of the data, as the specific grouping was known to the operator performing the bronchoscopy. Pathologists interpreting histological and cytological specimens, as well as laboratory technicians performing methylation analyses, were blinded to group allocation and clinical background. Enrollment flow chart. Firstly, the final inclusion patients were divided into the pre-biopsy BALF group and the post-biopsy BALF group, which was the training set. Subsequently, 137 patients with post-biopsy BALF were included as a validation set. Examinations and clinical information The examination items included bronchoscopy under ultrasound guidance, chest CT scanning, and BALF methylation. Tumor markers comprised Carcinoembryonic antigen (CEA), Cytokeratin fragment antigen 21 -1 (CYFRA21-1), Squamous cell carcinoma antigen (SCCA), Neuron-specific enolase (NSE), and Pro-gastrin-releasing peptide (Pro-GRP). Clinical information collected included age, gender, smoking history, clinical stage, pathologic type, lesion size, bronchoscopic biopsy pathology, and pathological cytology. Procedure of bronchoalveolar lavage in each group All patients received radial ultrasound-guided bronchoscopy, using 1% tetracaine 5 mL nebulized inhalation. Once the bronchoscope was inserted through the nasal cavity, a thorough examination of both the healthy and affected side of the bronchial tubes was conducted, with particular attention to the suspected lesion site. When the radial ultrasound confirmed that the bronchoscope had reached the lung lesion, biopsies and BALF collection were performed according to the respective group assignments (Fig. 2). Flow diagram of BALF sampling. The enrolled patients were randomly divided into pre-biopsy and post-biopsy BALF groups according to the different lavage methods. HE, Histopathological examination; IHC, Immunohistochemistry. Models and equipment: radial ultrasound-guided bronchoscopy was performed using a CV290 bronchial endoscope and a MAJ-1720 endoluminal ultrasound ring-scanning small probe, both from Olympus, Japan. The bronchoscopy, radial ultrasound navigation, and bronchoalveolar lavage procedures were performed in adherence to the established guidelines provided by the American Thoracic Society21,22. High-resolution CT examination The patients underwent high-resolution CT examination with the following parameters: slice thickness of 1 mm, matrix of 512 × 512, pitch of 1.0 ~ 1.1, time of 1 s, current of 140 mA, and voltage of 120 kV. All CT images were interpreted by two senior qualified radiologists using a standard lung window (window width: 1500 HU; window level: − 500 HU). CT imaging features of lung lesions were documented, including the spiculation sign, lobulation, irregular margin, bronchial obstruction, bronchial truncation sign, vacuole sign, pleural retraction, and vessel passing through. RASSF1A, SHOX2 methylation detection The alveolar lavage samples were placed in a 15 ml centrifuge tube and spun at 2000 rpm for 5 min. The supernatant was discarded, and the precipitate was retained for DNA extraction using a standard kit from Tellgen (Shanghai, China). The extracted DNA was modified with sulfite, then amplified and sequenced via Real-time Polymerase Chain Reaction (PCR). Samples were categorized as positive or negative based on the threshold cycle of fluorescence signals from the PCR. Observation index Observation index: methylation positivity rate of the SHOX2 and RASSF1A gene in BALF, cytological positivity rate in BALF. Sample size calculation According to the pre-experiment result in this study, the positivity rate of BALF methylation of pre-biopsy group for lung mass patients was 62%, and that of pre-biopsy group for lung nodule patients was 30%, and it was expected that the positivity rate of the post-biopsy group for lung mass patients would be 87%, and the positivity rate of post-biopsy group for lung nodule patients would be 60%. In accordance with the pre-biopsy group: the post-biopsy group = 1:1, and setting α = 0.05, (1−β) = 0.90,considering 10% dropout rate, PASS 2021 software calculated at least 154 patients with lung masses and 126 patients with lung nodules needed to be included, totaling 280 patients. Statistical methods SPSS 26.0 was utilized for data analysis. Measurement data were expressed as (± s), and analyzed using t-tests for comparison between groups. Count data were presented as [n(%)] and analyzed using chi-square tests. The lung lesion prediction model was constructed by screening predictors through univariate and multivariate logistic regression analysis. To minimize overfitting, we calculated events per variable (EPV). The EPV was 20.7 for the lung mass model and 24.5 for the lung nodule model, exceeding the recommended threshold of 10 events per variable. The multimodal predictive model analysis was pre-specified as a secondary objective to evaluate the integration of post-biopsy BALF methylation with imaging and tumor markers. Subsequently, the receiver operating characteristic curve (ROC) was plotted to evaluate the model’s efficacy by using MedCalc 20.0. A two-sided test was applied, with P < 0.05 considered statistically significant. As all comparisons were based on pre-specified hypotheses with independent scientific rationale, no multiple comparison corrections were applied. Results Patient characteristics in each group In the lung mass group, 121 lung cancer cases were identified through histopathology, comprising 62 lung adenocarcinoma (LUAC), 42 lung squamous carcinomas (LUSC), 13 small-cell lung cancers (SCLC), and 4 unspecified types. Additionally, 49 benign lesions were diagnosed, including 43 chronic inflammations, 4 tuberculosis cases, 1 aspergillosis, and 1 neurosphincteroma. The lung nodule group revealed 96 lung cancer cases, comprising 68 LUAC, 13 LUSC, 7 SCLC, and 8 unspecified types. Furthermore, 44 benign lesions were diagnosed, including 33 chronic inflammations, 5 tuberculosis cases, 4 aspergillosis, and 2 inflammatory pseudotumor. The demographic, clinical, and radiological characteristics did not show a significant difference between the two groups, P > 0.05 (Table 1). Table 1Baseline clinical characteristics [n(%)].Lung mass (N = 170)P-valueLung nodule (N = 140)P-valuePre-biopsy BALF group (N = 85)Post-biopsy BALF group (N = 85)Pre-biopsy BALF group (N = 70)Post-biopsy BALF group (N = 70)Age(years)0.6370.455s61.49 ± 9.1160.78 ± 10.6461.41 ± 10.5462.70 ± 9.77Range30 ~ 7928 ~ 8119 ~ 8035 ~ 82Gender0.3660.732Male62 (72.94)68 (80.00)39 (55.71)42 (60.00)Female23 (27.06)17 (20.00)31 (44.29)28 (40.00)Histopathology0.7350.856Malignant59 (69.41)62 (72.94)47 (67.14)49 (70.00)Benign26 (30.59)23 (27.06)23 (32.86)21 (30.00)Smoking0.3740.310Yes54 (63.53)57 (67.06)30 (42.86)37 (52.86)No31 (36.47)28 (32.94)40 (57.14)33 (47.14)Pathological types0.8930.193LUAC32 (54.24)30 (48.39)37 (78.72)31 (63.27)LUSC20 (33.90)22 (35.48)5 (10.64)8 (16.33)SCLC5 (8.47)8 (12.90)1 (2.13)6 (12.24)Other types2 (3.39)2 (3.23)4 (8.51)4 (8.16)Stages0.9970.832Ⅰ4 (6.78)3 (4.84)21 (44.68)19 (38.78)Ⅱ5 (8.47)5 (8.06)3 (6.38)4 (8.16)Ⅲ20 (33.90)22 (35.48)11 (23.40)10 (20.41)Ⅳ27 (45.76)28 (45.16)8 (17.02)13 (26.53)Undefined stage3 (5.08)4 (6.45)4 (8.51)3 (6.12)Largest diameter (mm)Lung nodule0.20810–20––27 (38.57)19 (27.14)20–30––43 (61.43)51 (72.86)Lung mass0.19630–5048 (56.47)36 (42.35)––50–7025 (29.41)33 (38.82)––> 7012 (14.12)16 (18.82)––LUAC, Lung adenocarcinoma; LUSC, Lung squamous carcinoma; SCLC, Small-cell lung carcinoma. Baseline clinical characteristics [n(%)]. LUAC, Lung adenocarcinoma; LUSC, Lung squamous carcinoma; SCLC, Small-cell lung carcinoma. In both the pre-biopsy BALF group and the post-biopsy BALF group, no severe adverse events were observed. Minor bleeding occurred in 5 cases (3.23%) in the Pre-biopsy group and 4 cases (2.58%) in the Post-biopsy group (P > 0.05), all managed conservatively without intervention. No pneumothorax or other complications occurred. Comparison of BALF RASSF1A and SHOX2 methylation positivity rate in malignant cases In a comparison of malignant lung masses, post-biopsy BALF group showed 48 cases of SHOX2 methylation positivity rate, significantly higher than the 34 cases in the pre-biopsy BALF group (77.42% vs. 57.63%, P = 0.032). For RASSF1A methylation, 28 cases (45.16%) were positive in the pre-biopsy BALF group, while 25 cases (42.37%) were positive in the post-biopsy BALF group, with no significant difference (P = 0.855). The positivity rate for either RASSF1A or SHOX2 was significantly greater in the post-biopsy BALF group, with 55 cases compared to 37 in the pre-biopsy BALF group (88.71% vs. 62.71%, P = 0.001). When comparing malignant lung nodules, the post-biopsy BALF group had 22 cases of SHOX2 methylation positivity rate, significantly higher than the 6 cases in the pre-biopsy BALF group (44.90% vs. 12.77%, P = 0.001). Similarly, the post-biopsy BALF group had 20 cases of RASSF1A methylation positivity rate, compared to 9 cases in the pre-biopsy BALF group (40.82% vs. 19.15%, P = 0.021). The positivity rate for either RASSF1A or SHOX2 was greater in the post-biopsy BALF group, with 30 cases versus 11 in the pre-biopsy BALF group (61.22% vs. 23.40%, P < 0.001), as shown in Fig. 3. Comparison of BALF methylation positivity rate between two groups of malignant cases. (A) Comparison of BALF methylation positivity rate of the SHOX2 gene between two groups of malignant cases. (B) Comparison of BALF methylation positivity rate of the RASSF1A gene between the two groups. (C) Comparison of BALF methylation positivity rate of the SHOX2 and RASSF1A genes between the two groups. *P < 0.05; **P < 0.01; ***P < 0.001. Comparison of BALF pathologic cytology positivity rate of malignant cases In the malignant lung mass group, the positivity rate of post-biopsy pathological cytology was significantly higher than the pre-biopsy rate (51.61% vs. 30.51%, P = 0.026). In the malignant lung nodule group, the positivity rate of post-biopsy pathological cytology also significantly exceeded the pre-biopsy rate (32.65% vs. 2.13%, P < 0.001), as shown in Fig. 4. Comparison of BALF cytological positivity rate between two groups of malignant cases. *P < 0.05; **P < 0.01; ***P < 0.001. Comparison of BALF pathologic cytology positivity rate of malignant cases In a comparison of methylation positivity rates for RASSF1A and SHOX2 across clinical stages of malignant lung lesions, stage IV showed a significantly higher SHOX2 positivity rate than stages III, II, and I (61.84% vs. 58.73% vs. 41.18% vs. 25.53%, P = 0.001). Similarly, RASSF1A positivity rate was significantly higher at stage IV compared to stages III, II, and I (48.68% vs. 38.10% vs. 29.41% vs. 19.15%, P = 0.015). Furthermore, the positivity rate for either RASSF1A or SHOX2 at stage IV surpassed that of stages III, II, and I (72.37% vs. 68.25% vs. 52.94% vs. 34.04%, P < 0.001), as presented in Table 2. In the lung mass group and the lung nodule group, the positivity rates for either RASSF1A or SHOX2 were 52.94% and 34.04% in stage I–II tumors, and 80.00% and 52.38% in stage IV tumors, respectively. Table 2Relationship between two-gene methylation and clinical stages [n(%)].Clinical stagesN = 217SHOX2+RASSF1A+RASSF1A+/SHOX2+I4712 (25.53)9 (19.15)16 (34.04)II177 (41.18)5 (29.41)9 (52.94)III6337 (58.73)24 (38.10)43 (68.25)IV7647 (61.84)37 (48.68)55 (72.37)Undefined stages147 (50.00)7 (50.00)10 (71.43)P value0.001**0.015*< 0.001****P < 0.05; **P < 0.01; ***P < 0.001. Relationship between two-gene methylation and clinical stages [n(%)]. *P < 0.05; **P < 0.01; ***P < 0.001. In comparing the positivity rates of two-gene methylation across malignant lung lesions by type, LUSC showed a significantly higher SHOX2 positivity rate than SCLC and LUAC (85.45% vs. 85.00% vs. 33.08%, P < 0.001). Conversely, SCLC had a notably higher RASSF1A positivity rate compared to LUAC and LUSC (85.00% vs. 34.62% vs. 29.09%, P < 0.001). Additionally, the positivity rate for either RASSF1A or SHOX2 in SCLC was significantly greater than in LUSC and LUAC (90.00% vs. 85.45% vs. 48.46%, P < 0.001), as presented in Table 3. Table 3Relationship between two-gene methylation and pathological types [n(%)].Pathological typesN = 217SHOX2+RASSF1A+LUAC13043(33.08)45(34.62)LUSC5547(85.45)16(29.09)SCLC2017(85.00)17(85.00)Other types123(25.00)4(33.33)P value< 0.001***< 0.001***LUAC, Lung adenocarcinoma; LUSC, Lung squamous carcinoma; SCLC, Small-cell lung carcinoma. *P < 0.05;**P < 0.01; ***P < 0.001. Relationship between two-gene methylation and pathological types [n(%)]. LUAC, Lung adenocarcinoma; LUSC, Lung squamous carcinoma; SCLC, Small-cell lung carcinoma. *P < 0.05;**P < 0.01; ***P < 0.001. Characterization of CT imaging features and their diagnostic value Among all 310 cases, 217 were malignant and 93 were benign. Compared to the benign group, the malignant group showed significantly higher positivity rates for the spiculation sign (38.25% vs. 16.13%, P < 0.001), lobulation sign (34.10% vs. 11.83%, P < 0.001), bronchial obstruction (34.56% vs. 11.83%, P < 0.001), and bronchial truncation sign (23.96% vs. 10.75%, P = 0.008), as shown in Fig. 5. Comparison of positivity rate of CT imaging features between groups. *P < 0.05; **P < 0.01; ***P < 0.001. Diagnostic value of two-gene methylation of BALF for malignant lung lesions in post-biopsy BALF group In the post-biopsy BALF group, the malignant masses had a significantly higher positivity rate of RASSF1A compared to benign masses (45.16% vs. 0.00%, P < 0.001). Likewise, the positivity rate of SHOX2 was also significantly higher in the malignant group (77.42% vs. 8.70%, P < 0.001). Furthermore, the positivity rate for either RASSF1A or SHOX2 was significantly greater in the malignant group than the benign group (88.71% vs. 8.70%, P < 0.001), as shown in Table 4. Table 4Diagnosis of lung masses in post-biopsy BALF group by methylation [n(%)].MethylationGroupTotal(n = 85)SensitivitySpecificityPPVNPVP valueMalignant(n = 62)Benign(n = 23)RASSF1A+2802845.16100.00100.0040.35< 0.001***−342357SHOX2+4825077.4291.3096.0060.00< 0.001***−142135RASSF1A/SHOX2+5525788.7191.3096.4975.00< 0.001***−72128PPV, positive predictive value; NPV, negative predictive value; *P < 0.05; **P < 0.01; ***P < 0.001. Diagnosis of lung masses in post-biopsy BALF group by methylation [n(%)]. RASSF1A/ SHOX2 PPV, positive predictive value; NPV, negative predictive value; *P < 0.05; **P < 0.01; ***P < 0.001. In the post-biopsy BALF group for lung nodules, the malignant group’s positivity rates for RASSF1A and SHOX2 were significantly higher than those in the benign group (40.82% vs. 0.00%, P = 0.001; 44.90% vs. 4.76%, P = 0.001). Furthermore, the positivity rate for either RASSF1A or SHOX2 was also significantly greater in the malignant group compared to the benign group (61.22% vs. 4.76%, P < 0.001), as outlined in Table 5. Table 5Diagnosis of lung nodules in post-biopsy BALF group by methylation [n(%)].MethylationGroupTotal(n = 70)SensitivitySpecificityPPVNPVP valueMalignant(n = 49)Benign(n = 21)RASSF1A+2002040.82100.00100.0042.000.001**−292150SHOX2+2212344.9095.2495.6542.550.001**−272047RASSF1A/SHOX2+3013161.2295.2496.7751.28< 0.001***−192039PPV, positive predictive value; NPV, negative predictive value; *P < 0.05; **P < 0.01; ***P < 0.001. Diagnosis of lung nodules in post-biopsy BALF group by methylation [n(%)]. RASSF1A/ SHOX2 PPV, positive predictive value; NPV, negative predictive value; *P < 0.05; **P < 0.01; ***P < 0.001. BALF methylation complementary to bronchoscopic pathology biopsy All cases underwent bronchoscopic biopsy guided by radial ultrasound. In cases where the pathologic results were inconclusive, CT-guided percutaneous lung aspiration biopsy or surgery was performed to determine the final pathological results. In the lung mass group, the missed diagnosis patients were 11 cases, with the missed diagnosis rate of 9.09% and the positive methylation rate of 63.64%. In the lung nodule group, the missed diagnosis cases were 30, with the missed diagnosis rate of 31.25% and the positive methylation rate of 23.33% (Fig. 6). BALF Methylation as a supplement to bronchoscopic pathological biopsy. (A) Comparison of bronchoscopic pathology results of lung masses with the final pathological results, of which 11 cases were benign and the final pathological results were malignant. (B) Comparison of bronchoscopic pathology results of lung nodules with the final pathological results, of which 30 cases were benign and the final pathological results were malignant. Univariate regression analysis to identify factors influencing prediction of malignant lung lesions Univariate logistic regression analysis of lung masses in the post-biopsy BALF group identified RASSF1A and SHOX2 methylation, lobulation, and bronchial obstruction in CT imaging features, and NSE, SCC, and CYFRA-211 in tumor markers as risk factors for the prediction model of malignant lung masses. Similarly, univariate logistic regression analysis of lung nodules in the post-biopsy BALF group screened RASSF1A combined with SHOX2 methylation and spiculation in CT imaging features as risk factors for the prediction model of malignant lung nodules (Fig. 7). Univariate logistic regression screening of risk factors for malignant lung lesions. OR, Odds ratio. Multimodal model constructed based on RASSF1A and SHOX2 methylation in BALF Risk factors identified by univariate logistic regression were further analyzed using multivariate logistic regression. This analysis revealed RASSF1A and SHOX2 methylation, lobulation in CT imaging features, and CYFRA21-1 in tumor markers as predictors in the multimodal prediction model for malignant lung masses (Table 6). The regression formula was Logit (P) = − 3.990 + 4.034 × lobulation + 4.705 × RASSF1A and SHOX2 methylation + 2.882 × CYFRA21-1. Table 6Predictors of malignant lung masses screened by multivariate logistic regression.VariablesMultimodal model predictors of malignant lung massβOR (95% CI)P valueRASSF1A and SHOX24.40581.843 (5.602 ~ 1195.658)0.001**Lobulation4.15363.613 (1.597 ~ 2533.540)0.027*CYFRA21-13.07121.568 (1.542 ~ 301.772)0.023**P < 0.05; **P < 0.01; ***P < 0.001; OR, odds ratio. Predictors of malignant lung masses screened by multivariate logistic regression. *P < 0.05; **P < 0.01; ***P < 0.001; OR, odds ratio. This analysis revealed RASSF1A and SHOX2 methylation and spiculation in CT imaging features as predictors in the multimodal prediction model for malignant lung nodules (Table 7). The regression formula was Logit (P) = − 1.076 + 3.762 × RASSF1A and SHOX2 methylation + 1.965 × spiculation. Table 7Predictors of malignant lung nodules screened by multivariate logistic regression.VariablesMultimodal model predictors of malignant lung nodulesβOR (95% CI)P valueRASSF1A and SHOX23.76243.044 (4.818 ~ 384.550)0.001**Spiculation1.9657.132 (1.822 ~ 27.918)0.005** *P < 0.05; **P < 0.01; ***P < 0.001; OR, odds ratio. Predictors of malignant lung nodules screened by multivariate logistic regression. *P < 0.05; **P < 0.01; ***P < 0.001; OR, odds ratio. The ROC curves for the multimodal models of malignant lung masses and lung nodules were plotted, respectively. The AUC of the multimodal model for post-biopsy lung masses was 0.985 (95% CI 0.926–0.999), with a sensitivity of 96.36%, a specificity of 100.00%, and a Yoden’s index of 0.9636 (Fig. 8A). Similarly, the AUC of the multimodal model for post-biopsy lung nodules was 0.903 (95% CI 0.809–0.961), with a sensitivity of 65.31%, a specificity of 100.00%, and a Yoden’s index of 0.6531 (Fig. 8B). Multimodal models of malignant lung masses and nodules after biopsy. (A) Multimodal model of malignant lung masses after biopsy. (B) Multimodal model of malignant lung nodules after biopsy. Calibration curves and the decision curves were plotted for the malignant lung occupying lesions’ prediction model. In the lung mass group, both the apparent curve and the bias-corrected curve were closer to the ideal curve, suggesting a good fit of the model (Fig. 9A). In the decision curve of the malignant lung mass prediction model, when the threshold probability was > 6%, the net benefit of the “All” strategy for patients was higher than that of the “None” strategy. This indicates that the malignant lung mass prediction model has certain clinical value (Fig. 9B). Calibration curve and decision curve of malignant lung mass prediction model after biopsy. (A) Calibration curve of the post-biopsy malignant lung mass prediction model. The diagonal line represents the ideal curve, fitting the predicted values with the actual values as the apparent curve, and correcting values as the bias-corrected curve. (B) Decision curve of the post-biopsy malignant lung mass prediction model. The X-axis denotes the range of threshold probability for high risk, and the Y-axis denotes standardized net benefit. “All” indicates all patients were predicted to have lung cancer, while “None” indicates all patients were predicted to have benign lesions. In the lung nodule group, both the apparent curve and the bias-corrected curve were closer to the ideal curve, indicating an excellent fit of the model (Fig. 10A). In the decision curve of the malignant lung nodule prediction model, when the threshold probability was > 26%, the net benefit of the “All” strategy for patients was higher than that of the “None” strategy. This indicates that the malignant lung nodule prediction model has certain clinical value (Fig. 10B). Calibration curve and decision curve of malignant lung nodule prediction model after biopsy. (A) Calibration curve of the post-biopsy malignant lung nodule prediction model. The diagonal line represents the ideal curve, fitting the predicted values with the actual values as the apparent curve, and correcting values as the bias-corrected curve. (B) Decision curve of the post-biopsy malignant lung nodule prediction model. The X-axis denotes the range of threshold probability for high risk, and the Y-axis denotes standardized net benefit. “All” indicates all patients were predicted to have lung cancer, while “None” indicates all patients were predicted to have benign lesions. Model validation 137 patients with post-biopsy lavage were re-incorporated as the internal validation set. This set included 88 patients with lung masses, of which 63 were diagnosed as malignant and 25 as benign by histopathology, as well as 49 patients with lung nodules, with 21 diagnosed as malignant and 28 as benign. The basic clinical characteristics of the internal validation set and the training set showed no significant differences in demographic, clinical, and radiological features (P > 0.05), as shown in Table S1. The validation set data were substituted into the constructed prediction model formula to obtain the predicted values for malignant lung masses and malignant lung nodules, respectively. The ROC curves were then plotted. The AUC for malignant lung masses was 0.930 (95% CI 0.854–0.974), with a sensitivity of 92.06%, specificity of 82.61%, and a Yoden index of 0.7467 (Fig. 11A). The AUC for malignant lung nodules in the validation set was 0.811 (95% CI 0.674–0.909), with a sensitivity of 61.90%, specificity of 96.43%, and a Yoden index of 0.5833 (Fig. 11B). Validation set ROC curves for lung mass and lung nodule prediction models. (A) ROC of the validation set for the lung mass model. (B) ROC of the validation set for the lung nodule model. Discussion RASSF1A and SHOX2 methylation in BALF have shown a certain level of sensitivity and specificity as a diagnostic method for lung cancer23. The traditional method of BALF acquisition typically performs before biopsy, which may lead to fewer tumor cells and tumor cell DNA in the BALF, causing a low methylation positivity rate. In this study, we innovatively employed the post-biopsy lavage method to collect BALF and found this method could significantly improve the positivity rate of BALF methylation and pathological cytology. Finally, by combining post-biopsy BALF methylation, lung CT imaging features, and tumor markers, we constructed and validated the value of a multimodal prediction model for malignant lung lesions. In this study, the positivity rate of RASSF1A and SHOX2 methylation in the malignant post-biopsy BALF group of lung masses was 88.71%, higher than that in the pre-biopsy BALF group, which was 62.71% (P < 0.05). Similarly, the positivity rate of pathological cytology in the malignant post-biopsy BALF group of lung nodules was 32.65%, higher than that in the pre-biopsy BALF group, which was 2.13% (P < 0.05). These findings align with the hypothesis of our study. This indirectly verifies that the post-biopsy lavage method could enhance the quality of BALF, leading to improvements in both BALF methylation positivity rate and pathological cytology positivity rate. Additionally, we observed that a small amount of blood was mixed in the BALF obtained after biopsy, indicating that the tumor tissue structure was indeed disrupted, and hemorrhage might flush tumor cells from the distal terminal airway to the proximal end. This could also be one of the reasons for the increased number of pathological cells in BALF24. Moreover, along with the increase in tumor cells in BALF, the tumor cell DNA content may also increase, which could contribute to the improved positivity rate of BALF methylation. Therefore, if the purpose of the examination is to determine whether the lesion is benign or malignant, the post-biopsy lavage method should be recommended. Studies have shown that DNA methylation positivity rate is associated with the clinical stage and pathological types of lung cancer25. A particular investigation including 137 patients revealed a significant increase in SHOX2 positivity rate in advanced cancer stages (stages III and IV) compared to early stages (stages I and II)26. Conversely, the increase in RASSF1A positivity rate was less substantial. These observations harmonize with prior research, indicating that the concurrent positivity rates for SHOX2 and RASSF1A might be more prominent in the later stages of the disease. Although the sensitivity of DNA methylation analysis in BALF is limited in early stages (stage I-II: 34.04–52.94%), likely due to lower tumor burden and limited DNA release in small lesions, its clinical value lies in complementing routine pathology of bronchoscopic biopsy, particularly given the 31.25% missed diagnosis rate in lung nodules where methylation detected 23.33% of otherwise missed malignancies. DNA methylation positivity rates are also relevant to the pathological type of lung cancer14. SCLC shows a higher positivity rate than LUAC or LUSC. SHOX2 is more frequently methylated in LUSC than in LUAC, whereas the methylation positivity rate of RASSF1A is higher in LUAC than in LUSC. This may suggest that SHOX2 and RASSF1A play different roles in different pathological types of lung cancer. For optimal detection, combined gene analysis is recommended. Using post-biopsy lavage for differentiating between benign and malignant lung lesions not only offers better sensitivity but can also complement bronchoscopic histopathologic biopsy27. In our research, there was a certain rate of missed diagnosis in bronchoscopic pathology biopsy. The positivity rate of lung mass methylation in 11 missed cases was 63.64%, and the positivity rate of lung nodule methylation in 30 missed cases was 23.33%. This suggests that even if the bronchoscopic pathology results are benign, the results may be false-negative due to factors such as sampling. Post-biopsy BALF methylation detection can help reduce the rate of missed diagnosis in bronchoscopic biopsy for lung cancer28. In this study, based on the multivariate logistic regression in constructing the prediction model for malignant lung lesions, RASSF1A and SHOX2 methylation, lobulation in CT imaging features, and CYFRA21-1 in tumor markers were identified as predictors for the malignant lung mass model. Likewise, RASSF1A and SHOX2 methylation and spiculation in CT imaging features were predictors for the malignant lung nodule model. Tumor markers were not included in the malignant lung nodule model, suggesting that tumor markers have low predictive value for malignant lung nodules. This may be related to the small diameter of lung nodules and the low levels of tumor markers in serum29. Elevated CYFRA21-1 is commonly found in NSCLC, which accounts for about 80% to 85% of lung cancers, and thus has high predictive value for malignant lung masses30,31. Lobulation and spiculation, two key CT imaging features, exhibit distinct discriminatory capabilities in distinguishing between benign and malignant lung lesions32,33. This study demonstrated the high accuracy of the predictive models for malignant lung masses (AUC: 0.985) and lung nodules (AUC: 0.903) in the training set. Calibration curves confirmed the models’ precision in estimating actual risk. Decision curves indicated superior net benefits when using the models compared to when not using them within specific thresholds. Internal validation confirmed these findings, with AUCs of 0.930 for malignant lung masses and 0.811 for malignant lung nodules, affirming the models’ exceptional predictive power. However, this study still has several limitations. Firstly, it is based on a single-center cohort and lacks external validation, which may introduce potential selection bias. We plan to conduct a multicenter study in the future. Additionally, operator awareness may have had a potential impact on sampling quality. BALF methylation has shown promising prospects for the diagnosis of lung lesions. However, the methylation positivity rates of different genes vary among pathological types. The methylation positivity rate of LUAC is generally lower than that of LUSC and SCLC. Although the combination of RASSF1A and SHOX2 can compensate for the differences in methylation positivity rates between the two genes in different pathotypes, further exploration of genes with higher methylation positivity rates in LUAC is necessary. This may represent a new direction in lung cancer methylation research. Conclusions Our research demonstrated that the post-biopsy lavage method can improve the positivity rate of RASSF1A and SHOX2 methylation and pathological cytology in BALF, which may improve the diagnostic value of BALF detection to identify benign and malignant lung lesions. Additionally, the multimodal model based on post-biopsy lavage BALF gene methylation results has good predictive value for malignant lung lesions, which can be expected to reduce the rate of missed diagnosis of lung cancer and improve the accuracy of early lung cancer prediction. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements This work was supported by the Technology Innovation and Application Development Special Key Project of Chongqing, China (No: CSTC2021jscx-gksb-N0029) and the Chongqing Talent Project (Innovation Leader) in China (No: CQYC20220303525). Author contributions J.L.: Conceptualization, project administration, data curation, writing-original draft. C.Y.: Conceptualization, project administration, supervision, formal analysis, writing-original draft. L.B.: Conceptualization, formal analysis, writing-review and editing. B.W.: Conceptualization, writing-original draft, writing-review and editing. Y.L.: Resources, data curation. W.Z.: Resources, data curation. Z.H.: Methodology, resources, data curation, writing-original draft. P.Z.: Resources, data curation. P.W.: Resources, data curation. Y.G.: Methodology, resources, data curation. D.Y.: Data curation. H.Y.: Data curation. B.Q.: Data curation. C.W.: Data curation. T.W.: Data curation. W.Z.: Formal analysis, funding acquisition, writing-original draft. Z.X.: Conceptualization, formal analysis, supervision, funding acquisition, writing-original draft, writing-review and editing. Funding This work was funded by the Technology Innovation and Application Development Special Key Project of Chongqing, China (No: CSTC2021jscx-gksb-N0029) and the Chongqing Talent Project (Innovation Leader) in China (No: CQYC20220303525). Data availability The datasets generated during this study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Institutional review board statement Approval of the research protocol by an Institutional Review Board: The study received approval from the Ethics Committee of the Second Affiliated Hospital of Army Medical University.Registry and the Registration No. of the study/trial: The study is registered with the Ethics Committee of the Second Affiliated Hospital of Army Medical University (No: 2022-Research 415-01). The clinical trial is registered with the Chinese Clinical Trial Registry (No: ChiCTR2300077871) on 22th November 2023. Informed consent Informed consent was obtained from all patients or their guardians. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Jincheng Liu and Chengyi You contributed equally to this work. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets generated during this study are available from the corresponding author on reasonable request.