Pulmonary sarcomatoid carcinoma (PSC) is rare among all lung malignancies, with an incidence of 0.1–0.5% (1, 2). According to the World Health Organization classification guidelines in 2021, PSC consists of giant cell carcinoma, pleomorphic carcinoma, spindle cell carcinoma, pulmonary blastoma, and carcinosarcoma (3). Clinically, newly diagnosed patients with PSCs are more likely to be male (77%), smokers (84%), and of advanced age, with a median age of 68 years at diagnosis (4). Similar to other subtypes of non-small cell lung carcinoma (NSCLC), the initial symptoms of PSC include cough, chest pain, weight loss, and dyspnea (5).

PSC is a highly aggressive malignancy. Recent studies have reported that the 5-year survival rate ranges from approximately 15% to 20% (6, 7), which is considerably lower than that in other subtypes of NSCLC. Targeted drugs, especially focusing on mesenchymal-epithelial transition (MET) exon 14 skipping mutations, have been shown to be beneficial in improving the median survival time to 10 months (8). However, distant metastasis (DM) still results in most patients having a shorter survival time (9–11). Owing to the rapid invasion into vasculature, 40–60% of PSC patients are diagnosed with DM at first presentation (6, 9, 12). The most common sites of DM are bone (16%), lungs (15%), brain (12%), and liver (8%), while 62% of patients present with multiple DM sites (13). The 1- and 3-year overall survival (OS) probabilities of PSC patients with DM are reported to be only 14.1% and 5.5%, respectively, which are significantly lower than those of patients without DM (58.2% and 38.1%, respectively) (13). Therefore, it is of great clinical significance to identify those at risk of developing DM upon initial stage of diagnosis.

Computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography-computed tomography (PET-CT) are commonly used diagnostic modalities for screening for DM in newly hospitalized patients with PSC. However, these screening methods are difficult to apply to all hospitalized PSC patients due to their high cost, including time-related costs. Meanwhile, due to its rare incidence, the prognostic factors of PSC patients with DM remain unclear. Consequently, research investigating the factors influencing survival is important as it can inform and guide clinical strategies. Since the emergence of big data analysis and machine learning (ML), it is possible to provide an alternative option for factors screening. There have been several predictive models with outstanding performance being built to apply in clinical practice by using big data and ML (14–16). The Surveillance, Epidemiology, and End Results (SEER) database (https://seer.cancer.gov/) covers geographically diverse patients with detailed information on the patients’ clinicopathological statistics and follow-up visits, providing an abundance of medical cases to analyze. This real-world-based database offers a golden opportunity to develop or examine ML models in this field. However, to our knowledge, no studies have focused on establishing models for the prediction and prognosis of DM in PSC patients.

Therefore, this study aimed to establish and validate diagnostic and prognostic models based on ML algorithms and Cox regression through large population and ML. Additionally, this study aimed to offer personalized predictors that could serve as effective tools for clinicians to preliminarily evaluate the risk and prognosis of PSC with DM, saving patients from unnecessary costs.

PSC presents a rare lung malignancy with a high probability for DM. Several studies having reported that DM is an independent risk factor for PSC (6, 17, 18). In this study, approximately 45.2% patients had DM at initial presentation, similar to that reported by Zombori-Tóth and Terra (19, 20). The high rate of DM may be attributed to its capacity to invade into vasculature and the high likelihood of DM recurrence. Lococo et al. analyzed the pathological results of 143 patients and found the high incidences of lymphatic (30%) and vascular emboli (68%) (21). Liang et al. found that MET mutations were commonly expressed in PSCs with a high rate of 16% (22), whereas the incidence of KRAS mutations was 22%. Similarly, Liu et al. have identified that eight out of 36 (22%) patients harboring MET mutations (23). This incidence was considerably more frequent than that in other NSCLC subtypes (3%) (24). A study on 77 Chinese patients with PSC indicated that positive METex14 skipping mutations rate was 20.8%, leading to worse disease-free survival (DFS) (25). In recent years, immune checkpoint inhibitors (ICIs) and targeted drugs for MET exon 14 skipping mutations have exhibited satisfying benefits in improving patient survival (6, 26, 27). However, patients who developed DM still had a significantly worse prognosis. Thus, early attention to DM is important to improve prognosis and help clinicians make the most optional targeted therapeutic decision.

CT, MRI, and PET-CT are conventional radiological screening methods for patients with NSCLC. However, performing all these radiological examinations is costly for newly diagnosed patients. In addition, time-consuming screening processes and potential side effects can delay the patients’ course from diagnosis to therapy. Therefore, there is an urgent need for a noninvasive, precise, and rapid method to assist in identifying potential DM at the initial hospitalization stage and to estimate the prognosis for PSC patients with DM. In this study, six ML algorithms for predicting DM in patients with PSC and a nomogram for evaluating the prognosis for those with DM were developed. An automatic calculator based on the best-performing algorithm was created and published online for public access. Moreover, the prognostic nomogram could accurately identify risk factors in PSC patients with DM and help clinicians evaluate survival.

Few studies have identified the risk factors for DM in patients with PSC due to its rarity. Logistic regression analysis was performed and found that histologic subtype, T stage, and N stage were independent risk factors for DM in patients with PSC. Among these factors, the N3 stage was most correlated with DM. Patients with the N3 stage had the highest OR of 9.72 (P < 0.001), followed by the T4 stage (OR 4.30, P < 0.001), N2 stage (OR 3.09, P < 0.001), giant cell carcinoma (OR 2.56, P < 0.001), T3 stage (OR 2.36, P = 0.022), and spindle cell carcinoma (OR 2.12, P < 0.001). These findings were similar to those reported by Xiao et al., who retrospectively analyzed 934 PSC patients using the SEER program database and reported that spindle cell carcinoma (OR 3.151, P < 0.001) and giant cell carcinoma (OR 4.023, P < 0.001) were independent risk factors for DM (13).

Similarly, T-stage and N-stage have been reported as risk factors for DM in lung adenocarcinoma and squamous cell carcinoma (28, 29). The two variables have also been identified as important factors for the development of DM in patients with PSC (28, 29). These three clinical features were then incorporated into the building procedures of the six ML models. The XGB model performed better than the other algorithms, with an AUC of 0.821, showing excellent predictive ability for DM in patients with PSC. In addition, calibration and DCA curves indicated that the model was highly consistent with actual observations and showed better clinical applicability (30).

Carcinoembryonic antigen levels, cytokeratin 19 fragment (CYFRA21-1), and other serologic indicators which were associated with DM in patients with NSCLC (31, 32) have not been included in this study due to missing data in the SEER program. However, the relatively high accuracy of the XGB algorithm still helped identify those PSC patients with a high risk for DM. To facilitate the use of the ML predictor, a web-based calculator was developed, which is likely to help clinicians detect DM early in patients with PSC in a convenient manner.

Compared to other subtypes of NSCLC, PSC is less sensitive to chemoradiotherapy. However, studies have confirmed that chemoradiotherapy significantly prolonged the CSS of PSC patients with DM. Xiao et al. analyzed the prognostic risk factors of 512 patients with metastatic PSC diagnosed between 1975 and 2016 in the SEER program (13). They found that having received chemotherapy (HR 0.308, P < 0.001) was an independent prognostic risk factor for PSC patients with DM, which was similar to our study findings (HR 0.29, P < 0.001). However, they did not identify radiation therapy and surgical resection as independent prognostic factors for PSC patients with DM, while these two therapies were indicated to prolong CSS significantly in our study. We speculate that this discrepancy was due to diagnosis time of the enrolled patients were prior to the year 2000 while surgical and radiation techniques were not so effective. Our study identified T-stage and N-stage as important prognostic risk factors for PSC patients with DM. Advanced T stage (T4, HR 2.52, P = 0.005; T3, HR 2.10, P = 0.03) and N-stage (N3, HR 1.88, P = 0.002; N2, HR 1.60, P = 0.005) correlated with poor CSS in PSC patients suffered from DM.

Interestingly, we have also found that the nodule site appeared to affect the prognosis for PSC patients with DM. CSS was longer in patients whose primary site was located on an upper rather than a lower lobe (HR, 0.68; P = 0.011). The reason may be that the nodules on the lower lung will develop more DM lesions leading a worse prognosis (13). Notably, we established a nomogram for predicting the prognosis for PSC patients with DM, with a relative higher AUC at 3 months and 6 months respectively. Its high consistency to actual observations indicates that this nomogram could precisely predict CSS in PSC patients with DM.

There are some limitations in our present study. First, this was a retrospective study, and selective bias could not be avoided. Second, characteristic information about DM was collected during initial hospitalization, which may have led to an underestimated percentage of DM in patients with PSC. Lastly, the predictive model was externally validated using patients’ information from Southwest Hospital in China. However, since this hospital mainly treats Chinese patients, the model’s application value should be further validated in cohorts involving differing ethnicities.

Our study indicated that PSC patients with advanced T-stage, N-stage, histology of giant cell carcinoma and spindle cell carcinoma were risk factors for DM and should receive more attention in terms of preventative therapeutic strategies. The AUC, accuracy, sensitivity, and specificity of the XGB model reached 0.821, 0.755, 0.757, and 0.754, respectively. A diagnostic model for DM based on the ML algorithm and a web-based predictor was then established, which could conveniently and precisely predict the risk of DM in PSC patients. Our study provided initial predictive and prognostic models for PSC patients with DM. Future studies may focus on further improving the models by adding other potential variables and developing more detailed models to predict the risk and prognosis for specific metastatic sites.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/ Supplementary Material .

The studies involving human participants were reviewed and approved by Southwest Hospital, Third Military Medicine University. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

XY and WX designed the study. GT, LZ, and KW collected the primary data. XY drafted the manuscript. HL and XZ reviewed and edited the paper. All authors contributed to the article and approved the submitted version.