H-X model: A novel tool for predicting pulmonary embolism in lung cancer

To the Editor: Lung cancer (LC) is the most common solid tumor and the leading cause of tumor-related deaths worldwide.^[1] Survival among patients with LC is influenced not only by the progression and treatment of the tumor itself but also by complications. Venous thromboembolism (VTE) is one of the main causes of cancer-related death, with cancer patients complicated by VTE showing a 3−4-fold increase in mortality compared with patients without VTE. Pulmonary embolism (PE) is a subtype of VTE in which an embolus blocks the blood vessels of the pulmonary artery tree. The risk of PE is nearly 20 times greater among cancer patients than among cancer-free patients.^[2]

Although computed tomographic pulmonary angiography (CTPA) serves as the current first-line diagnostic and evaluative tool for PE, its application in primary hospitals is limited by prolonged examination times and high costs. Thus, it is important to develop a clinical prediction model based on risk factors. Although multiple factors influence the occurrence of PE in patients with LC, few models have incorporated these risk factors into predictive and diagnostic models. The Wells score, Geneva score, Khorana score, and their modified versions are the most widely used VTE evaluation models^[3]; however, they are not specific for PE among patients with LC.

This study aims to comprehensively analyze the risk factors for PE among patients with LC by using data from two independent clinical centers to establish a convenient, highly specific, and sensitive clinical model for predicting the occurrence of PE.

This study was conducted in accordance with the Declaration of Helsinki and approved by the West China Hospital Ethics Committee (No. 2024-1763). The requirement for informed consent was waived.

This retrospective study enrolled patients with LC from West China Hospital (January 2011 – December 2021) and Sichuan Provincial People’s Hospital (January 2017 – December 2021). Follow-up data were censored in January 2023. Details of the study design, data collection, and statistical analysis are provided in Supplementary Figure 1 and the Supplementary Methods, http://links.lww.com/CM9/D3.

Of the 72,235 patients with pathologically confirmed primary lung cancer, 304 were diagnosed with PE. We then calculated the prevalence of PE and found that it showed a steadily rising trend year by year, as illustrated in Supplementary Figure 2, http://links.lww.com/CM9/D3. We subsequently excluded 12 PE patients with incomplete clinical information, and 292 patients with PE were ultimately included in the West China Hospital cohort [Supplementary Table 1, http://links.lww.com/CM9/D3]. After applying propensity score matching (PSM), 292 primary LC patients without PE were enrolled as matched controls. Baseline characteristics were compared using the chi-square test for categorical variables and the independent-samples t-test/Mann–Whitney U test for continuous variables; no significant intergroup differences were observed after matching (P >0.05) [Supplementary Table 2, http://links.lww.com/CM9/D3]. The 584 patients were randomly assigned to the training set and the internal validation set at a ratio of 3:1. Most factors did not significantly differ between the training and internal validation sets [Supplementary Table 3, http://links.lww.com/CM9/D3]. In the training set, significant differences (P <0.05) in thrombosis-related symptoms, tumor metastasis or progression, pathological type, antitumor treatment, and coagulation-related indicators were observed between the PE and non-PE groups. In addition, 100 patients from Sichuan Provincial People’s Hospital (37 LC patients with PE and 63 LC patients without PE) composed the external validation set. The clinical features of all PE cases are summarized in Supplementary Table 1, http://links.lww.com/CM9/D3.

Univariable logistic regression was performed to screen predictors of PE by comparing clinical characteristics between PE and non-PE patients in the training set [Supplementary Table 4, http://links.lww.com/CM9/D3]. Only variables with both odds ratio (OR) >1 and P <0.1 in univariable analysis were retained for subsequent LASSO regression and multivariable logistic regression (LR) modeling [Supplementary Table 5, http://links.lww.com/CM9/D3]. Several predictors were eliminated sequentially: four variables were excluded as their regression coefficients were reduced to zero; hypertension and the number of metastatic organs were dropped due to insufficient evidence linking them to LC-associated VTE; prior DVT was also removed to avoid selection bias caused by missing data. We constructed models using combinations of the remaining 11 risk factors: heart rate (HR) >100 beats per minute (bpm), tumor progression, chemotherapy, bilateral lung metastasis, platelet (PLT) count, white blood cell (WBC) count, neutrophil percentage, monocyte percentage, prothrombin time (PT), activated partial thromboplastin time (APTT), and fibrinogen (Fg) level.

We subsequently used traditional LR, decision tree, and random forest (RF) models to predict the occurrence of PE among patients with LC. The RF model achieved the best performance in both the training set and internal validation set, with area under the curve (AUC) values of 0.97 and 0.87, sensitivity of 0.87 and 0.84, and specificity of 0.91 and 0.87, respectively [Figure 1A−B and Supplementary Table 6, http://links.lww.com/CM9/D3]. This model, which was chosen as the final prediction model and named the H-X model, was validated in the external validation set (AUC = 0.85) [Figure 1C]. In addition, the calibration curves [Supplementary Figure 3, http://links.lww.com/CM9/D3] verified the agreement between predicted and observed PE risk. The observed curve closely matched the ideal diagonal across the training, internal validation, and external validation cohorts, demonstrating favorable model calibration. Finally, we compared the H-X model with commonly used evaluation methods (the Wells score and Khorana score). The H-X model (AUC = 0.87, sensitivity = 0.84, specificity = 0.87) developed herein showed superior performance in predicting LC-related PE, with superior predictive performance compared with the Wells score (AUC = 0.80, sensitivity = 0.75, specificity = 0.72) and Khorana score (AUC = 0.70, sensitivity = 0.65, specificity = 0.62) [Figure 1D and Supplementary Table 6, http://links.lww.com/CM9/D3]. Detailed predictive performance indicators are summarized in Supplementary Table 6, http://links.lww.com/CM9/D3.

Figure 1.

ROC curves for different prediction models. (A) Comparison of logistic regression, decision tree, and random forest models in the training set. (B) Comparison of the three models in the internal validation set. (C) ROC curves of the H-X model in the training set, internal validation set, and external validation set. (D) Comparison of discriminative performance between the H-X model and two classical predictive scores (Khorana score, Wells score). AUC: Area under the curve; ROC: Receiver operating characteristic.

LC is one of the malignancies with the highest risk of PE. A meta-analysis encompassing 41 studies reported an overall incidence of LC-related PE of 3.7%, with individual study incidences ranging from 1.3% to 23.7%. Moreover, the actual incidence may be much higher due to asymptomatic occult PE.^[4] Previous studies revealed that vascular endothelial cell damage, altered blood flow status, and hypercoagulability are the three major elements of thrombosis.^[5] LC progression simultaneously disrupts all three components, thereby facilitating thrombogenesis.

Hematological derangements unique to LC patients further exacerbate tumor-induced hypercoagulability, which served as a critical reference for variable selection during model construction. Patients with LC frequently exhibit leukocytosis and thrombocytosis.^[6] Leukocytosis promotes the formation of neutrophil extracellular traps (NETs). NETs increase the risk of thrombosis by releasing extracellular chromatin and helping to trap PLT and microvesicles to stabilize thrombi. Moreover, thrombocytosis can lead to an increased risk of thrombosis through its own action or in combination with WBCs. PLT contributes to the adhesion among leukocytes, PLT and tumor cells through enhanced P-selectin release. PLT promotes tumor cell metastasis, and this process also promotes the formation of a hypercoagulable state. Numerous studies have demonstrated that D-dimer serves as a negative predictor for thrombosis. However, D-dimer elevation can be triggered by multiple factors, including various anti-tumor therapies. To date, there is no well-established strategy to adequately adjust for treatment-related and other confounding factors, making it impossible to isolate the signal of thrombus formation reflected solely by D-dimer. In our cohort, D-dimer concentrations were elevated above the normal threshold across all subgroups. No significant intergroup difference in D-dimer levels was observed between patients with LC complicated by PE and those without PE (P >0.05). Accordingly, D-dimer was excluded from our predictive model.

Beyond hematological indicators, we integrated tumor features and treatment parameters into the H-X model to improve its predictive performance for LC-related PE.^[7] Advanced LC patients with distant metastasis exhibit increased expression of tumor procoagulants, prolonged hypercoagulability, elevated prothrombin/cytokine levels, and increased thrombotic risk, which supports the inclusion of bilateral lung metastasis in our model. Future large-sample prospective studies are needed to determine whether bilateral lung metastasis confers a greater thrombotic risk than other metastatic sites. In addition, chemotherapy increases thrombotic risk across LC stages, potentially through vascular endothelial damage, cytokine release, reduced endogenous anticoagulants, and blood cell alterations (e.g., WBC count/PLT count/coagulation function), all of which would influence PE formation.

Given the above pathological and clinical rationales for variable selection, the H-X model integrates tumor characteristics, HR >100 bpm, treatment history, and blood indicators—all of which are readily available in LC clinical care settings. As an ensemble learning method, RF offers unique advantages in handling complex data and enhancing prediction stability via multiple decision trees. Its incremental learning capability enables dynamic risk prediction using longitudinal follow-up data, which fits well with the progressive nature of LC. The model outperforms traditional VTE scores (Wells, Khorana), which facilitates prompt PE identification in LC management, improving patient outcomes.

The limitations of this study are as follows. First, although this study was an independent two-center study, the sample size was relatively small. In the future, more research centers could be added, and the sample size could be expanded to verify our results. Second, the model does not include imaging features (e.g., computed tomography [CT] changes) or further subcategorize patients with LC. Third, during parameter selection, the model only included risk factors with an OR >1, while protective factors were excluded. In real‑world clinical practice, the protective effects of these factors against pulmonary embolism should not be underestimated. Consequently, this limitation may lead to an overestimation of risk when the model is applied clinically. To ensure accuracy and comprehensiveness, future iterations of the predictive model should jointly incorporate both risk and protective factors.

In conclusion, we observed a rising trend in the incidence of LC-related PE. The H-X model, which included eleven easily accessible parameters (HR >100 bpm, tumor progression, chemotherapy, bilateral lung metastasis, PLT count, WBC count, neutrophil percentage, monocyte percentage, PT, APTT, Fg level), demonstrated excellent predictive performance in the validation sets and outperformed existing PE models. This tool may assist clinicians in assessing the risk of PE during LC management.

Funding

The study was supported by the Noncommunicable Chronic Diseases-National Science and Technology Major Project (No. 2023ZD0506102/2023ZD0506100); Special Fund of State Key Laboratory of Respiratory Health and Multimorbidity (Nos. 2060204 and 2060204-2026); 1·3·5 Project of State Key Laboratory of Respiratory Health and Multimorbidity, West China Hospital, Sichuan University (No. RHM24101).

Conflicts of interest

None.

Supplementary Material

cm9-139-2997-s001.docx ^{(251.9KB, docx)}