The accurate prediction of survival in patients with LPADC is essential for patient counseling, follow-up, and treatment planning. Previous studies have revealed multiple prognostic factors that affect the survival time of patients with pulmonary papillary carcinoma, including patient age, grade classification, lymph node status, tumor size, distant metastases, and surgical treatment9, 11. Machine learning is increasingly utilized in research for the prediction of survival of patients with cancer25,26,27, with relatively favorable results. Although CoxPH is the classical method utilized for the analysis of survival data, the use of this method requires linear relationships between variables. As a result of the continuous advances achieved in recent years, machine learning is widely applied to the medical field28,29,30. In this study, we used ensemble machine learning models to accurately predict CSS in patients with LPADC, and obtained satisfactory results.
Consistent with the findings reported by You et al., the four models developed in this study confirmed that surgery is an important prognostic factor for patients with lung adenocarcinoma3. Similarly, distant metastases have an important impact on the prognosis of patients with LPADC. In conjunction with previous analyses, the findings demonstrate that patients who developed distant metastases had poorer survival rates than other patients26, 27. A higher N-stage also plays a crucial role in the model, indicating poor prognosis28. Other characteristics (e.g., tumor size, grade, sex, chemotherapy, primary site, etc.) have different degrees of importance in various models11, 23, 27. These results suggest that the selection of appropriate treatment modalities (e.g., surgery, radiotherapy, and chemotherapy) may be more important for predicting CSS in patients with LPADC than TNM staging alone.
Interestingly, the ensemble models (i.e., GBS, EST, and RSF) did not demonstrate a markedly better ability for predicting CSS in LPADC in the validation cohort compared with the CoxPH model. This indicates that the machine learning approach may only offer advantages when traditional models are limited. Therefore, there are several possible explanations for the comparable predictive performance observed between the ensemble and CoxPH models in this study. Firstly, the number of predictors used to construct the model was not sufficiently large, and the advantages of machine learning in analyzing large samples and multivariate data are not fully realized. Secondly, the SEER database collects variables derived from clinical experience; many of these variables are linearly correlated with outcomes. Therefore, the data may be better qualified for the application of parametric (CoxPH) models. The GBS, EST, and RSF models developed in this study achieved the predictive efficacy of the CoxPH model under a broader condition. The web calculator constructed for the study is based on the training dataset, and care should be taken when applying the EST model that may be overconfident. Hence, it is not recommended to use this algorithm for the prediction of survival. In this study, the CoxPH model had poorer long-term predictive power than the ensemble models. Therefore, use of the RSF model is recommended for the prediction of LPADC CSS beyond 10 years.
This study had several limitations. Firstly, in the SEER database, there was a lack of data regarding established predictors of survival in patients with LPADC (e.g., chemotherapy regimens and biological markers). Secondly, due to the retrospective nature of this study and data processing, samples with missing information were excluded; this may have led to considerable bias. Thirdly, the work related to the measurement of prediction model errors in the study is not yet complete. Finally, the results of this study were not externally validated; although we randomly split the study sample during the development of the models, the generalizability and reliability of this approach should be further validated with external datasets. The prognostic value of this approach should be improved in the future by adding more predictors, increasing external validation, and conducting prospective studies.
In conclusion, a geometric model and a CoxPH model were developed and evaluated for the prediction of CSS in patients with LPADC. Overall, all four models showed excellent discriminative and calibration capabilities; in particular, the RSF model and GBS model showed excellent consistency for long-term forecasting. The integrated web-based calculator offers the possibility to easily calculate the CSS of patients with LPADC, providing clinicians with a user-friendly risk stratification tool.