Stage I NSCLC is the very early stage of lung cancer. Due to its high heterogeneity and the development of technology of CT diagnosis, the treatment choices for different patients shown considerably large discrepancy (Donington et al 2017). With the continuous progress in precision and individualized treatment, the TNM staging system based solely on anatomical classification was not accurate enough for clinical usefulness. Therefore, it is urgent to develop a new prediction model to assist the surgeon in selection of suitable treatment strategies based on routinely clinicopathological variables. In this study, a predictive nomogram was established and verified to predict the prognosis of patients with stage IA NSCLC based on the pretreatment clinicopathological characteristics.
Many clinical factors could affect the survival of the patients of NSCLC (Merritt et al 2021; Koike et al 2013). Notably, the treatment management was an important prognostic factor for the lung cancer patients. Whitson et al. reported that lobectomy conferred superior overall (P < 0.0001) and cancer-specific (P = 0.005) 5-year survivals compared with segmentectomy in stage I adenocarcinoma (Whitson et al 2011). However, Dai et al. found that for T1aN0M0 NSCLC patients not suitable for lobectomy, segmentectomy should be recommended to those with tumor size less than 2cm (Dai et al 2016). For patients who were not surgery candidates, thermal ablation demonstrated better results in overall survival and acceptable local control. Mimae et al. reported that the 3-years OS rates were slightly better after wedge resection than segmentectomy plus lobectomy for patients over 80-year (89.4%, 95% CI, 73.8–95.9% vs 75.8%, 95% CI, 62.0–85.2%; P = 0.14) (Mimae et al 2021). Similarly, Linda Willen et al. found that surgeries were more common in the younger age group and the usage of stereo-tactic body radiotherapy (SBRT) increased with the increase of age (< 69years 5.4%; >85years 35.8%) (Willen et al 2021). Gender had also been confirmed as an independent prognostic factor in surgically managed patients. Chansky et al, reported that female patients who underwent surgery alone had 5-year survival rates of 56.8%, which was better than their male counterparts with only 48.3% (P < 0.0001) (Chansky et al 2009). In conclusion, the selection of treatments should be made based on multiple factors after the comprehensive evaluation.
By incorporating these factors, our nomogram showed perfect discriminative ability. The C-indexes were 0.704 (95%CI, 0.694 to 0.714) in the training cohort and 0.713 in the test cohort (95%CI, 0.697 to 0.728), respectively. It was better than the C-indexes in TNM, which were 0.550 in the training set (95% CI, 0.408–0.683) and 0.548 in the test set (95%CI, 0.401–0.672, P < 0.001), respectively. The 3-year and 5-year validation curves also showed high degree of agreement with the actual situation. Besides, the DCA curves revealed favorable potential clinical usefulness. All the evaluation above confirmed that our nomogram was an excellent model with a powerful prognostic performance to predict the OS and assist the making of treatment decision.
Several nomograms had been established to predict the OS of NSCLC after surgery (Birim et al 2006; Zhang et al 2014; Liang et al 2015). However, there is no model developed to predict the treatment efficacy for previously untreated patients. Our model was designed to help doctors and patients choosing the best treatment. For individual, surgeon calculated scores according to patient’s physical condition and tumor characteristics and then provided treatment recommendation based on the predicted survival. Our model compared the influences of different treatment strategies on the prognosis of NSCLC patients, which was easily practicable and in line with the actual situation of lung cancer.
Although some previous studies had demonstrated prognostic models for NSCLC, the results were slightly worse than ours. Yuan et al. showed a nomogram for cancer-specific survival of stage I NSCLC in 2019 with C-index of 0.64 (95%CI, 0.63–0.65) (Zeng et al 2019). The result of another study based on the lung adenocarcinoma in the stage I to III was only 0.69 (95%CI, 0.64–0.73) (Xie and Zhang 2021). The most important reason for the better result in this study was the large sample size. Our source of population-based data for model establishment and validation was the SEER program which currently captured 400,00 cancer cases annually and stored cancer data for approximately 34.6% of U.S. population from 18 (SEER) cancer registries. Our samples not only included different races and centers but also could be updated regularly, which guaranteed the timeliness of the model and minimized selection bias.
At present, many nomograms had been reported in the literature, but few were clinically applied. The possible reasons could be the difficulty in obtaining factors and high cost (Zeng et al 2019). David MJ et al (2012) reported a quantitative-PCR-baded assay to predict survival in resectable lung cancer. Some models were based on radiomics (Huang et al 2021) and artificial intelligence algorithm (Churchill et al 2021). These prediction models functioned well, but were too expensive with variables not easily available in all clinical settings. The variables in our model were easily available, which efficiently decreased the cost and increased the real-world practicality.
There were some limitations of this study. Firstly, this was a retrospective research and the data selection bias was unavoidable. Secondly, the cancer-specific survival (CSS) would be a more suitable metric compared to the OS, but determining the cause of death was unachievable in the SEER databases (Weiser et al 2011). Although we used multivariable analysis to reduce the impact of confounding, there were still some unobtainable factors in SEER databases such as smoking, pulmonary function and gene mutation. As the SEER database was from the United States, more than 85% of our cases involved white people. Therefore, the prediction for the Asia-Pacific population needs further verification (Shi et al 2014).