According to the data in 2021, despite the overall incidence remains relatively low, the incidence of UTUC and its proportion within UCs were significantly higher than before[20]. This rise may be attributed to advances in detection technology and improved survival rates of patients with bladder cancer (BC)[19–20]. As UTUC is biologically aggressive, more than half of the patients have an invasive state at the time of diagnosis [1, 19]. Furthermore, there is a noticeable increase in the number of patients in advanced stages[23].
Due to the rarity and individual heterogeneity of patients with UTUC, it is difficult for clinicians to assess their prognosis. However, the threat of UTUC to patients’ life underscores the need for useful prediction tools. In addressing this issue, evidence based on statistical analysis are recommended to be provided to make more objective assessment. For example, nomograms based on modeling can predict the probability of specific events at the individual level[21].
Previous studies have established various models in the field of UTUC, leading to great progress in predicting the survival outcomes of patients[3, 11–14, 16, 24–25]. However, most of these studies targeted on patient in the specific therapeutic stage, such as high-grade UTUC patients after radical nephroureterectomy[25]. In addition, the variables included in certain studies were relatively incomplete, focusing solely on one aspect such as clinicopathology or treatment strategy[3, 16, 24]. In this article, the study scope is broader, encompassing various disease stages, treatment approaches, and outcomes (including OS and CSS at 1-, 2-, 3-, and 5-year, respectively). And variables involved aspects such as demographic information, clinical pathology, treatment details, and follow-up data. Besides these, this study is the first to establish prognostic models for patients with UTUC utilizing 6 different machine learning algorithms (including LR, RF, SVM, KNN, ID3, and mainly XGBoost). Furthermore, the use of SHAP and the development of clickable web calculators can also be considered as innovations.
In this study, 12 variables were identified as independently associated with prognosis of patients with UTUC. Among these, three variables ("race, NHIA, and SEER stage") were selected through multivariable COX regression analysis (P < 0.05), while the remaining nine variables were screened by Lasso, including "sex, annual household income, months from diagnosis to treatment, tumor grade, T stage, side of primary tumor, examination of lymph nodes, radiotherapy, and chemotherapy."
According to our multivariable analysis findings, non-white patients exhibited a higher likelihood of experiencing poorer OS (P = 0.031), which aligns with previous studies[1, 27, 29]. For instance, Raman et al.[7], based on data spanning from 1973 to 2005 in the US, found that African-American patients with UTUC had poorer outcomes compared to white patients with UTUC. What’s more, a study in 2012 demonstrated that Asians tended to be diagnosed with a more advanced stage and a higher tumor grade of UTUC[27]. However, there were also conflicting views on the impact of race on the prognosis of UTUC, as an international study showed that ethnicity may not significantly affect the 5-year recurrence free survival (RFS) and CSS probabilities, necessitating further evidence[28]. As for NHIA, our study identified a link between non-Hispanic patients and poorer OS, in line with certain prior studies[7, 29]. A study based on SEER database reported that Hispanic patients were prone to have improved 5-year survival compared to black non-Hispanic ones[7]. However, the results of Hosain et al.[27] indicated that Hispanics usually have larger tumors of UTUC, therefore further discussions are needed to draw concrete conclusions. “SEER stage” was another significant variable in the multivariable analysis, categorized into three levels (including localized UTUC, regional metastasis, and distant metastasis). Our study showed that regional metastasis was associated with worse OS when compared to localized lesion. This finding was consistent with the conclusion drawn in a previous review[4].
Turning to the results of the LASSO regression analysis. In terms of demographic information, “sex” was included in the final model, suggesting its strong impact on prognosis. This finding aligned with previous studies conducted by Li et al. [14] and Ruvolo et al. [23], which reported that being female may be a risk factor for survival (P < 0.01). Another variable selected by LASSO was "annual household income." It is well-established that income significantly affects cancer prognosis, as conducting an economic evaluation of patients can help gauge their willingness and cooperation levels throughout the entire treatment process [31–32].
Clinicopathological characteristics including “T stage, tumor grade, and the tumor side” were regarded to be associated with prognosis of patients with UTUC by LASSO. The predictive ability of “T stage” and “tumor grade” had been well-documented [1, 11, 34]. For instance, a retrospective study in 2009 reported that pT played a critical role in affecting clinical prognosis, with patients in pT3 and pT4 stages had worse CSS [11]. Additionally, this study found that patients with grade 1/2 tumors had higher 5-year CSS rates (93.3% ~ 95.1%) when compared to those with grade 3/4 tumors (84.2% ~ 85.6%) [11]. Margulis et al.[34] also reported that tumor grade accurately predicted the recurrence and CSS of patients after radical RNU. Regarding “tumor side”, Ferro et al.[35] found that this variable had no independent effect on both the OS and CSS of patients with UTUC after RNU. In this study, although the LASSO results indicated that “the side where the primary is” was meaningful, the findings of both uni- and multivariable COX regression showed no statistical significance[34]. Therefore, the potential impact of tumor side on the prognosis of patients with UTUC remains to be determined.
In terms of treatment strategy, one of the key predictors identified was "months from diagnosis to treatment," emphasizing the importance of receiving treatment timely and appropriately. For instance, the latest European Association of Urology (EAU) guideline recommends that radical nephroureterectomy should be performed within 12 weeks after diagnosis[1, 30]. Lymphovascular invasion (LVI) occurs in approximately 20% of aggressive UTUC cases[38]. Recent reviews have indicated that LVI is a strong prognostic predictor for patients with UTUC[36–37]. What’s more, lymph node examination (LNE) can help remove undetected micrometastases and provide essential pathological information for diagnosis and treatment[39]. Given these findings, although there have been much fewer studies on how regional LNE affects the clinical outcomes in the field of UTUC, its importance can still be seen[33, 39]. Notably, Li et al. [33] founded that regional LNE was a protective predictors for patients with high-grade bladder cancer, providing valuable insight. Another two treatment-related predictor screened by LASSO were chemotherapy and radiotherapy. These treatment approaches are often utilized as neoadjuvants to RNU, aiming to improve the prognosis of patients with adverse clinical and/or pathological features[40–41]. Through the multivariable COX regression analysis and Kaplan-Meier curves, Zhang et al.[12] also observed that chemotherapy after surgery was associated with a better prognosis. Additionally, radiotherapy may play a important role together with chemotherapy[1, 42].
Despite the innovations of this study, there are several limitations. Firstly, as a public database comprising 18 individual registries in the US, most of the data originated from white individuals, resulting in a certain degree of information bias. Secondly, while this study includes numerous variables, some key factors like genetic characteristics have not been covered, and the details regarding treatment were limited. Finally, machine learning-based models require more interpretability when compared to traditional nomograms, and additional external validations are encouraged.