In our study, the use of different ML supervised models showed good predictive value for oncological outcome of UTUC after RNU. Using readily available clinical parameters, ML supervised models may provide accurate predictions of prognosis, and pTNM staging alone could be implemented to guide postoperative treatment. Among various experiments, the LR ML supervised model obtained the best results in predicting CSS at both 3 and 5 years, and the maximum AUC during training reached 0.85 and 0.84, respectively. On the other hand, LR(SVC) is more reliable during validation and has the best results with CSS 3 years. Although we acknowledge that we are not the first attempt to propose a post-RNU predictive model, existing models cannot be fully compared: (1) To date, we used one of the largest patient cohorts (n= 3129) for predicting the prognosis of UTUC. (2) We intentionally included two sets of patients with different ethnicities. (3) We investigated the applicability of ML supervised models in the UTUC field. (4) performed full external validation; In fact, the only external validation of the model was that of Yates et al. using 200 bootstrap resamples.
Adjuvant therapy is invasive and toxic, especially in patients with a single kidney, and the existence of better predictive tools may eliminate the need for adjuvant therapy. Many efforts have been made to create predictive models for the postoperative prognosis of UTUC. Nevertheless, due to lack of validation, these models remain unreliable in clinical practice, and no model is yet recommended with strong evidence in current European guidelines. POUT trial3are currently interested in validating adjuvant chemotherapy after RNU for UTUC, using only selected porpoise pTNM staging data. This includes pT2 to T4, pN0 to N3 M0 or pT any N1 to 3 M0. However, his preliminary subgroup analysis demonstrated wide variation in benefit for patients receiving adjuvant chemotherapy, highlighting the need for better stratification strategies after RNU that account for additional characteristics.
Several prognostic nomograms have been proposed, and these tools exceed AJCC/TNM staging in terms of survival prognosis in internal validation. Two of these studies4,6 We include UTUC patients undergoing surgery, whether RNU or other conservative surgery. Ku et al.11 Although limited to external validation studies, instead Krabbe et al.7We used different results from our study regarding recurrence-free survival and therefore cannot be compared. 4 models in total5, 8, 9, 10 Equivalent, but one uses the older WHO 1973 grading system8. Seven different independent prognostic factors (age, pT, LVI, location, CIS, architecture, pN) are used variably in these nomograms, with Cha’s model being the more comprehensive (7 features), followed by Seisen (6 features) and Roupret (5 features). All models evaluated 5y-CSS. Cha’s model also evaluated 2y-CSS. Neither the training nor the internal validation set has an AUC trade-off of more than 0.81 for CSS predictions. This supports the hypothesis that ML can implement existing models.
Furthermore, our study represents the first attempt to generate a reliable model in multiple ethnicities. Most of these nomograms did not take into account existing differences in Asian patients, who appear to have more advanced and more severe disease than other ethnicities.12. This may be explained by differences in genetic and epigenetic factors such as environmental and occupational exposures, lifestyle choices, and socio-economic factors.twenty three. We aim to move towards race-conscious healthcare, as suggested by Cardena et al.twenty four Because clinical research should be used to examine structural barriers, we decided to use two sets of patients with different ethnicities rather than using race as a proxy for biology. Therefore, our model has been tested on both European and Asian patients and is reliable regardless of the patient’s origin.
Various machine learning techniques are used in the field of urology, most of which are used in the lower urinary tract environment. (1) Regarding radiomics, AI has been implemented that can differentiate between bladder tumors and normal bladder in multiparametric magnetic resonance imaging (mpMRI).twenty five or computed tomography (CT) to determine the stage of bladder cancer26(2) Regarding prognosis, only the experience of Lam et al. Wang et al created and tested a number of AI algorithms that use clinicopathological evidence to estimate 5-year survival after radical cystectomy.27,28. This is, so far, the first experience investigating the potential application of ML supervised algorithms to UTUC, especially for post-RNU prognosis prediction.
Finally, this model can help clinicians address the challenge of stratifying UTUC patients and understanding clinical aggressiveness based on the baseline characteristics of this particular tumor. It can be used to enhance follow-up strategies for patients at high risk of recurrence, as well as to stratify candidates for adjuvant and subsequent therapy.
This study has some limitations. First, the multicenter nature of the study may have resulted in discrepancies in surgical technique, type of bladder cuff performed, use of intraoperative or perioperative mitomycin, use of neoadjuvant chemotherapy, and pathological diagnosis. Second, because the cohort spanned 2004, the use of two different pathological gradings may have influenced the algorithm. In addition, 2 patients in training and 15 patients in the validation cohort were diagnosed as pT0 in their final histopathology specimens. Even though this may reflect real-world data, on the one hand the prognosis for these patients is by definition excellent. Moreover, there are non-negligible differences in gender expression between the two cohorts, and therefore, differences in the underlying biology may influence treatment response and prognosis. Finally, the lack of centralized pathological correction of images and specimens can introduce bias.
